A conceptual framework proposal for the implementation of Prognostic and Health Management in production systems

Abstract

Prognostic and Health Management (PHM) is an emerging maintenance concept that is highly regarded by the scientific community and practitioners, as its adoption can bring economic, technical and environmental benefits to a company. PHM fully reflects the smart maintenance paradigm encompassing data collection, data manipulation, state detection, health assessment, prognostic assessment and advisory generation. Despite the undeniable benefits, there is still a large gap between the scientific and the real world. Several authors have investigated on the barriers to PHM implementation for companies, highlighting among them the lack of systematic approaches to its design and implementation. As a first contribution to this topic, the authors conducted a systematic literature review (SLR) to investigate the use of Decision Support Systems (DSSs) to support the PHM implementation. The SLR highlighted that few DSS had been developed and were limited to critical unit identification, maintenance strategy selection and data acquisition phase of PHM. Therefore, a conceptual framework for PHM implementation was provided as a second contribution. This framework summarises the decisions that should be addressed by a practitioner wishing to implement PHM services; moreover, it could lay the foundations for the development/improvement of the missing/existing DSSs for PHM implementation.

INTRODUCTION

Maintenance is a core function of manufacturing companies as it keeps production assets in good condition ensuring that they can perform at their optimal level. In the past the maintenance function was considered only as a cost [1], whereas, nowadays, maintenance has acquired a positive meaning as it could be considered as a strategic function contributing to the achievement of sustainable manufacturing operations [2]. As a matter of fact, good management of maintenance could lead to economic, environmental and social benefits [3], increasing, in this way, the ability of a company to be competitive in offering low prices, high quality and performance. In addition, good maintenance could mitigate one of the fearer aspects in manufacturing, namely unplanned downtime, which represents a huge cost in terms of production, revenue and reputation [4]. In this context [5, 6], Prognostic and Health Management (PHM) is gaining great success both in scientific community and in the industrial environments due to its ability to identify fault conditions in advance allowing to (i) reduce the maintenance activities, since they are performed only when they are necessary; (ii) reduce the unplanned downtime and maximising in this way the efficiency; (iii) maximise the equipment life; (iv) optimise the management of the spare parts and, (v) improve the safety of worker and environment. PHM services aim to address three main tasks which are detection, diagnosis, and prognosis of the fault condition [7]. The former focuses on the state of the monitored unit to detect abnormal conditions, the second instead, deals with the identification and isolation of the fault conditions, while the latter concerns the evolution of the failure conditions over the time, that is, the forecasting of the Remaining Useful Life (RUL). However, achieving the benefits promised by PHM requires a great effort for companies since its implementation involves the ownership of several technologies and skills. Indeed, sensing and data management systems to acquire, transmit, store, and manipulate the acquired data are a pre-requisite to setting up a diagnosis and prognosis system. Moreover, elaboration units and skilled practitioners in data analysis matter are needed to use data for PHM tasks [8, 9]. Finally, the full exploitation of the PHM outputs requires flexible production management in order to quickly reorganise it according to the recommendations generated by PHM solutions [10, 11].

In light of the above considerations, it appears clear that the implementation of the PHM solutions involves considerable challenges for companies who want to approach it, especially for Small and Medium Enterprises (SMEs) that often lack the necessary technologies and knowledge. Moreover, implementing PHM might require huge investment without the possibility to make an accurate trade-off analysis between the incurred cost and the achieving benefits [12–14]. Indeed, as proved by the authors in ref. [15] even if 80% of the surveyed manufacturing companies are interested to develop predictive maintenance (PdM), its applicability was still considered low. An interesting survey among 21 SMEs that operates with CNCs in United Kingdom in matter of PdM was conducted in ref. [16]. The questions of the survey aimed to investigate the following points: (i) the feasibility of adopting PdM for SMEs; (ii) the cost effectiveness for SMEs to use PdM; and the (iii) the value creation for SMEs by adopting PdM. The research showed that among the surveyed SMEs only 14% of them use a PdM-based maintenance strategy and that 42% of them rated the effectiveness of their current maintenance strategy as ‘moderate’ (3/5 of the scale) in terms of costs and impact on downtime reduction. Thus, the surveys confirmed on one hand the growing interest of the companies in investing in PHM solutions and on the other hand their hesitancy in its implementation.

The paper is organised as follows. Section 2 describes the context of the study, providing information about the PHM-oriented implementation path and the Decision Support Systems (DSSs), while Section 3 describes the main open issues in the PHM implementation and the contribution of the study. Section 4 describes the research methodology adopted into the paper, whereas Sections 5 and 6 reports respectively the systematic literature review (SLR) results and the proposed conceptual framework. Finally, Section 7 summarises the challenges and the future works.

CONTEXT OF THE STUDY

This section presents an overview of the context in which the SLR was conducted with the aim of introducing the key concepts of PHM, making clear the choices made by the authors in the selection of relevant papers and in the definition of the method for the information extraction from the selected papers. Indeed, a detailed description of all phases and steps involved in the implementation of PHM services are provided and then, some insights on the DSSs are discussed.

PHM-oriented implementation path

The implementation of the PHM involves two main phases named, respectively, ‘Pre—PHM phase’ and ‘PHM implementation phase’ as shown in Figure 1.

[IMAGE OMITTED. SEE PDF]

The ‘Pre—PHM phase’ aims to decide whether to implement the PHM strategy on a unit of the shop floor, typically the one considered as critical for the achievement of the company mission. This phase includes two steps: (1) the identification of the critical unit (CU); and (2) the selection of the most appropriate maintenance strategy to maintain it. The ‘PHM implementation phase’, instead, concerns the practical implementation of the PHM services and involves six steps in compliance with the best practices provided by standards, such as the International Standard Organisation (ISO)-13374 ‘Condition Monitoring and Diagnostics of Machines’ and the MIMOSA Open System Architecture for Condition Based Maintenance (OSA-CBM) standard [17]. The ISO-13374 defines the six top-level functions that the PHM should accomplish while the MIMOSA OSA-CBM provides the data structures and defines the interface methods for the functionality blocks defined by the ISO standard. Below, a description of the six functionalities is given.

Identification of the critical asset/unit: The critical asset/unit is the one whose failure has the greatest impact on the mission or objective of the company [18, 19]. For example, in ref. [20] the critical asset was identified based on failure frequency since the overall objective was to identify the maintenance strategy to maximise the availability of the production plant. It is worth noting that an ‘asset’ represents a system that performs an operational task and is composed of several units, where the ‘unit’ is the part of a system subjected to maintenance activities that cannot be further subdivided into sub-units [21]. Due to its importance for the company, its identification allows maintenance practitioners to focus their efforts on it to ensure its reliability [19, 22]. Thus, the goal of the step is the identification of the unit that has the most significant impact on the company in terms of performance and/or cost (due to downtime).

Maintenance strategy selection (MSS): Once the CU was identified, the next step concerns the identification of the most appropriate maintenance strategy to maintain it. Several maintenance strategies were developed over the years and the identification of the most appropriate one is a multi-objective problem since the needs of different stakeholders should be considered in the decision-making process [23]. Among these maintenance strategies, if the PHM solution is selected, the following six steps should be addressed for the implementation of its functionalities (detection, diagnosis, and prognosis).

Data acquisition (DA): All the PHM services to accomplish their tasks need consistent, clean, and reliable data. Indeed, acquiring and communicating such data accounts for 90% of the development of a PHM system. Failure to accomplish these tasks represents a bottleneck in realising the full potential of a PHM system [24]. Thus, this step aims to identify, install and configure all the technologies and tools to acquire, transmit and store data from the shop floor.

Data manipulation (DM): The acquired data need to be pre-processed in order to make them ready for the next steps in which knowledge is extracted from them to accomplish detection, diagnosis and prognosis tasks [24]. In particular, the typical operations performed on data in DM may concern the merge of data coming from different sources, the clean against outliers, the management of the missing values, the coupling of data with timestamps, and finally the conversion of data from analogue to digital format [25].

State detection (SD): The first service of a PHM system is related to the state of the monitored unit, that is, the monitoring system should be able to recognise the state of the monitored unit among all the possible states in which it could be [7]. The outputs of this service could be the current state indicator, threshold boundary alerts and severity of threshold deviation [25].

Health assessment (HA): Once an anomaly condition is identified at the previous steps, a PHM system should be able to make a diagnosis, that is, assess the presence of existing fault conditions and isolate it. Thus, the monitoring system should be able to infer the health status of the monitored unit by means of the ‘symptoms’ detected in the previous steps [26]. The outputs of this service could be the health grade, diagnosed fault, the root cause(s), the failed unit in multi-unit systems, recommendations, and an explanation about the failure [25].

Prognostic assessment (PA): The capacity to predict the future health state and failure modes based on the current health grade and planned usage loads is the powerful ability that a PHM system should have, that is, a prediction about the future trend of the degradation phenomena should be assessed [26]. The outputs of this service could be the future health grade and the RUL [25].

Advisory generation: PHM system should be able to exploit the information given in outputs from all the previous services. Indeed, the operational and maintenance advisories based on the information coming from the SD, HA and PA functions are generated [27]. These advisories should be processed as a ‘work request’ into the management system so that the spare parts and required tools can be prepared in advance. The operational advisory could be immediate or more strategic. The first case, for example, deals with the notification of some alerts to the operators while the latter case, for example, deals with the rearrangement of the production plan based on the high risk of failure but also with the forecast of the likelihood of accomplishing a production run based on the current system information. The outputs of this phase are the operation and maintenance (O&M) advisories, and the capability forecast assessment [25].

Decision Support Systems

The concept of the DSS is related to the decision-making process and refers to a computer technology solution that helps the user to take a decision when the choice is among different alternatives, and when it is very difficult to evaluate the pro and cons of each alternative [28]. The decision-making process involves three main phases that are [29, 30]: (1) the definition of the objective; (2) the identification of the alternatives; (3) the evaluation of the alternatives. Among these phases, a DSS should [31] (i) support the decision makers, rather than replace them; (ii) use data about the application domain to become aware of it; (iii) use data analysis techniques to find useful relationships among them; and (iv) evaluate the effect of each alternative against the defined criteria.

DSSs are widely recognised in the scientific community as tools to support the decision-making process in identifying a solution that meets the requirements of multiple stakeholders and/or takes multiple factors into account. In ref. [32], a DSS to assist multiple stakeholders in designing cost-effective O&M activities for offshore wind farms was developed. Experts' knowledge was used to identify the maintenance activities performed on the components of the offshore site. Technical, structural and environmental information of the offshore wind farm was considered as input while variables related to the personnel, vessel type and weather restriction were used as criteria to identify the optimal O&M plan. In ref. [33], a DSS was developed to enhance the maintenance and renewal management of railway tracks supporting the track managers and engineers to decide when to perform a maintenance or renewal activity taking quality and cost criteria into account. The decision rules were inferred from interviews with experts while data collected from the system were used as input to suggest the appropriate decision. The aircraft maintenance planning problem was addressed by means of a DSS by the authors in ref. [34]. Indeed, the developed DSS aimed to schedule the aircraft maintenance checks, activities and work shifts reducing the airlines' maintenance operations cost. The maintenance planning document provided by the aircraft manufacturer was used to define the types of checks and related activities and required skills, while the information related to the fleet's status and operational constraints were used to identify the optimal schedule. In ref. [35], the authors used a DSS based on a fuzzy Multicriteria decision-making method, that is, Analytical Hierarchy Process (AHP) and Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), to identify the most appropriate maintenance strategy for the CU in a manufacturing industry. The choice was among four maintenance strategies and their impact on four relevant factors for the company, that is, cost, safety, added value and feasibility, was assessed by experts' judgement.

The variety of applications described above demonstrates the flexibility in the use of DSS and their great contribution to facilitating decision-making in the presence of numerous stakeholders and constraints. In this paper, the use of DSS was investigated to support the implementation phase of the PHM within real contexts where several stakeholders and requirements are involved.

OPEN ISSUES IN PHM IMPLEMENTATION AND CONTRIBUTION OF THE STUDY

Several authors tried to address the issues related to the difficulties associated with the practical implementation of the PHM services. Arena et al. [36] reported three major issues towards PHM implementation: (i) poor quality of the collected data that make them unusable for diagnostic and prognostic purposes; (ii) difficulty in conducting a cost-benefit analysis; (iii) lack of skilled operators. The authors contributed to the second issue by developing a cost model able to link the PHM investment costs and related savings with the performance of the machine learning (ML) algorithm used to predict the state of the unit. Also, Zio et al. [12] identified the data collection process as an open issue, plus other two challenges related to (i) the physics of the problem since often the degradation process is not known and, (ii) the interpretability of PHM's model outputs especially when ML algorithms are used to make the predictions. As a contribution to the topic, the authors summarised the commonly used methods and tools to address the previous issues, reporting also, several examples of applications to clarify their meaning. Indeed, the most used fault detection, diagnosis and prognosis techniques were reported with respect to the challenges of missing fault data, missing label data and run-to-failure trajectories, respectively. Similarly, Lee et al. [18] provided a review of the commonly used algorithms and visualisation tools to support the selection of the most appropriate tools to be used in the implementation of PHM services. For each algorithm and visualisation tool, the application range was provided as well as the advantages and disadvantages. Moreover, a selection tool for wind turbines was developed to suggest which tools to use according to customer requirements and application conditions. Hu et al. [22] outlined three issues related to the design of PHM that are: (i) unclear requirements since there are still some difficulties in identifying and filtering the PHM requirements from an operational point of view transferring them from the conceptual high-level to the technical low level; (ii) often the design logic is not properly addressed as the tendency is to adapt some enabler technologies to PHM rather than select and acquire appropriate technologies; (iii) lack of verification, validation and feedback phases during PHM life cycle. Consequently, the authors proposed a high-level framework to show the critical activities and issues in the design and development of the PHM phases. Also, Saxena et al. [37] investigate the issue of the requirements definitions for PHM solutions, particularly in transferring high-level customer requirements to the lower technical level. To this purpose, the V-Model for PHM system development was proposed and suggested by the authors as a useful tool for the choice of the metrics and the requirements definition. Furthermore, the challenges related to cost-benefit analysis and performance specifications for PHM were also highlighted and to this purpose several cost-benefit analyses was reported and categorised.

Concluding, several reviews outlined that still nowadays the researches on PHM solutions are more focused on the technological part of PHM, the one related to the development of models and algorithms, for specific applications, rather than on the creation of a systematic approach for design and introduce PHM into an existing maintenance management system [38].

Summarising the scientific literature analysed so far, some authors focused on the identification of the appropriate algorithms to be used to perform the tasks of the PHM services, while others investigated the impact of the machine algorithms' performances in trade-off analyses of the PHM services implementation. Finally, several authors developed high-level frameworks and procedures to model the interactions and the criticalities among the different dimensions of the PHM service implementation and to support the selection of the most appropriate metrics and specifications.

Nevertheless, a framework covering the whole decision-making process related to the implementation of PHM services is still missing.

With this awareness in mind, this study aims to contribute to the development of a systematic procedure through two contributions. First, a SLR was conducted to deeply investigate the use of DSSs in the PHM implementation process. The results of the literature review (Section 5) showed that, despite the great interest in the PHM, there is still a lack of clear methodologies to support its implementation in industrial contexts. Indeed, among 146 reviewed papers only 10 DSSs were identified to support the critical decisions that must be taken during PHM implementation. As second contribution, a conceptual framework was developed to summarise all the decisions that should be addressed during the PHM services implementation showing at the same time, how there are linked among each other and with the stakeholders involved in the decision-making process (Section 6). The conceptual framework covers all the steps that go from the identification of the CU until the selection of the features and algorithms needed for detection, diagnosis, and prognosis tasks. The proposed framework could lay the foundations for the development or improvement, respectively, of the missing and existing DSSs for PHM implementation.

METHODOLOGY

This study conducted a SLR to investigate how DSSs were used in the scientific community to support the practitioner during the implementation of the PHM in real contexts. Thus, the research question (RQ) to which the SLR aims to answer is:

RQ) How DSSs are currently used in the PHM implementation?

The following sub-sections describe the steps through which the SLR was carried out.

Identification of research databases and keywords definition

Two scientific databases were used to find the scientific papers related to the RQ: ‘Scopus’ and ‘Web of Science’. As seen in Table 1, three groups of keywords were used to compose the research string: (i) Topic group includes two keywords related to the tool on which this literature review is focused on; (ii) Area of interest group includes four keywords related to the area of interest within which the use of the tool was to be investigated; (iii) Application group includes four keywords to narrow the scope to only industry and manufacturing contexts. The ‘*’ operator was used for some keywords to include all their possible declinations. Finally, the research string was composed by combining the keywords and groups through the two Boolean operators ‘AND’ and ‘OR’; in particular, the keywords of each group are linked with the ‘OR’ operator while the three groups are linked to each other with the ‘AND’ operator. In the following, the research string: <<TITLE-ABS-KEY ((‘predictive maintenance’ OR ‘Prognostic and Health Management’ OR ‘Prescriptive Maintenance’ OR ‘Proactive Maintenance’) AND (compan* OR industr* OR manufacturing OR producti*) AND (‘decision support system’ OR DSS))>>.

TABLE 1 Clusters of keywords used to create the research string.

	Topic group	Area of interest group	Application group
← OR →	‘Decision Support System’	‘Predictive Maintenance’	Compan*
DSS	‘Prognostic and Health Management’	Industr*
	‘Prescriptive Maintenance’	Manufacturing
	‘Proactive Maintenance’	Producti*
	← AND →

Literature research and papers selection

The keywords of Table 1 were searched within title, abstract and keywords of the papers. No range of years was defined. The relevant papers for the RQ were identified through two screening phases in which four exclusion criteria (EC) were defined. The first screening phase excluded (EC1) papers that are not in English and whose full text was not available and, (EC2) entire conference proceedings, books, books chapter, editorial and literature reviews. The second screening phase, instead, aimed to exclude papers that did not focus on the main two topics of the SLR, that is, with the Topic and Area of interest groups of keywords. Thus, two further EC were defined: (EC3) papers whose main topic was not consistent with keywords belonging to ‘Area of interest group’; (EC4) papers which did not clearly provide the key elements of a DSS.

Information extraction strategy

Figure 2 provides the details of the adopted information extraction strategy. The information from relevant papers were extracted taking inspiration from the Section 2.1 PHM-oriented implementation path: papers were first classified according to the phases and steps (Figure 1) which DSS deal with, then, the key elements of the DSS were extracted. Thus, the extraction process was guided from the following three questions: (i) Which of the two phases does the DSS refer to? (ii) Which step does the DSS refer to? and (iii) What are the key elements of the developed DSS?

[IMAGE OMITTED. SEE PDF]

SLR RESULTS

The search was run on 12 December 2022. The review process is shown in Figure 3. The research string produced 177 papers across the two scientific databases. The scanning for duplicated revealed 31 duplicates papers, thus, after removing them, 146 unique papers were identified. In the first screening phase 38 papers were excluded according to the EC1 and EC2. The remaining 108 papers were carefully read to assess their eligibility according to the RQ and the objective of the study. The second screening phase led to the exclusion of 28 papers due to the EC3 and, 70 papers due to the EC4. Therefore, at the end of the process, 10 papers were recognised as eligible to answer the RQ.

[IMAGE OMITTED. SEE PDF]

It is worth noting that the papers excluded by the EC3 were most focused on (i) the development of optimisation problems to create a link between maintenance and production plans taking advantage of the advisories generated by diagnosis and prognosis tools like in ref. [39]; (ii) the designing of DSSs for a field different from the manufacturing one, like in ref. [40] where ML based DSS was developed to predict the risk for projects in engineering procurement and construction domain; (iii) the optimisation of preventive maintenance (PM) parameters like the optimal inspection cycle, like in ref. [41] or (iv) the development of a new algorithm to optimise the opportunistic maintenance like in ref. [42] where the reinforcement learning was used, and (v) the modelling of complex systems like in ref. [43] where the system was modelled by means of the Bayesian network in a maintenance-oriented point of view. Instead, paper excluded by EC4 were most focused on (i) the implementation of new algorithms for diagnosis and prognosis tasks like in refs. [44–47]; (ii) the development of tool to identify the most appropriate ML algorithm according to the available data like in refs. [48, 49] where a suite of ML algorithms were provided and the most appropriate one was chosen through an iterative procedure; and finally on (iii) the description of success stories of PHM implementation in real contexts like in ref. [50] where a monitoring system was implemented in a thermal power plant and [51] where a predictive system was developed in the steel manufacturing lines of TATA company.

In none of the above-mentioned cases, the main elements of a DSS were provided and, for this reason, they were excluded from the analysis.

Results overview

The SLR identified 10 relevant papers that developed a DSS to support the implementation of PHM. As can be seen from Figure 4, 8 out of the 10 developed a DSS for the Pre-PHM phase addressing the decision related to (i) the identification of the CU with three DSS and (ii) the selection of the appropriate maintenance strategy to apply on it with five DSSs. The remaining two papers developed a DSS to address the first step of the PHM implementation phase concerning DA.

[IMAGE OMITTED. SEE PDF]

Based on the results and the PHM-oriented implementation path outlined in Section 2.1, it can be concluded that the Pre-PHM phase was addressed in the literature while there is a lack of DSSs for the PHM implementation phase since only two DSSs were developed to address the first steps. It is worth noting that the scientific community are not neglecting the other steps of this phase as proved by the huge number of papers considered in the EC4 EC. The issue with them is related to the fact that they are not addressed with a systematic approach and are specific for the case study, therefore it is not possible to identify the key elements of a DSS.

Criteria and methods used in the DSSs

This section outlines the criteria and methods used within DSSs to rank and select the most effective alternatives according to boundary conditions. In effect, criteria are used to evaluate qualitative and quantitative characteristics of alternatives within the DSS. They can express both the needs of the stakeholders involved in the decision-making process and the characteristics of the production unit considered. On the other hand, methods provide systematic approaches to evaluate each alternative against each criterion in order to find the alternative with the best trade-off against all criteria.

Eighty decision criteria were used within the developed DSSs to evaluate the fitness of the alternatives to the requirements of the company. Table 2 summarises the identified criteria, which were grouped into seven macro-areas: economic, strategical, technological, reliability, operational, human resource, and safety. They are described below according to the percentage of the decision criteria belonging to each of them (descending order).

TABLE 2 Decision criteria grouped per macro area.

Area	Decision criteria	Reference
Economic	Value of the unit; total operation cost resulting from unit failure; cost of corrective maintenance action; consequence (time and cost to restore the functionality); inventory cost; estimated maintenance cost; cost of downtime; variable cost of maintenance; fixed cost of maintenance; cost of overtime; expected cost for design and development of the maintenance; required payback period; cost of the diagnostic technique; cost of preventive maintenance action; yearly cost of the current predictive maintenance; yearly cost of predictive maintenance in the best and in the worst case; ratio between expected cost due to corrective maintenance actions and planned replacement	[36, 52–58]
Strategical	Criticality priority; regulatory compliance; outsourcing of job; outsourcing of manpower; proactive management; management skill; quality of product; expected frequency of preventive inspections; size of the company; type of ownership; type of capital; the way of supervision; rejection rate of false positive	[36, 52–56]
Technological	CMMS for documentation, communication, monitoring, testing and control; automated inspection system; accuracy of the tools; external factor influence; technology maturity; early failure prediction; support from the OEM; number of detectable failures; portability; integration with company informative systems; data scoring frequency; false positive rate; true positive rate; security (data privacy issues)	[36, 54, 56–58]
Reliability	MTTR; MTBF; MTTF; Expected probability of failure without maintenance; expected probability of failure with maintenance; occurrence; downtime for failure; total numbers of failures per year; availability; expected probability of failure after PdM implementation; maintainability; Weibull distribution parameters	[36, 52–56, 58–60]
Operational	Yearly working hours; quality of design; material quality; work in process material quality; component specific: Load, vibration and temperature; capability; quantity needed; type of industry; type of production; type of machines owned	[36, 54–56, 59, 60]
Human resource	Training of workers and personal; ease acceptance by workforce; flexible workforce; empowered worker; cross functional team; ease of use; job uncertainty; training needed for the tool; effect on the human resources	[54, 57, 58]
Safety	Worker safety; machine safety; safety of material; safety of environment	[54, 57]

The economic area includes 21% of the decisional criteria which aim to evaluate both the purchase and design costs of the alternatives and their economic impacts towards the achieving of the company's goals, such as the costs related to the failure of a unit or the return period of the investment. The strategical area includes 18% of decisional criteria which aims to provide some characteristics of the company like its size, ownership, capital. Moreover, the strategical criteria aim to define some acceptable thresholds for parameters like the quality of the products and the rejection rate of false positive (management of the misclassified prediction). The technological macro area includes 18% of criteria. These criteria aim to assess the capability and the performance of the alternatives from a technological point of view, for example, for an algorithm its false and true positive rates were considered, while for the DA technologies their accuracy, data scoring frequency and early failure detection were assessed. The reliability area includes 15% of decision criteria, which aim to collect information on (i) the reliability of the unit such as mean time between failures (MTBF) and MTTF; (ii) the expected increase in the reliability of the unit by comparing, for example, the probability of failure with and without the maintenance activity and (iii) the performance of the maintenance management by considering the mean time to repair (MTTR). The operational area included 13% of decisional criteria and they are related to the production aspects of the company like the type of production, the yearly working hours, and the capability of the machines. The human resource includes 11% of decision criteria which aim to investigate the impact of the alternative on the workers and their interaction with them such as the ease of acceptance or the quantity of needed training sessions for a new tool. Finally, the area linked with the safety aspect includes 5% of decisional criteria related to different dimensions of safety, that is, worker, environment, and machine. Figure 5 gives a visual presentation of the number of criteria for each macro-area.

[IMAGE OMITTED. SEE PDF]

Moreover, the ratio of the number of decision criteria belonging to each macro area and the total number of decision criteria found for each DSS objective was computed. For example, if we consider the DSS for CU identification, 14 decision criteria were identified distributed for macro areas as follows: 7 decision criteria belonging to the Reliability macro area, 2 to the Strategical macro area, 1 to the Operational macro area and 4 to the Economic macro area; so that the percentage for Economic macro area can be computed as follows: 4/14 = 29%. Such a distribution highlighted how the importance of each decision criteria macro area is closely linked to the final purpose of the DSS. As shown in Figure 6, each of the three scopes, that is, CU identification, MSS and DA technology, relies on a different decisional criteria macro-area to take the final decision. For the decision of the CU, the most used decision criteria were those belonging to the reliability area, followed by economic and strategical ones. This indicates that determining the CU within a manufacturing plant depends on its reliability parameters such as MTBF, MTTF and Weibull parameters, as well as the economic consequences of its failure and its impact on achieving company goals. For the decision on the most appropriate maintenance strategy, the economic and strategic decision criteria were the most used, followed by operational and reliability criteria. This suggests that this type of decision is taken by company managers from a strategical point of view to maximise the company's profits. Finally, for the decision on the most feasible DA technology, technological criteria were the most used, followed by those related to human resources. Therefore, in this case, the technological capabilities of the tool have the greatest impact on its choice, and its interaction with workers is also considered.

[IMAGE OMITTED. SEE PDF]

Unlike the decision criteria, for the methods used in the developed DSSs it was not possible to infer a trend in their use due to the limited number of papers and the fact that each paper used a different method. Several Multi Criteria Decision Making (MCDM) methods were used to evaluate the effectiveness/impact of each alternative against the mission of the company like as Analytical Hierarchy Approach (AHP), or Fuzzy Logic based methods. Moreover, two classification algorithms were used, that is, Decision Tree (DT) and Rough Set Theory (RST). Finally, some mathematical frameworks were developed to investigate the different alternatives such as Markov decision model and costs model.

DSSs developed for critical unit identification

Three DSSs were found for the identification of the CU within an industrial plant according to the mission of the company. The common objective of the developed DSSs was to find the unit on which focus the maintenance efforts according to the mission of the company. The company in ref. [52] wanted to minimise the cost of the Shutdown Projects that consists of the cost associated with the production loss plus the costs related to the maintenance activities, thus the objective of the DSS was to identify which operational unit include in the project. Instead, the maximisation of the availability led the identification of the CU in refs. [53, 60], where the DSS aimed to identify the CU of the belt conveyor system on which install a condition monitoring system to enable PdM, and to identify the subsystem on which optimise the parameters of autonomous PM on the most critical subsystem, respectively. In all the DSSs the CU among those owned by the company was identified, that is, the alternatives, are the operational units owned by the company in refs. [52, 60] and the subcomponents of the system under investigation in ref. [53]. In two DSSs, the MCDM methods were used to rank the alternatives according to several criteria and experts' judgements such as Fuzzy logic in ref. [52] and failure modes and effects criticality analysis (FMECA) method in ref. [53]. While in ref. [60] the whole system was modelled by means of the Markov decision model and the simulation was used to evaluate the impact of the failure of each subsystem on the system's availability. As can be seen from Figure 6, the most used decisional criteria are the ones belonging to the reliability macro area aimed to characterise (i) the reliability of the unit with MTBF and MTTF; (ii) the performance of the maintenance activities with MTTR and expected probability of failure with and without maintenance; (iii) the impact of unit failure on the availability of the system. The economic criteria are the second most used, and they aimed to evaluate (i) the economic impact of the unit failure and (ii) the cost of the maintenance activities. Even if there are differences in the methods used for the CU identification, a common approach for the identification of the CU can be identified and it is summarised in Figure 7 while the relevant information about the DSSs are reported in Table 3.

[IMAGE OMITTED. SEE PDF]

TABLE 3 DSSs developed for the identification of the CU in Pre-Prognostic and Health Management phase.

DSS for CU identification
Reference	Objective	Alternatives	Criteria	Methods
[52]	Decide whether include or not an operational unit in the shutdown project	List of operational units in the shop floor	Cost of corrective maintenance action; criticality; expected probability of failure with maintenance; expected probability of failure without maintenance; MTBF; MTTR; total operation cost resulting from unit failure; value of the unit	J-Factor, Fuzzy logic
[53]	Identification of the CU to equip with condition monitoring system	List of subcomponents that could lead to a stoppage of the considered system	Consequence (time and cost to restore the functionality); occurrence; priority	FMECA
[60]	Identification of the CU on which optimise the maintenance parameters	List of subsystems included in the considered systems	Downtime for failure; MTTR; total numbers of failures per year; yearly working hours	Simulation and Markov decision model

DSSs developed for maintenance strategy selection

Five papers developed a DSS to support the decision about the appropriate maintenance strategies to implement on the CU. The common objective of the DSSs was to identify the most appropriate maintenance strategy to apply to the CU according to the mission of the company. Although the DSSs shared a common objective it was achieved in different ways across the five papers as can be seen in Table 4. Indeed, in ref. [54] were used five MCDMs methods based on Fuzzy logic, that is, Fuzzy TOPSIS, Fuzzy line graph, Fuzzy spider graph, Fuzzy digraph and matrix approach and 26 criteria belonging to the seven macro areas defined in Table 2 to identify the most appropriate maintenance strategy among five ones to maximise the availability of the critical subsystem. A set of equations was developed by the authors in ref. [55] to select the most appropriate maintenance strategy among the corrective maintenance (CM), PM and PdM to implement on the CU to maximise the profitability of the company over a chosen period. The equations aimed to investigate four aspects of the unit and the company and are related to (i) the unit efficiency and its proximity to the critical path; (ii) the cost aspects; (iii) the breakdown information and (iv) the maintenance system design to select the appropriate maintenance strategy to implement on the CU. The identification of the maintenance strategy with respect to the maximisation of the company's profitability was also addressed by the authors in ref. [56] but with special attention to the performance of the ML algorithm used within the PdM strategy to predict the health status of the monitored unit. Indeed, the authors developed a cost model to compute the yearly expected cost for PdM based on an ML classification algorithm considering as performance indicators the true positive rate (H) and false positive rate (F) that represents the fraction of correct and wrong decisions provided by the algorithm, respectively. Thus, the suitable maintenance strategy between CM and PdM against the best and the worst case for the ML algorithm performance was identified through the DSS. The previous PdM cost model was also used by the authors in their work [36] to assess when PdM is more convenient than the CM according to the performance of the ML algorithm and occurrence and severity levels of the FMECA analysis. Indeed, three levels of occurrence and severity were considered, and a DSS based on DT was built for each pair of them (occurrence, severity) to identify the conditions for which the PdM is more convenient against the CM. In addition to the performance of the ML algorithm other parameters were considered as DT attributes, that is, the parameters of Weibull distribution, the data scoring frequency, MTBF and the yearly working hours for the occurrence and the ratio between expected cost due to CM actions and planned replacement and the cost of the CM action for severity. It is worth noting that for a few pairs of values (occurrence, severity) the performance of the algorithm was not considered by the DT as a decisive criterion for the choice of maintenance strategy. A DSS based on the DT was also developed by the author in ref. [59] to identify the Lean Maintenance (LM) method to be implemented in the maintenance area of a company according to its characteristics to maximise availability. Indeed, methods such as 5S, Single Minute Exchange Die, Kaizen, reliability centred maintenance, Total Productive Maintenance, Kanban system, Poka Yoke are part of LM aiming to reduce or eliminate the waste and losses involving several areas of the company management [61]. The authors surveyed 65 manufacturing companies to investigate the impact of the LM method against the Number of Unplanned Downtime (NUD) and the Overall Equipment Effectiveness. Then, through a DT they established a link among the used methods, the NUD and the characteristics of the company like its size, the type of capital, industry, machines owned, ownership and production. Finally, the authors compared the results of the DT with the ones provided by the RST algorithm.

TABLE 4 DSSs developed for maintenance strategy selection in Pre-Prognostic and Health Management phase.

DSS for maintenance strategy selection
Reference	Objective	Alternatives	Criteria	Methods
[54]	Identification of optimal maintenance strategy to maximise the plant availability	CM; PM; CBM; PdM; RCM	Inventory cost; estimated maintenance cost; automated inspection system; CMMS for documentation, communication, monitoring, testing and control; component specific: load, vibration and temperature; cross functional team; ease acceptance by workforce; empowered worker; flexible workforce; management skill; material quality; MTBF; MTTR; out sourcing of job; out sourcing of manpower; proactive management; quality of design; regulatory compliance; safety of environment; safety of material; safety to machine; safety to worker; training of workers and personal; work in process material quality	Fuzzy TOPSIS, Fuzzy line graph, Fuzzy spider graph and Fuzzy diagraph and matrix approach
[55]	Identification of the optimal maintenance strategy to minimise the payback period for its investment	CM; PM; PdM	Availability; capability; cost of downtime; cost of overtime; downtime for failure; estimated maintenance cost; expected cost for design and development of the maintenance; expected frequency of preventive inspections; expected probability of failure after PdM implementation; expected probability of failure with maintenance; Fixed cost of maintenance; MTTF; MTTR; quality of product; quantity needed; required payback period; variable cost of maintenance	Analytical method
[59]	Identification of the optimal method of the lean maintenance to maximise the plant availability	5S; SMED; Kaizen; RCM; TPM; Kanban system; Poka Yoke	MTTR; size of the company; the way of supervision; type of capital; type of industry; type of machines owned; type of ownership; type of production	CART Decision Tree and Rough Set Theory
[56]	Identification of the most profitable maintenance strategies	CM and PdM	MTTF; cost of corrective maintenance action; cost of preventive maintenance action; data scoring frequency; rejection rate of false positive; Weibull distribution parameters; yearly cost of predictive maintenance in the best and in the worst case; yearly cost of the current predictive maintenance; yearly working hours	Analytical method
[36]	Identification of the most profitable maintenance strategies	CM and PdM	Cost of corrective maintenance action; data scoring frequency; false positive rate; MTTF; ratio between expected cost due to corrective maintenance actions and planned replacement; rejection rate of false positive; true positive rate; weibull distribution parameters; yearly working hours	Decision Tree

The common approach addressed by the developed DSSs for the identification of the most appropriate maintenance strategy to apply on the CU is summarised in Figure 8.

[IMAGE OMITTED. SEE PDF]

DSSs developed for data acquisition technology selection

Two DSSs were found for the identification of the most appropriate data acquisition technology (DAQ) to instal on the CU to enable the PdM. A three stages framework for identifying the DAQ to enable PdM in the railway infrastructures was developed in ref. [57]. The first stage of the framework aims to identify all the feasible DAQs for the unit under investigation listing its failure mode, the second stage concerns with the definition of the evaluation criteria to rank the DAQs while the third stage concerns the acquisition strategy for the selected DAQs, that is, develop technology in-house, externally source or a mix of them. In addition to the previous workflow the authors in ref. [58] classified the identified DAQs in four levels according to their diagnostic capacity, that is, the levels go from 0 which includes elementary portable systems with a low diagnostic capacity until level 3 which includes on-line sophisticated system with an excellent diagnostic capacity. In both studies AHP method were used to rank the DAQs against the mission of the company. As can be seen from Figure 6, the most used decision criteria belong to technological macro area and they aimed to evaluate the capability of each DAQ in terms of accuracy, failure prediction, number of detectable failures and portability. Also, the criteria related to the human resource were considered aimed to evaluate the ease of interaction between the workers and DAQs. It is worth noting that, even though in both cases the DAQ was chosen to enable the PdM, the compatibility between the DAQ and the algorithm that could be used in the next steps (diagnosis and prognosis tasks) was neglected. The common approach addressed by the developed DSSs for the identification of the CU is summarised in Figure 9 while the relevant information was reported in Table 5.

[IMAGE OMITTED. SEE PDF]

TABLE 5 DSS developed to identify the most appropriate data acquisition technology in Prognostic and Health Management implementation phase.

DSS for data acquisition
Reference	Objective	Alternatives	Criteria	Methods
[57]	Identification of the appropriate DAQ to enable PdM	List of all feasible DAQs according to the CU failure modes	Accuracy of the tools; cost of the diagnostic technique; early failure prediction; ease of use; external factor influence; job uncertainty; safety; security; support; technology maturity	AHP
[58]	Identification of the appropriate DAQ to enable PdM	List of all feasible DAQs according to the CU failure modes	Accuracy of the tools; cost of the diagnostic technique; ease of use; effect on the human resources; integration with company informative systems; maintainability; needed training for the tool; number of detectable failures; portability	AHP

Success stories of PHM implementation

Since there were only 10 papers in the SLR that proposed a DSS for one or more steps of PHM, the authors decided to propose a framework for PHM implementation. For designing the framework in the best way possible, the authors also reconsidered the papers excluded before in the SLR process (because they did not present a DSS for PHM implementation) to analyse if there were some describing success stories in PHM implementation.

Among the results of the literature review, six papers that describe the implementation of some PHM's functionality in real contexts were found. Although these papers do not provide any of the main elements of a DSS, they represent a source of knowledge, thus, similar to the authors in ref. [7] where the success stories were used to prove the potential that PHM could bring in different business contexts, in this paper, the success stories are used as a source to comprehend what are the decisions that must be addressed in the implementation of the PHM solutions. A brief description of the stories is then provided in the following.

The implementation of a monitoring system to enable the assessment and visualisation of the health conditions of a thermal power plant's main subsystems is described in ref. [50]. An overall health index was created for the whole plant by combining the health indexes created for each of its six sub-systems. Instead, the authors in ref. [51] described the implementation of a data-driven approach to predict the wear of the CU, namely the pot gear, in the TATA steel company. Before the project, the unit was maintained by a time-based maintenance strategy (every 4 weeks) then PdM was implemented to maximise the life of the unit and to enhance the performance of the maintenance activities. An approach to enhance the quality of machined products by enabling predictive monitoring on a machining machine is described in ref. [62]. The authors used the robotic arm that loads and unloads pieces from the machine to measure their dimensions and create in such a way a method to label the operative parameters collected from the machining machine. The main purpose was to establish a link between the dimensional errors of the machined products and the operating parameters of the machining machine. Also, the authors in ref. [63] dealt with a machining machine and, in their study, they described how an acoustic sensor was used to detect the wear state of the machining tool. An approach, capable of predicting a failure condition 7 days in advance, was developed for a centrifugal pump in the oil and gas industry [64]. Finally, the authors in ref. [65] described the SIMPREBAL project that aimed to implement condition-based maintenance to monitor and diagnose hydrogenators' machinery malfunctions.

A CONCEPTUAL FRAMEWORK TOWARDS A DSS FOR PHM IMPLEMENTATION

The review highlighted the lack of DSSs to support the PHM implementation phase. This awareness prompted the authors to define and report on all the decisions that must be taken during the implementation of the PHM services in order to lay the foundations for the development/improvement of DSSs related to the PHM implementation phase. To this end, a ‘conceptual framework’ (Figure 10), towards a DSS for PHM implementation, was developed exploiting the information coming from several sources, reported in the following: (i) the 10 relevant papers (Sections 5.1–5.5); (ii) six papers found among the results of the literature review that described in detail success stories of PHM implementation (Section 5.6); (iii) PHM-related papers and (iv) knowledge of the authors in the PHM domain.

[IMAGE OMITTED. SEE PDF]

The proposed conceptual framework highlights the challenges in each PHM step implementation and aims to be a ‘roadmap’ for practitioners to help them identifying the steps and information needed to achieve their goals. The challenges to be addressed for each step have been reported in the conceptual framework in the form of questions above the arrows (see Figure 10). For example, to implement the SD service, it is necessary to define the states in which the production unit could be in, so the challenge is: ‘Which operational states are relevant to the needs of the company?’ Another example is of challenge is related to the data gathering: to obtain useful data to perform the PHM services, it is necessary to define the monitoring parameters related to the failure modes, so the challenge is ‘What are the symptoms of each failure mode?’

Furthermore, for most of the challenges, some key technologies and methodologies/examples were provided considering the scientific literature. For example, to define the states for SD service, two examples from the literature were reported below, while the FMECA analysis was suggested to identify the monitoring parameters related to the failure modes.

Below, a detailed description of conceptual framework for PHM implementation is provided. Then, a simple scenario of the framework utilisation is provided, at the end of Section 6, to show the feasibility of the framework.

As can be seen from Figure 10, the decisions (white boxes) included in the conceptual framework involve three main actors (yellow boxes) that are the production units owned by the company and the stakeholders belonging to two company areas, that is, production and maintenance. The production stakeholders are mainly interested in maintaining and improving the productivity levels and quality of the manufactured product, while the maintenance stakeholders are mainly focused on maintenance activities. In addition, the questions to be answered to take the decision (over the arrows) and the correlation among the decisions (arrows) were provided (Figure 10). The decisions were grouped into three main areas by green, blue and grey boxes and are reported following the steps shown in Figure 1.

The first step concerns the identification of the CU, this is the unit whose failure has the most negative impact on the achievement of the company's goals. Thus, this step aims to answer the question ‘What is the most impacting unit?’. The CU is selected among the ones owned by the company through the identification of the failure modes that can affect them (What is (are) the failure mode(s) and its(their) effects?). To this scope the FMECA analysis was used in ref. [53] while maintenance sheets were used in refs. [50, 64]. Then, evaluation metrics should be defined to measure the impact of the failure against the stakeholders' needs (How much does failure impact the stakeholders' needs?). Indeed, the metrics aim to convert the needs into quantifiable values to use them in the decision-making process. It can be noted that the impacts are evaluated against the needs of the production stakeholders since they are mainly linked with the production losses and safety aspects. Indeed, as summarised in Table 2 such criteria could be representative of different aspects of the company's management, for example, economic, strategic, reliability, and safety.

Once the CU has been identified, the most appropriate maintenance strategy to maintain it must be chosen. Thus, this step aims to answer the question ‘What is the most appropriate maintenance strategy to maintain the CU?’. As discussed in Section 5.4, this is a multi-objective problem since several, often conflicting needs, must be considered in the choice of the most appropriate maintenance strategy. For example in refs. [36, 56], the economic and technical constraints related to the development of ML algorithms were considered in the selection of the maintenance strategy, while in ref. [54] MCDMs approaches were used to select the best maintenance strategy to increase the availability of the system against 26 criteria. The feasible maintenance strategies should be identified based on the CU and its failure modes (What maintenance strategies can be used to maintain the CU?). Indeed, the most appropriate maintenance strategy is the one that both contributes to the reduction/deletion of the failure negative effects (How much does the maintenance strategy contribute to reducing the failure impact(s)?) and matches with the company's goals and constraints (How well does the maintenance strategy meet the needs of the stakeholders?). Both production and maintenance stakeholders are involved in the decision-making process. In particular, the latter owns a deep knowledge of the failure modes and the ways to mitigate and prevent them.

If the PHM approach is selected, the decisions included in the blue and grey boxes must be addressed in order to enable the detection, diagnosis and, prognosis services. These services make use of data acquired from the unit to accomplish their tasks so that the appropriate data acquisition technology(ies) must be identified and installed. Therefore, the blue box aims to answer the question ‘What is the most appropriate data acquisition technology(ies)?’. As described in Section 5.5 the monitoring parameters should be identified based on the symptoms of the failure mode (What are the symptoms of each failure mode?). For example, experts' judgement was used in ref. [51] to identify the most useful variables to predict the wear state of the pot gear, while a piezoelectric sensor was installed near the machining tool in ref. [63] to monitor its wear state through acoustic emissions. Once the monitoring parameters have been identified the feasible DAQs are identified by answering the question ‘How can they be acquired?’. Thereafter, the most appropriate DAQ is chosen by also considering the needs of the stakeholders through the evaluation metrics (How well does the DA technology meet the stakeholders' needs?), for example, four macro-area of criteria were considered in ref. [57] to measure the fitness of the DAQ against the company's needs.

The acquired data are pre-processed in the DM step to use them in the next steps by means of pre-processing methods (Does data require pre-processing?). For example, the outliers analysis was conducted in refs. [50, 51]. In addition, the under-sampling technique was applied in the former while the normalisation was applied in the latter. Instead, linear interpolation was used in ref. [64] to fill the missing values. Furthermore, the data coming from different sources should be merged and synchronised [62].

The three services due to their specific purpose require additional information as input from the data. The left-hand side of the grey box is dedicated to this purpose. The SD service aims to detect the state of the unit, that is, the set consisting of the input data, the features and the algorithm must be able to distinguish the state of the unit among a list of states in which it could be in. For this reason, its implementation requires the identification of the relevant states in which the unit could be and their link with the monitoring parameters (Which operating states are relevant to the needs of the company?). In general, the number of states depends on what can be measured and it is closely linked to the state conditions that are relevant to the company's needs [21]. For example, seven classes of wear were defined for the machining tool in ref. [63] while four operational states were established in ref. [65] namely: normal, alert, alarm and trip.

Once an abnormal condition has been detected in the previous step, the HA service aims to diagnose the existing fault condition, that is, assess the presence of the fault condition and isolate it. This functionality wants to provide a measure of the unit's degradation through a health index like in ref. [50] where a health index was created for each monitored sub-system and for the whole system. In addition, the diagnosis task could also distinguish among the failure modes the one that is affecting the unit by analysing the symptoms. To this scope, it is necessary to provide the ambiguity groups in which the relationship between the symptoms and the failure mode is established (Can the failure modes be separated by monitoring parameters?). The FMECA analysis could be used for this scope, that is, to create the rules between the symptoms and the root causes [65].

Finally, after the fault condition has been detected and isolated, the PA functionality aims to predict the degradation of the unit over the time providing in output its RUL. The RUL should be adequate with respect to the response time of the company to be exploited effectively by it. Indeed, the minimum forecasting windows should be provided as a requirement for the prediction (What is the minimum useful length of the forecast window?). For example, a period of 1 week was set as a requisite for the prognostic algorithm in ref. [64] since it corresponds to the minimum time required to apply the maintenance to the unit before its failure.

Each service to accomplish its task requires feature(s) and an algorithm. The central part of the grey box, named ‘Data Analysis’, is dedicated to this purpose. Indeed, for each service, it is necessary to check if the input data are adequate to create features, that is, measurable properties and characteristics of the analysed phenomena which are indispensable for the PHM services (Are the information provided by data relevant for the goal?). Furthermore, also the algorithms to extract features from the data must be chosen (How can relevant information be extracted from data?). For example, the Fast Fourier Transform [63], the unsupervised learning techniques like hierarchical clustering and K-means clustering [51], and the backward simple moving average [64] could be used to create features. Lastly, suitable algorithms must be identified according to the task and the characteristics of the data (Which algorithms can be used to achieve the goal?). Indeed, several regression algorithms were tested in refs. [51, 62], while classification algorithms were used in refs. [50, 64]. As can be seen from the conceptual framework (Figure 10), the identification of the best features and the algorithm to accomplish the task involves an iterative procedure where the performance of the several solutions was tested against the evaluation metrics selected on the basis of the type of algorithm and task (How can the performance be assessed?). For example, the Root Mean Squared Error and the R squared value could be used to assess the performance of regression tasks while F1 score and confusion matrix could be used to measure the performance of classification task. Finally, the acceptance thresholds for the computed performance should be set considering the stakeholders' needs (How well does the performance of the method satisfy the stakeholders' needs?).

It is worth noting that the methods identified within the blue and grey boxes are strictly connected, that is, the decisions in one box affects the ones in the other box. In fact, the data acquired and processed by the methods identified in the blue box may prove to be inadequate for the implementation of certain PHM service or may not allow for good output performance. Therefore, in such cases, it is necessary to go back and revise previous decisions. Concluding, all the decisions included in the conceptual framework should be addressed for all the units owned by the company and for all the failure modes, in order to select the most appropriate maintenance strategy for each of them.

To show the feasibility of the framework, a simple scenario is provided in the following. Let us imagine that a practitioner wants to implement the SD service on a production unit. Certainly, some questions arise in his/her mind, such as: (1) What is the first step I should take? (2) What do I need to implement the SD service?

From the scenario described, we can assume that he/she has already identified the CU and has decided to implement one of the PHM services; thus, within the framework he/she is at the beginning of the grey area, at the ‘Is the SD to be implemented?’ rhombus and wants to reach the ‘SD feature(s) and algorithm’ block. The first block (States) asks him/her to define the states in which the CU could be, that is, the states that the SD service must recognise. These states delimit the operating ranges of the production unit according to the company's objectives, for example, they can be related to product quality or production speed.

The second block (Data Analysis) deals with the data to be collected and the data analysis methods to be developed in order to implement the SD service. Specifically, the practitioner must, in order: (1) define the monitoring parameters; (2) install the DA technologies; (3) check the quality of the acquired data and, if necessary, pre-process them; (4) identify the feature to characterise the states; and (5) identify the algorithms to extract knowledge from the data. Therefore, the framework guides the practitioner towards the definition of the main steps that should be implemented for the SD service.

DISCUSSION AND CONCLUSIONS

PHM is a hot topic among the scientific community and practitioners in real contexts as its adoption could bring several benefits to a company in achieving economic, technical and environmental goals. Two main regulations, ISO 13374 and OSA-CBM, define the functionalities and data structure of PHM. The benefits of the PHM solutions are due to its three main services, namely, SD, HA and PA. SD deals with the detection of the current state of the monitored unit evaluating the variance between some parameters and their baseline. The HA, instead, deals with the failure conditions. Indeed, once an abnormal condition is detected by SD, the HA aims to assess the existent fault condition and isolate it by identifying the root cause(s), the failure mode and the affected unit. Finally, the PA deals with the evolution of the degradation process over time by providing the RUL of the unit.

However, despite the efforts of the scientific community and regulators, most industries consider the PHM as a concept that is still far from real applicability. One of the main limitations, as outlined by the scientific literature, concerns the lack of systematic approaches to guide companies and practitioners during the implementation of PHM solutions.

To contribute to this gap, first of all, this study investigated the use of DSSs as tools to support the implementation of PHM solutions. The entire PHM implementation process was considered, two main phases named ‘Pre PHM-phase’ and ‘PHM implementation phase’ and eight steps were defined (Figure 1). The results of the SLR showed that, despite the great interest in PHM, there is still a lack of clear methodologies to support its implementation in industrial contexts. Indeed, among the 146 reviewed papers only 10 DSSs were identified to support the critical decisions that must be taken during PHM implementation. Specifically, 3 DSSs were found to support the identification of the CU, 5 DSSs were developed to identify the most appropriate maintenance strategy to maintain the CU, and only 2 DSSs were built to select the most appropriate DA technologies to enable PHM services. Furthermore, most of the selected DSSs belong to the ‘Pre PHM phase’ while only 2 DSSs belong to the ‘PHM-implementation phase’. Among the DSSs, 80 decision-making criteria were identified and grouped in seven macro-areas, namely, economic, strategical, technological, reliability, operational, human resource, and safety. It is worth noting that the ‘criteria’ aim at assessing the stakeholders' needs, enabling their involvement in the decision-making process. A cross-analysis between the criteria macro-area and scope of the DSS showed that criteria belonging to the area of reliability, economic and strategical are the most used to identify the CU. While criteria belonging to the economic, strategic and operational areas are the most used to select the most appropriate maintenance strategies. Finally, technological criteria were the most considered for the choice of the DA technologies. Regarding the methods used to evaluate the alternative against the criteria, MCDM methods are the most used, but simulation, ML algorithms and mathematical models were also used.

In summary, some open issues emerged from the SLR, as reported in the following:

• Systematic approaches to guide and support the practitioners during the implementation phase of PHM solutions are lacking, as only 2 DSSs were found belonging to the PHM implementation phase.

• DSSs to support practitioners in the implementation of the DM, SD, HA and Prognosis Assessment phases are lacking.

• Iterative procedures such as the ‘try and error procedure’ are unavoidable, given the close dependency between methods and case studies, especially when data-driven methods are considered. In this regard, it is worth noting that the two DSSs developed to identify the most appropriate DA technologies to enable PHM solutions neglected their integration with the methods that will be used within the PHM service to analyse the collected data.

• Not all companies are interested in achieving the highest level of functionality in PHM. Therefore, a DSS could be developed to support the decision on which PHM's level of functionality is the most suitable for a company according to its characteristics (e.g. goals, skills and technological readiness, type of production) and requirements.

As an initial contribution to fill the first three highlighted gaps, the authors proposed a conceptual framework (Figure 10) that considers the entire process leading to the implementation of one or all of the PHM services. The framework starts from the identification of the CU, then allows for the identification of the most appropriate maintenance strategy to maintain it and enables the identification of the features and algorithms to perform the tasks of PHM services. For each phase and step, the challenges and the relative decision to be addressed are provided. In addition, the relationships among all the parties involved in the decision-making process were identified. Two main iterative procedures were identified in the framework, which deal with the identification of the features and algorithms to perform the task of PHM services and the selection of the DA technologies and pre-processing methods. In these phases, indeed, it is crucial to have data to test solutions in the field to assess their performances.

A future research step of this work includes, first of all, the application of the conceptual framework in several case studies in order to highlight its strengths and weaknesses, thus improving its completeness. Then, a further step can concern the development of maturity models and DSSs related to one or more of the challenges highlighted in the steps of the conceptual framework. In this vision, the conceptual framework will provide the roadmap for PHM implementation, listing all the necessary steps, while the maturity models will assess the level of maturity (readiness) to face off with each challenge and the DSSs will support the identification of the most appropriate alternative to be taken to address each challenge.

AUTHOR CONTRIBUTIONS

Raffaele Abbate: Conceptualization; data curation; formal analysis; investigation; methodology; resources; software; validation; visualization; writing—original draft; writing—review and editing. Chiara Franciosi: Conceptualization; formal analysis; investigation; methodology; resources; validation; writing—review and editing. Alexandre Voisin: Conceptualization; formal analysis; investigation; methodology; resources; supervision; validation; writing—review and editing. Marcello Fera: Conceptualization; formal analysis; methodology; resources; supervision; validation; writing—review and editing.

ACKNOWLEDGEMENTS

None.

Open access publishing facilitated by Universita degli Studi della Campania Luigi Vanvitelli, as part of the Wiley - CRUI-CARE agreement.

CONFLICT OF INTEREST STATEMENT

The authors have no competing interests to declare that are relevant to the content of this article.

DATA AVAILABILITY STATEMENT

All data generated or analysed during this study are properly reported and cited in the article.

References

Voisin, A., et al.: Generic prognosis model for proactive maintenance decision support: application to pre‐industrial e‐maintenance test bed. J. Intell. Manuf. 21(2), 177–193 (2010). [DOI: https://dx.doi.org/10.1007/s10845-008-0196-z]