Design and realization of compressor data abnormality safety monitoring and inducement traceability expert system

Abstract

Centrifugal compressors are widely used in the oil and natural gas industry for gas compression, reinjection, and transportation. Fault diagnosis and identification of centrifugal compressors are crucial. To promptly monitor abnormal changes in compressor data and trace the causes leading to these data anomalies, this paper proposes a security monitoring and root cause tracing method for compressor data anomalies. Additionally, it presents an intelligent system design method for fault tracing in compressors and localization of faults from different sources. This method starts from petrochemical big data and consists of three parts: fault dynamic knowledge graph construction, instrument data sliding fault-tolerant filtering, and the fusion and reasoning of fault dynamic knowledge graph and instrument data variation monitoring. The results show that this method effectively overcomes the problems of false alarms and missed alarms based on fixed threshold alarm methods, and achieves 100% classification of two types of faults: non starting of the drive machine and low oil pressure by constructing a PCA (Principal Component Analysis)—SPE (Square Prediction Error)—CNN (Convolutional Neural Network) classifier. Combined with dynamic knowledge graph and NLP (Natural Language Processing) inference, it achieves good diagnostic results.

1. Introduction

In actual production activities, the occurrence of system failures is inevitable due to internal factors (such as part damage) or environmental factors (such as ventilation and sealing). The compressor system consists of a lubrication system, cooling system, power system, and other components. These complex systems, composed of many interconnected units, can significantly impact normal operations when a failure occurs and may further lead to considerable losses that are difficult to estimate. Therefore, it is important to study the fault diagnosis methods of compressor systems. Expert systems usually consist of an inference engine, a knowledge base, an interpretation facility, and a user interface. In practice, most of the information required to create rules comes from expert experience or maintenance manuals. However, due to the lack of comprehensive experience and information, this presents some inherent constraints on the development of the rule base. Limited knowledge acquisition is a bottleneck for traditional rule-based expert systems, restricting their scope of application. Additionally, traditional rule-based fault diagnosis expert systems lack flexibility.

In the petrochemical production process, instrumentation is not only an important basis for understanding the state of equipment, operating conditions, material quality and emissions, but also an important basis for monitoring the operation of the entire device faults and potential accidents, and discovering abnormal working conditions and disaster risks. In recent years, data-driven methods based on deep learning [1–6] have made significant progress in fault diagnosis of complex systems due to their ability to efficiently and economically collect large amounts of data. For example, the method of combining Z numbers with fuzzy logic improves the accuracy results of expert systems applying risk analysis [7]. Combined with intuitionistic fuzzy set theory and expert-inspired frameworks, TFTs of dynamic systems with uncertain data can be quantitatively analyzed [8]. A methodology for constructing a knowledge base for an operational guidance expert system based on HAZOP analysis and accident analysis [9]. Artificial neural network based fault diagnosis system for detecting combustion faults, wear faults, fuel injection faults [10]. Analyzing the root cause of faults using the degree of change and the average variable threshold limit [11]. Combining multi-scale principal component analysis with signature based directed graph method for process monitoring and fault diagnosis [12]. Application of Threshold Recursive Graph and Texture Analysis in Diagnosing Static Friction in Process Control Circuits [13]. A Multi Kernel Support Vector Machine (MK-SVM) algorithm is proposed for diagnosing simultaneous faults in distillation towers [14]. Combining multivariate index weighted moving average (MEWMA) monitoring scheme with PCA modeling to improve anomaly detection performance [15]. Single class support vector machine (KPCA-OCSVM) based on kernel principal component analysis and various kernels to learn an anomaly free training set, and then classify the test set [16]. An engineering based approach that can be improved upon traditional natural language processing to construct a domain knowledge graph based on the ontology of petroleum exploration and development [17]. Addressing the issues of label dependency, sparse interaction data, and irrelevant knowledge in community-based software expert recommendations, paper propose an Expertise Preference-Aware Network model for Software Expert Recommendation (EPAN-SERec) with knowledge graph [18]. Adopting knowledge graph best development practices to achieve scalable and optimized manufacturing processes [19] etc. However, because there are many types of equipment instruments in the petrochemical process industry, there are many hidden dangers that affect compressor safety, and there are many signs and manifestations of compressor failure, there are mostly complex interlocking relationships, feedback relationships, or causal relationships between instrument data and failures. The effects of faults are intertwined and change dynamically over time, making it difficult for conventional fault tree analysis, cause-and-effect tree analysis and simple hierarchical network topology diagrams to fully reflect their complex multiple couplings and mutual effects, which increases the complexity of the compressor and Difficulty of process safety risk analysis.

In the production process of the existing large-scale petrochemical plant, whether the monitoring platform adopts the DCS mode or the FCS mode, due to the above-mentioned complexity and technical difficulty, the overrun alarm based on the fixed threshold or the overrun grading alarm method based on the multi-level threshold that is most widely used in the field of instrument data anomaly monitoring is not only difficult to set accurately, but also the monitoring method itself is difficult to avoid the limitations such as false positives, false alarms or missed judgments and missed alarms. Especially in the early stage of the occurrence of the abnormality or when the amplitude of the abnormality is less than 3 times the variance of the measurement error, it is difficult to detect the early fault signs and the weak fault is difficult to detect.

To this end, this paper proposes a compressor data security monitoring and trigger source method based on a dynamic knowledge graph to break through the weak links in existing petrochemical instrument safety monitoring technology and traditional expert systems, and solve the problem of difficulty in separating instrument abnormalities from equipment failures. It is difficult to catch small faults in time and the problem of frequent false alarms and missed alarms in the monitoring process is of great significance to ensuring petrochemical safety.

In summary, the main contributions of this study are as follows:

Based on the sample data obtained by sliding window processing, a sliding window three-segment mid-section fault-tolerant filtering method is implemented to implement a safe and reliable sliding window fault-tolerant filter, as well as a residual generation and data mutation monitoring algorithm based on fault-tolerant filtering.(2)Based on the fault data obtained by sliding window processing, a deep learning model for system fault classification combining PCA-SPE and CNN is proposed to realize 100% classification of compressor faults.(3)Combination of model-based classification results and expert system realizes dynamic knowledge graph-based instrumentation data security monitoring and causative origins.

2. Construction of a residual generation and data variance monitoring algorithm based on fault-tolerant filtering

Accurate extraction of intrinsic variations in instrument data is an important technical aspect of accurate monitoring of data variations and early detection of small variations. For the petrochemical instrument sampling data sequence {Y(t_i),i = 1,2,…,N}, this paper proposes and implements a safe and reliable sliding window fault-tolerant filter, as well as a residual generation and data mutation monitoring algorithm based on fault-tolerant filtering:

Design a sliding window three-segment mid-section fault-tolerant filtering method: Specifically, construct a sliding window with window width s = 3k+2, where k is an adjustable parameter (the default value is 5), For the instrument sampling data fragments in the half-open interval time window [t_i,t_i+s), X_i = {Y(t_j),t_j∈D_i} are sorted in order from small to large, The sorted data fragment is , and the mean filter and filter residuals of k points in the middle of are calculated.(1)The above sliding filter is strongly tolerant to abnormal data Outliers and localized spots. Briefly note that the residual sequence formed by the residuals obtained by the above sliding window fault-tolerant filter is , and the threshold parameter C_i is constructed by using a recursive algorithm(2)where indicates that two permutations are arbitrarily selected at i+1 time points, and their weights decrease with the increase of i when the statistics of multiple points are accumulated. j is the weight coefficient, which means that the later the moment j has a greater effect on the threshold value. is a piecewise function used to adjust the relationship between and C_j−1, as shown in Formula (3), represents the filtered residual result of the low j-th calculation in the residual sequence, and C_j−1 represents the threshold value of the j-1st calculation.(3)And based on this, monitor and diagnose instrument data changes:

When , the data is judged to be normal; When , the data is suspected to be abnormal, when , data reports a high-level abnormality, and when , data reports a very high-level abnormality. When , data is suspected to be abnormal; when , data reports a low-level abnormality; when , data reports a very low-level abnormality. The complete implementation process is shown in Fig 1.

[Figure omitted. See PDF.]

3. Construction of a deep learning based system fault classifier

The framework of the system fault classifier based on deep learning proposed in this paper is composed of two parts: PCA-SPE monitoring and CNN classifier. Due to the significant differences in the data selected in this article, it is difficult to distinguish between known edge normal data and abnormal data. Therefore, first, PCA-SPE is used to screen the data for fault samples and obtain corresponding fault samples. Then, combined with the constructed CNN classifier, the fault classification is achieved. The detailed construction process of the corresponding algorithm is as follows.

3.1 PCA-SPE-based fault monitoring

Firstly, according to the offline data, the data H is normal, H is H(t₁)…H(t₁₅₀₀), and H is 12, and the dimension is the same as that of the monitoring sample. Then the matrix H is subjected to error tolerant filtering [20], and the normalization process is carried out, and the data of each column of matrix H are normalized to the data with a mean value of 0 and a root mean square of 1 to obtain the matrix .

The samples used for monitoring are obtained continuously from {Y(t_i)i = 1,2,…,N} through a sliding window. Y(t_i) is 12-dimensional data. The data segment with a sliding window of 100 is recorded as Y[t_k∶t_k+100]. The data Y[t_k∶t_k+100] is normalized, the mean, and variance are the mean and variance of the training data, and each column of data in the matrix Y is normalized to the data with a mean value of 0 and a root mean square of 1 to obtain matrix .

1) Calculate the covariance matrix of the normalized training data and the monitoring data respectively:(4)(5)where n is the number of the training samples, g is the number of the monitoring samples, and the resulting covariance matrix is a characteristic 12*12-dimensional matrix.

2) The eigenvalues are sorted from large to small, and the top k features with a sum of more than 85% eigenvalues are selected for PCA dimensionality reduction.

(6)

The first k eigenvalues from large to small form a diagonal matrix S_k*k, and k corresponding eigenvectors form a matrix P_m*k.

The training set diagonal matrix and the corresponding eigenvectors form the matrix:(7)(8)The diagonal matrix of the monitoring set and the corresponding feature vector form a matrix:(9)(10)For , after dimensionality reduction, the number of samples remains unchanged, but the number of features becomes K. After dimensionality reduction, it is:(11)(12)The formula for calculating the matrix of obtained by reconstructing and is as follows:(13)(14)where n is the number of training samples, g is the number of monitoring samples, m is the number of corresponding key variables, and k is the number of corresponding variables after dimensionality reduction.

The training sample used in this paper is 1500 sampled values and the monitoring sample window is a large sample of 100 data and the two samples are independent of each other and conform to a normal distribution and conforms to a normal distribution with zero mean. We used a PCA-based process monitoring methodology with the squared prediction error (SPE) statistic [21, 22] to make the determination.

The control limits for the SPE statistic were proposed by Jackson and Mud Holkar in 1979. conforms to the multivariate normal distribution, and the SPE statistic [23] is established for , that is, for (i = 1,2… g) Construct statistic Q:(15)where is the row vector of row i of , I_m*m is the identity matrix, and is the matrix of k eigenvectors corresponding to k eigenvalues from large to small from (1) and (2).

Since obeys a normal distribution with zero mean, and the sample sampling points are independent of each other, it has been proved in the literature [24] that Eq (13) is approximately in line with the normal distribution.(16)(17)(18)Therefore, the approximate mean of c is a normal distribution of 0, which is calculated as follows:(19)Then the control limit of Q is:(20)c_α is the confidence limit of the standard normal distribution, c_α is 0.996 confidence, λ is the eigenvalue of the corresponding covariance matrix R_ta, and h₀ is the adjustment coefficient.

1) According to Eqs (15) and (20), the SPE statistic Q and the statistical limit Q_α.

2) Fault determination

If the system is functioning normally, the Q-value of the sample should meet Q <Q_α, otherwise it is considered to be faulty.

If Q<Q_α, the corresponding fault data .

3.2 CNN module

The CNN module consists of a CNN layer, a pooling layer, and a fully connected layer, as shown in Fig 2, which are described in detail below.

[Figure omitted. See PDF.]

The CNN layer consists of two operations, the convolution operation and the activation operation. The input to the convolution operation is a 3D data Z∈R^w×h×c, where w, h, and c represent its width, height, and number of channels, respectively. In particular, when c = 1, it is reduced to 2D data as input to the proposed model ∈R^w×h×c. Specifically, the horizontal axis of two-dimensional data represents the feature dimension, and the vertical axis represents the time dimension. The axis scale is characterized by the number of attributes and the width of the sliding window, respectively. In order to clearly illustrate the convolution operation, we first define four hyperparameters, namely the number of convolution kernels K, the size of each convolution kernel F, the step size of the convolution S and the number of zeros P. Then, for the convolution operation of each convolution kernel, the data ∈R^w×h×c is filled with zeros to ensure that the shape of the output and input are the same, resulting in the transformed data ∈R^{(w+p)×(h+p)×c}. After that, the slicing operation is:(21)where t_w∶ t_w+ F represents the operation of extracting the subset located between row t_w and row t_w+ F, t_h ∶ t_h+ F represents the operation of extracting the subset located between column t_h and column t_h+F, and ∶ represents all the data in the direction of the cutting channel. t_w and t_h start at 1 and change with the number of steps S. According to (t_w,t_h), data can be extracted from . Let the convolution kernel be denoted as kernel∈R^F×c, then is defined as kernel and multiplied element by element, and the formula is as follows:(22)Therefore, can be obtained through the following methods:(23)where denotes the element of , and can be obtained from ordered by the size of t_w, t_h. It can be noted that(24)For each convolutional kernel, a channel such as B is obtained, and then the above operation is repeated on K convolutional kernels to get a new 3D data S∈R^w*×h*×K.After the convolution operation, the activation operation is essential. It enables the network to obtain nonlinear expressions of the input to enhance representation capabilities and make the learned features more separable. Use the Rectifier Linear Element, which is a widely used linear element, which can activate each element of the S in the following ways:(25)where S[i,j,k] is the element of S and A[i,j,k] is the element of A∈R^w*×h*×K, the data obtained after the activation operation.

Dropout regularization:

The Dropout layer reduces model overfitting by randomly zeroing out a portion of the elements of the input data. The formula for the operation of Dropout is as follows:(26)where Z is the input data, p is the discard rate, and M is the mask matrix generated from the Bernoulli distribution.

Fully Connected Layer (Dense):

The fully connected layer maps the output of the globally averaged pooling layer to the probability of the output class through a weight matrix and bias vectors. The operation formula for the fully connected layer is as follows:(27)where is the output of the global average pooling layer, W is the weight matrix, b is the bias vector, and softmax (⋅) is the softmax function to normalize the raw scores into category probabilities.

4. Fault diagnosis based on expert system

In the actual situation of compressor online monitoring, because the parameters are outside the allowable setting range, a large amount of information has surfaced before the alarm, which is identified by PCA-SPE-CNN in this paper. This study explores the correlation of deviating parameters, equipment components, and failure modes based on the data collected in this paper, which is derived from published literature and compressor data from PetroChina West-East Gas Pipeline Co., Ltd. at Gaoling Station. Regarding the diagnosis and traceability issues of compressors, a network-based intelligent diagnostic program was developed to find abnormalities by processing monitoring data. Once the early failure mode is identified, the program will use the knowledge base to try to infer the possible failure and achieve fault traceability and location.

4.1 Construction of knowledge graph

Constructing a dynamic knowledge graph is the premise and foundation for identifying the causes of instrument variation. This paper proposes a bottom-up method for constructing a dynamic knowledge graph of petrochemical instrument data variations: based on petrochemical big data, a petrochemical data flow structure classifier is established to extract structured data, semi-structured data and unstructured data from petrochemical big data. structured data; Perform preprocessing such as cleaning, deduplication, and formatting on the data stream to form corresponding expandable data sets of structured, semi-structured, and unstructured data; Extract knowledge from the data set to form an entity set E of materials, working conditions, modalities, and production processes related to petrochemical instruments, a relationship set R that reflects process relationships, structural relationships, and association relationships, as well as a relationship set R that reflects the quantification degree of process conditions. An attribute set A such as the object’s normality and so on; Using the three sets of E, R, and A, the triplet {(ě,§,ǔ): ě,ǔ∈E,§∈R} and {(ě, ζ, ǔ): ě,ǔ∈E; ζ ∈A}, and the entity disambiguation, co-signifier dissolution and knowledge merging are carried out on the above-mentioned triplet sets to achieve knowledge fusion; Define and construct the concepts, terms and relationships in the petrochemical instrument fault knowledge map to form a fault knowledge base structure for petrochemical instruments; Expand existing knowledge about instrument faults and obtain new knowledge through information extraction, knowledge fusion and knowledge reasoning; A graph database is used to store the obtained knowledge graph; While continuously receiving the instrument data and process data sent by the petrochemical DCS system, combined with the incremental data information, the iterative optimization of the knowledge graph is realized, the iterative optimization in this article is mainly achieved by constructing a PCA-SPE-CNN classifier to classify faults in real-time monitored instrument data, thereby realizing inference iteration, and combining expert knowledge in the knowledge graph to update and iterate knowledge.The dynamic knowledge graph of the in-service process of the petrochemical instrument is completed, the technical process is shown in Fig 3, and the construction schematic diagram is shown in Figs 4 and 5.

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

4.2 Construction of a compressor failure cause traceability system

Identifying the source of variations in instrument data is critical to preventing fail-safe risks. This paper proposes a multi-induced anomaly traceability method for petrochemical instrument data based on improved CNN classifier, deep integration of natural language processing and dynamic knowledge graph, which uses dynamic knowledge graph and integrates intelligent information processing technologies such as cognitive computing, feature extraction, knowledge representation, knowledge reasoning, and information retrieval, as well as convolutional neural network and natural language processing model to realize the source of inducement of instrument data variation. Specifically, the method of inducing data mutation sources consists of the following three parts:

1) Inference engine design based on natural language processing NLP. Based on the above dynamic knowledge graph, a collection of heterogeneous knowledge graph triples describing the fault sources of petrochemical equipment and compressors, the heterogeneous patterns of instrumentation data, and the relationship between the fault sources and the data heterogeneity are constructed to form the graph dataset; A graph dataset is used as the basis for modeling instrumentation data variability and compressor faults in combination with an improved CNN classifier, and a priori logic rules are introduced in the modeling process; Based on the historical data of petrochemical instruments and the dynamic knowledge graph of faults, the summary module of the relationship between given entities and the inference module of state inference based on this are the inference modules. Relational NLP is constructed as an editor to realize the specific classification learning of the relationship between faulty entities and neighboring entities in the knowledge graph, and combined with the set question words, the matching is more efficient to obtain the corresponding classification representation, and then the NLP inference engine based on dynamic knowledge graph is constructed.

2) Inference engine design for multi-mode fusion. In this paper, an inference method based on multi-mode fusion of inference layer is designed to realize the source and location of instrument and equipment faults. Specifically, a dynamic knowledge graph-based NLP inference engine is used to construct a premise set and a consequence set for failure risk; The historical data and historical fault cases of instruments and equipment are used for model training to obtain the fuzzy correlation of the linkage between the premise set and the consequence set; Based on the equipment data, a supervised machine learning method is used to establish a fault detection and diagnosis model based on resolved redundancy, from which fault types are determined; Absorbing the technical kernel of fault inference engine based on NLP computing, we establish a natural language inference engine based on combining semantic relevance and premise set for fault tracing, and integrate it with machine learning model inference engine and NLP computing inference engine at the inference layer to realize fault tracing and localization. The above reasoning layer fusion diagnostic reasoning method can effectively avoid the limitations of conventional forward reasoning and reverse reasoning which are difficult to be used in situations such as causal inter-network correlation, and effectively overcome the influence of multiple coupling mutual cause and effect relationships between detection devices as well as between detection devices and the inspected object and the controlled process, and realize the traceability and localization of the instrumentation data and the equipment faults.

3) Design of multi-induced anomaly traceability system based on the fusion of instrument data and dynamic knowledge graph. Starting from petrochemical big data, through the construction of dynamic knowledge graph of instrument faults, the sliding fault-tolerant filtering of instrument data, and the integration and reasoning of dynamic knowledge graph of instrument faults and instrument data anomaly monitoring, a safety monitoring and inducement traceability system for instrument data anomalies is established, and an intelligent support platform is formed to serve the source of instrument safety detection and inducements.

5. Analysis of experimental results

This section mainly uses the data provided by China Petroleum West East Gas Pipeline Co., Ltd. for the Gaoling Station compressor to analyze the experimental data of motor speed, motor power, motor stator temperature, and lubricating oil main pressure. Using the above data for the analysis and validation of the algorithm, the detailed settings and analysis validation of the corresponding algorithm dataset are described below. The residual generation and data anomaly monitoring algorithm based on fault-tolerant filtering is mainly used for detecting the stationarity of each dimensional feature.The construction of classifier algorithms mainly considers the influence of twelve dimensional features on second-class faults.The construction of knowledge base based on knowledge graph is mainly used for fault tracing, and the specific analysis is as follows.

5.1 Residual generation based on fault-tolerant filtering and monitoring process analysis of data anomaly detection algorithm

Taking the structured data collected by the instrument, as well as the feature change data extracted from the semi-structured data and the unstructured data as the object, the sliding window fault tolerance rate wave algorithm (2) and the filter residual generation algorithm (3) established by the invention are used to realize the data-driven monitoring and diagnosis of the abnormal change of each feature state. He test dataset consisted of 428 samples. Each sample has 12 characteristics, the test sample points are 428*12 = 5136, and the data sampling frequency is 1 time/min. The schematic diagram of data sampling, filtering, and filtering residuals in the process of data anomaly monitoring is shown in Figs 6–8. The monitoring of each feature of the data is shown in Figs 9–11.

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

5.2 Introduction to classifier data settings

For each fault and normal state, the compressor status is reflected in a training data set and a test data set. Each training data set contains 1500 samples, and each test data set contains 228 samples. Each sample has 12 features, the original training sample points are 1500 × 12 = 18000, and the original test sample points are 228 × 12 = 2736. The data sampling frequency is 1 time/min. For the training samples containing faulty samples and normal samples, the samples are randomly sorted and normalized in order to ensure good learning results.

5.3 Analysis of data reasons

The motor fault and its related indicators are selected for verification. After the motor fault, the faulty motor cannot maintain a relatively constant speed and power as before the fault occurred, and the motor stator temperature also gradually changes with the motor fault. They can serve as key indicators of motor failure. The motor was in a shutdown state before 10:00 on 2022/10/14, and ran normally after 10:50. An abnormality occurred around 19:30, and returned to normal around 5:30 on 10/15. The temperature of the motor stator remains unchanged when the motor is stopped, rises after the motor operates normally, drops after an abnormality occurs in the motor, and recovers after the motor recovers. The motor speed, power diagram and motor stator temperature are shown in Figs 12 and 13. After the oil pressure failure, the lubricating oil main pipe pressure remained normal before 2:54 on 2022/10/23. An abnormality occurred after 2:54, and the main pipe pressure changed as follows As shown in Fig 14.

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

5.4 Network architecture setup

The input shape of the CNN network is batch 50 on the vertical axis, the matrix is feature 12 on the horizontal axis, the convolution kernel is 1*3, the number of convolution kernels is 16, 16, 64, 64, 256, 256, the activation function is the relu function, and the pooling layer is the maximum pooling and average pooling. The network structure is a stack of three consecutive convolutional layers and a maximum pooling, and finally an average pooling, regularization and fully connected layers are used for fault classification. The number of training rounds is 25 and the training time is 6 seconds, which meets the monitoring needs.

5.4.1 Parameter tuning.

The selection of convolutional layers in this article is based on the structure and complexity of the data, and is continuously adjusted through experiments. For the situation where 12-dimensional features affect two faults, multi-layer convolution needs to be used to extract features at different levels. Through the combination of convolutional layers and pooling layers, high-level features can be effectively extracted. This paper uses two convolutions combined with one pooling to improve the generalization ability of the model. In addition, regularization techniques are combined to prevent overfitting, and triple stacking is performed. Small kernels can capture local features, and when multiple small kernels are stacked and used, global features can be captured at a deeper level. Large kernels can directly capture a wider range of features, but require greater calculations. This paper prefers to use small kernels to deepen the layers to achieve the same effect. This method can effectively capture local features and has high computational efficiency. Many successful networks use small cores and gradually deepen the network, which is often an effective strategy. A lower initial learning rate of 0.0002 was chosen and the Adam optimizer was used to adapt to the learning rate. For the Dropout rate, 0.3 is a moderate Dropout rate that prevents overfitting while maintaining good model performance, and the number of filters is given an initial value space. Finally, grid search technology is used to optimize parameters.

5.5 Comparative analysis with other algorithms

The PCA-SPE-CNN, PCA-SPE-KNN, PCA-SPE-SVM, CNN classifiers have been validated with data and their performance has been evaluated. The accuracy, precision, recall, f1 score, and AUC-ROC and ROC curves were comprehensively evaluated by cross-validation with k-fold. The ROC curve is derived from the signal detection theory and represents the relationship between the true positive rate (TPR) and the false positive rate (FPR) of the classifier (TPR on the x-axis and FPR on the y-axis), and the validation results and evaluation results are described below.

Four additional important terms are also used in the calculation of various assessment parameters, namely true positive (TP), true negative (TN), false positive (FP), false negative (NP). True Positive Rate (TPR): The number of true positives (TP) divided by the number of samples that are true positive (TP + NP). False Positive Rate (FPR): The number of false positives (FP) divided by the number of samples that are truly negative (FP + TN). TP: These are positive datasets that are accurately labeled by classification methods. TN: These are negative datasets that are accurately labeled by classification methods. FP: These are negative datasets that have been incorrectly labeled as positive datasets by the classification method. NP: These refer to positive datasets that have been mislabeled as negative datasets by the classification method.

5.5.1 Classification accuracy and performance analysis of PCA-SPE-CNN classifier.

First, offline normal data is used for training, then a sliding window is used to perform fault monitoring on the monitoring data, and finally the monitored data is screened for fault data. According to Eqs (15) and (20), the SPE statistic Q and the statistical limit Q_α of the monitoring data are calculated respectively, and the fault monitoring is carried out by comparing the magnitude of the Q and Q_α values. The monitoring results are shown in Fig 15. Those that are higher than the limit value are fault data. The corresponding normal data filtering part is shown in Fig 16.

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

After training the PCA-SPE-CNN algorithm, the confusion matrix was constructed. The improved algorithm achieved 100% classification accuracy. The accuracy of the training set and verification set and the classification results of the model are shown in Fig 17. In addition, the model passed an evaluation based on the ROC curve. The ROC curve is derived from the theory of signal detection and represents the relationship between the true positive rate (TPR) and the false positive rate (FPR) of the classifier (TPR on the x-axis and FPR on the y-axis). This curve reflects the model performance under different thresholds, and the results are shown in Fig 18. The model underwent k-fold cross validation as shown in Fig 19.

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

In order to comprehensively evaluate the performance of the model, analysis was conducted on accuracy, precision, recall, F1 score, and AUC-ROC indicators. The results are shown in Table 1. And the Bootstrap method was used to estimate the performance metrics, so as to improve the robustness of the analysis.

[Figure omitted. See PDF.]

By resampling 1000 times, ensure that the model performs more robustly under different validation set samples. Finally, calculate the mean and 95% confidence interval of the performance indicators to measure the fluctuation of the model’s performance under different sampling conditions. The calculated result is Mean AUC: 1.000,95% CI: [1.000,1.000]. Mean Accuracy: 1.00, 95% CI: [1.00, 1.000].Mean AUC: represents the average AUC value of the model among all Bootstrap resamples, with a value closer to 1 indicating better discriminative ability of the model. 95% CI for AUC: The 95% confidence interval of AUC, indicating that the AUC value of the model will fall within this interval with a 95% probability under different resampling conditions. If the interval is small, it indicates that the model performance is relatively stable; If the interval is large, it indicates that the model exhibits significant fluctuations under different data. Mean Accuracy: Refers to the average accuracy of the model’s predictions across all resamples. 95% CI for Accuracy: The 95% confidence interval for accuracy, representing the fluctuation range of the model’s accuracy under different resampling conditions.

5.5.2 Classification accuracy and performance analysis of PCA-SPE-KNN, PCA-SPE-SVM, CNN classifiers, and expert system diagnostic results.

Due to significant differences in the selected data, it is difficult to distinguish between known edge normal data and abnormal data. After training PCA-SPE-KNN, PCA-SPE-SVM, and CNN algorithms, a confusion matrix was constructed, and complex models were compared based on accuracy, precision, recall, and F1 score. According to the confusion matrix, the classification performance is far inferior to the improved PCA-SPE-CNN model, and the accuracy, recall, and F1 score of the three compared models are not as good as the improved model. Figs 20–22 show their classification results, and Table 2 shows their metric evaluations. By comparing the above, the best solution can be achieved.

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

Expert system diagnosis results are shown in Fig 23 below.

[Figure omitted. See PDF.]

6. Conclusions

This paper proposes and implements a safe and reliable sliding window fault-tolerant filter, as well as a residual generation and data mutation monitoring algorithm based on fault-tolerant filtering. The performance of the proposed PCA-SPE-CNN classifier is tested based on process data from an industrial centrifugal compressor. This classifier is also combined with an expert system based on knowledge graph and NLP inference engine to realize the causal traceability of the expert system. This method greatly improves the following problems:

First, the above monitoring method can quickly realize early symptom detection, small-amplitude variance identification and fault isolation, effectively overcoming the existing DCS or FCS system instrumentation data monitoring commonly used based on a fixed threshold alarm method of false alarms and missed alarms, as well as difficult to realize the limitations of within-threshold variance monitoring and so on; The second is to improve the accuracy of fault classification; Third, it improves the problems of difficulty in knowledge base maintenance and the lack of flexibility of traditional rule-based fault diagnosis expert systems. The system is able to learn from actual operational data and adapt adaptively to changing circumstances. This system can effectively avoid the limitations of conventional forward reasoning and reverse reasoning that are difficult to apply to situations such as network correlation between causes and effects, and effectively overcome the multiple coupling interactions between detection devices, as well as between detection devices, the detected object, and the controlled process. The influence of the mutual relationship enables the traceability and location of instrument data and equipment faults.

In future research, more advanced machine learning algorithms can be explored to improve the real-time and diversity of fault monitoring; Examine the applicability and effectiveness of this method in other industries such as manufacturing, and promote the widespread use of the technology.

Supporting information

S1 Data.

https://doi.org/10.1371/journal.pone.0315917.s001

(CSV)

S2 Data.

https://doi.org/10.1371/journal.pone.0315917.s002

(CSV)

S3 Data.

https://doi.org/10.1371/journal.pone.0315917.s003

(CSV)

S4 Data.

https://doi.org/10.1371/journal.pone.0315917.s004

(CSV)

References

1. 1. LEI Y, JIA F, LIN J, et al. An Intelligent Fault Diagnosis Method Using Unsupervised Feature Learning Towards Mechanical Big Data [J]. IEEE Transactions on Industrial Electronics, 2016, 63: 3137–47.

* View Article

* Google Scholar

2. 2. LI C, LI C, TAO Y, et al. Improving forecasting accuracy of daily enterprise electricity consumption using a random forest based on ensemble empirical mode decomposition [J]. Energy, 2018.

* View Article

* Google Scholar

3. 3. LI X , ZHANG W, DING Q. Understanding and improving deep learning-based rolling bearing fault diagnosis with attention mechanism [J]. Signal Process, 2019, 161: 136–54.

* View Article

* Google Scholar

4. 4. LIU C-L , HSAIO W, TU Y-C. Time Series Classification With Multivariate Convolutional Neural Network [J]. IEEE Transactions on Industrial Electronics, 2019, 66: 4788–97.

* View Article

* Google Scholar

5. 5. NATH A G, UDMALE S S, SINGH S K. Role of artificial intelligence in rotor fault diagnosis: a comprehensive review [J]. Artificial Intelligence Review, 2020, 54: 2609–68.

* View Article

* Google Scholar

6. 6. TAO X , GONG X, ZHANG X Y, et al. Deep Learning for Unsupervised Anomaly Localization in Industrial Images: A Survey [J]. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 1–21.

* View Article

* Google Scholar

7. 7. YAZDI M, HAFEZI P, ABBASSI R. A methodology for enhancing the reliability of expert system applications in probabilistic risk assessment [J]. Journal of Loss Prevention in the Process Industries, 2019.

* View Article

* Google Scholar

8. 8. KABIR S, GEOK T K, KUMAR M, et al. A Method for Temporal Fault Tree Analysis Using Intuitionistic Fuzzy Set and Expert Elicitation [J]. IEEE Access, 2020, 8: 980–96.

* View Article

* Google Scholar

9. 9. WANG F, GAO J. A novel knowledge database construction method for operation guidance expert system based on HAZOP analysis and accident analysis [J]. Journal of Loss Prevention in The Process Industries, 2012, 25: 905–15.

* View Article

* Google Scholar

10. 10. BASURKO O C, URIONDO Z. Condition-Based Maintenance for medium speed diesel engines used in vessels in operation [J]. Applied Thermal Engineering, 2015, 80: 404–12.

* View Article

* Google Scholar

11. 11. SHAHID M, ZABIRI H, TAQVI S A A, et al. Fault root cause analysis using degree of change and mean variable threshold limit in non-linear dynamic distillation column [J]. Process Safety and Environmental Protection, 2024.

* View Article

* Google Scholar

12. 12. ALI H, MAULUD A S, ZABIRI H, et al. Multiscale Principal Component Analysis-Signed Directed Graph Based Process Monitoring and Fault Diagnosis [J]. ACS Omega, 2022, 7: 9496–512.

* View Article

* Google Scholar

13. 13. KOK T L, ALDRICH C, ZABIRI H, et al. Application of unthresholded recurrence plots and texture analysis for industrial loops with faulty valves [J]. Soft Computing, 2022, 26: 10477–92.

* View Article

* Google Scholar

14. 14. TAQVI S A A, ZABIRI H, UDDIN F, et al. Simultaneous fault diagnosis based on multiple kernel support vector machine in nonlinear dynamic distillation column [J]. Energy Science & Engineering, 2022, 10: 814–39.

* View Article

* Google Scholar

15. 15. HARROU F, KADRI F, KHADRAOUI S, et al. Ozone measurements monitoring using data-based approach [J]. Process Safety and Environmental Protection, 2016, 100: 220–31.

* View Article

* Google Scholar

16. 16. CHENG T, DAIRI A, HARROU F, et al. Monitoring Influent Conditions of Wastewater Treatment Plants by Nonlinear Data-Based Techniques [J]. IEEE Access, 2019, 7: 108827–37.

* View Article

* Google Scholar

17. 17. ZHOU X-G, GONG R, SHI F, et al. PetroKG: Construction and Application of Knowledge Graph in Upstream Area of PetroChina [J]. Journal of Computer Science and Technology, 2020, 35: 368–78.

* View Article

* Google Scholar

18. 18. TANG M, WU D, ZHANG S, et al. EPAN-SERec: Expertise preference-aware networks for software expert recommendations with knowledge graph [J]. Expert Syst Appl, 2024, 244: 122985.

* View Article

* Google Scholar

19. 19. JAWAD M S, DHAWALE C A, AZHAR A, et al. Adoption of knowledge-graph best development practices for scalable and optimized manufacturing processes [J]. MethodsX, 2023, 10.

* View Article

* Google Scholar

20. 20. Shaolin Hu, Na Fu, Guo Wenming High fidelity error Q-filtering algorithm for dynamic measurement data [J]. Journal of Astronautics, 2016, 37(01): 112–7.

* View Article

* Google Scholar

21. 21. BAKDI A, KOUADRI A. A new adaptive PCA based thresholding scheme for fault detection in complex systems [J]. Chemometrics and Intelligent Laboratory Systems, 2017, 162: 83–93.

* View Article

* Google Scholar

22. 22. TAMURA M TSUJITA S. A study on the number of principal components and sensitivity of fault detection using PCA [J]. Comput Chem Eng, 2007, 31: 1035–46.

* View Article

* Google Scholar

23. 23. ZHOU F, PARK J H, LIU Y. Differential feature based hierarchical PCA fault detection method for dynamic fault [J]. Neurocomputing, 2016, 202: 27–35.

* View Article

* Google Scholar

24. 24. MUDHOLKAR J G S. Control procedures for residuals associated with principal component analysis [J]. Technometrics, 1979, 21(3): 341–9.

* View Article

* Google Scholar

Citation: Wang Y, Hu S (2025) Design and realization of compressor data abnormality safety monitoring and inducement traceability expert system. PLoS ONE 20(1): e0315917. https://doi.org/10.1371/journal.pone.0315917

About the Authors:

Yuan Wang

Roles: Conceptualization, Data curation, Methodology, Software, Validation, Visualization, Writing – original draft

Affiliation: School of Information and Control Engineering Jilin Institute of Chemical Technology Jilin, Jilin, China

Shaolin Hu

Roles: Conceptualization, Formal analysis, Investigation, Methodology, Writing – review & editing

E-mail: [email protected]

Affiliation: Automation School Guangdong University of Petrochemical Technology, Maoming, Guangdong, China

ORICD: https://orcid.org/0000-0003-0565-7450

References

1. LEI Y, JIA F, LIN J, et al. An Intelligent Fault Diagnosis Method Using Unsupervised Feature Learning Towards Mechanical Big Data [J]. IEEE Transactions on Industrial Electronics, 2016, 63: 3137–47.

2. LI C, LI C, TAO Y, et al. Improving forecasting accuracy of daily enterprise electricity consumption using a random forest based on ensemble empirical mode decomposition [J]. Energy, 2018.

3. LI X , ZHANG W, DING Q. Understanding and improving deep learning-based rolling bearing fault diagnosis with attention mechanism [J]. Signal Process, 2019, 161: 136–54.

4. LIU C-L , HSAIO W, TU Y-C. Time Series Classification With Multivariate Convolutional Neural Network [J]. IEEE Transactions on Industrial Electronics, 2019, 66: 4788–97.

5. NATH A G, UDMALE S S, SINGH S K. Role of artificial intelligence in rotor fault diagnosis: a comprehensive review [J]. Artificial Intelligence Review, 2020, 54: 2609–68.

6. TAO X , GONG X, ZHANG X Y, et al. Deep Learning for Unsupervised Anomaly Localization in Industrial Images: A Survey [J]. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 1–21.

7. YAZDI M, HAFEZI P, ABBASSI R. A methodology for enhancing the reliability of expert system applications in probabilistic risk assessment [J]. Journal of Loss Prevention in the Process Industries, 2019.

8. KABIR S, GEOK T K, KUMAR M, et al. A Method for Temporal Fault Tree Analysis Using Intuitionistic Fuzzy Set and Expert Elicitation [J]. IEEE Access, 2020, 8: 980–96.

9. WANG F, GAO J. A novel knowledge database construction method for operation guidance expert system based on HAZOP analysis and accident analysis [J]. Journal of Loss Prevention in The Process Industries, 2012, 25: 905–15.

10. BASURKO O C, URIONDO Z. Condition-Based Maintenance for medium speed diesel engines used in vessels in operation [J]. Applied Thermal Engineering, 2015, 80: 404–12.

11. SHAHID M, ZABIRI H, TAQVI S A A, et al. Fault root cause analysis using degree of change and mean variable threshold limit in non-linear dynamic distillation column [J]. Process Safety and Environmental Protection, 2024.

12. ALI H, MAULUD A S, ZABIRI H, et al. Multiscale Principal Component Analysis-Signed Directed Graph Based Process Monitoring and Fault Diagnosis [J]. ACS Omega, 2022, 7: 9496–512.

13. KOK T L, ALDRICH C, ZABIRI H, et al. Application of unthresholded recurrence plots and texture analysis for industrial loops with faulty valves [J]. Soft Computing, 2022, 26: 10477–92.

14. TAQVI S A A, ZABIRI H, UDDIN F, et al. Simultaneous fault diagnosis based on multiple kernel support vector machine in nonlinear dynamic distillation column [J]. Energy Science & Engineering, 2022, 10: 814–39.

15. HARROU F, KADRI F, KHADRAOUI S, et al. Ozone measurements monitoring using data-based approach [J]. Process Safety and Environmental Protection, 2016, 100: 220–31.

16. CHENG T, DAIRI A, HARROU F, et al. Monitoring Influent Conditions of Wastewater Treatment Plants by Nonlinear Data-Based Techniques [J]. IEEE Access, 2019, 7: 108827–37.

17. ZHOU X-G, GONG R, SHI F, et al. PetroKG: Construction and Application of Knowledge Graph in Upstream Area of PetroChina [J]. Journal of Computer Science and Technology, 2020, 35: 368–78.

18. TANG M, WU D, ZHANG S, et al. EPAN-SERec: Expertise preference-aware networks for software expert recommendations with knowledge graph [J]. Expert Syst Appl, 2024, 244: 122985.

19. JAWAD M S, DHAWALE C A, AZHAR A, et al. Adoption of knowledge-graph best development practices for scalable and optimized manufacturing processes [J]. MethodsX, 2023, 10.

20. Shaolin Hu, Na Fu, Guo Wenming High fidelity error Q-filtering algorithm for dynamic measurement data [J]. Journal of Astronautics, 2016, 37(01): 112–7.

21. BAKDI A, KOUADRI A. A new adaptive PCA based thresholding scheme for fault detection in complex systems [J]. Chemometrics and Intelligent Laboratory Systems, 2017, 162: 83–93.

22. TAMURA M TSUJITA S. A study on the number of principal components and sensitivity of fault detection using PCA [J]. Comput Chem Eng, 2007, 31: 1035–46.

23. ZHOU F, PARK J H, LIU Y. Differential feature based hierarchical PCA fault detection method for dynamic fault [J]. Neurocomputing, 2016, 202: 27–35.

24. MUDHOLKAR J G S. Control procedures for residuals associated with principal component analysis [J]. Technometrics, 1979, 21(3): 341–9.