ShuffleNet‐based comprehensive diagnosis for insulation and mechanical faults of power equipment

Abstract

Despite the complicated fault mechanism of power equipment, with the increasing promotion and development of the Ubiquitous Power Internet of Things (UPIoT), the fault information of power equipment can be instantly saved, which makes possible intelligent diagnosis via fault samples. This study proposes a new method to implement comprehensive intelligent diagnosis by adopting the ShuffleNet lightweight convolution neural network (SLCNN). Considering the requirements of the UPIoT intelligent terminal, this study constructs six models, which are measured in terms of recognition accuracy, model storage, and calculation cost when applied to insulation and mechanical datasets. Compared with the existing models, the SLCNN outperforms them significantly in terms of recognition accuracy, with an accuracy of 95.77% and 99.9% in insulation and mechanical fault diagnoses, respectively. The SLCNN also demonstrates obvious advantages in other performance indicators, all of which contribute to its use in the accurate and reliable fault diagnosis of power equipment in the UPIoT context. Furthermore, through comparing feature maps, it is discovered that the aliasing degree and boundary of insulation defects are not as obvious as those of mechanical faults, which means that insulation fault diagnosis is much more difficult than mechanical fault diagnosis.

INTRODUCTION

The basic components of the entire power system, including gas-insulated switchgear (GIS), transformers, circuit breakers, generators, and other power equipment, are of fundamental importance for the safe, reliable, and orderly operation of the entire power process, which includes power generation, transmission, distribution, and utilisation [1–3]. The failure of power equipment leads to power outages and inconveniences residents. Even worse, it can give rise to explosions and other accidents that severely threaten personal and property safety. As a result, reliable online monitoring and fault diagnosis of equipment is needed, so as to tackle problems in time and ensure the safe and reliable operation of the whole power system [4–6].

Statistics suggest that power equipment faults mostly take place in the insulation, machinery, and temperature rise [7]. And insulation and machinery faults are the two main types of faults. Insulation faults mostly manifest in the form of partial discharge (PD). With the advancement of the Ubiquitous Power Internet of Things (UPIoT), power equipment fault information can be instantly retained, which offers the prospect of power equipment fault diagnosis technology based on online monitoring. To build a reliable insulation and mechanical fault diagnosis system, intensive research has been conducted over the past two decades, mainly in regard to two aspects: (i) using signal pre-processing, such as the fast Fourier transform, wavelet transform, wavelet packet transform, empirical mode decomposition, and Hilbert–Huang transform, to construct perfect fault signal characteristics [8, 9]; and (ii) using pattern recognition methods to classify the features of the above structure, mainly including the support vector machine, artificial neural network, random forests, and other deep learning methods [10–14].

In recent years, with the rapid development of deep learning technology, a series of deep learning models have been increasingly introduced and adopted for intelligent fault diagnosis. Regarding insulation and mechanical vibration, the convolutional neural network (CNN) [15], deep belief network [16], stacked denoising auto-encoders (SDAE) [17], and recurrent neural network [18] have achieved unprecedented success in the field of fault diagnosis with the advantage of automatic learning of the high-level features. Lee et al. have investigated the use of CNNs on raw signals and found that the filters are in a position to detect unique features of each classification category [19]. Han et al. have proved that an adversarial learning framework can not only boost the generalisation ability of the trained model, but also avoid overfitting with a small number of labelled samples [8]. Yu et al. demonstrated the advantage of learning the combination of multiscale information, which aims to automatically capture complementary and discriminative feature representations from different scales of vibration signals [20]. Li et al. have used a multi-column CNN that incorporates ultra-high frequency spectra of multiple resolutions with excellent PD identification ability [21]. Wang et al. have discovered that the PD pattern recognition of SDAE shares high similarity among encoders [22]. Wang et al. have validated the feasibility of the MobileNet CNN for GIS PD pattern recognition in the application of big data and the UPIoT context [23].

The studies mentioned above, however, are largely based on the diagnosis of a single insulation or mechanical fault. Meanwhile, these networks fail to meet the network parameters and storage requirements of UPIoT mobile terminals. In addition, the overfitting and even vanishing gradient problem makes it difficult to train the models as the number of convolution layers increases to improve recognition accuracy.

As a network specifically designed for the mobile terminal, the ShuffleNet network greatly reduces the computational and storage cost of the model while ensuring the model's accuracy [24]. Since the ShuffleNet lightweight CNN (SLCNN) has not yet been used for intelligent fault diagnosis, in this study, a new method is proposed by adopting the SLCNN to implement the comprehensive intelligent insulation and mechanical fault diagnosis of power equipment. As for the proposed model, the concurrent mechanical and insulation fault samples have been considered, and meanwhile, model parameters have also been transferred and adjusted through transfer learning, which not only improves the accuracy of fault diagnosis but also accelerates the training speed of the model. By adopting lightweight convolution instead of traditional convolution, the number of parameters and calculation load of the model are effectively reduced. The introduction of residual structure alleviates the overfitting and even vanishing gradient problem. In addition to effectively improving the fault diagnosis accuracy, the method greatly reduces the calculation and storage costs of the model, so that it can be effectively deployed to the intelligent terminal of UPIoT, thus realising accurate and reliable diagnosis of power equipment.

The structure of this study is as follows: Section 2 systematically describes UPIoT. Section 3 describes the SLCNN proposed in this study. Section 4 clarifies the experimental details, results, and discussion. Finally, conclusions are drawn in Section 5.

UBIQUITOUS POWER INTERNET OF THINGS

As an important current concern, the UPIoT is widely considered critical to the transformation and upgrading of power grid enterprises. According to the definition of the UPIoT, the role that data communication and processing technologies, such as fibre optic networks, wireless private networks, and big data, play is of vital significance, as they can provide powerful technical support for the interconnection and high intelligence of the power grid [25]. Meanwhile, millions of pieces of power equipment transmit a tremendous amount of data to the UPIoT through intelligent sensing to characterise the operating status of devices, thus enabling the UPIoT to control its running status by providing massive data feedback.

In this context, the volume and types of data processed by the UPIoT have increased dramatically due to access to massive UPIoT nodes. To function within advanced application scenarios in the UPIoT context, such as intelligent operations and maintenance and reliable situational awareness, the key is to collect, store, and process massive amounts of UPIoT data. In this case, centralised data centre processing and the traditional cloud computing operational mode are difficult to adapt to the rapidly expanding data scale, which poses a great demand for better computing capacity in power data centres. For this reason, it is of great significance to seek rapid reliable processing of massive amounts of information by means of processing data via the Internet of Things (IoT) intelligent terminal, supported by edge computing, and then transferring the key information to the cloud [26].

The construction of the UPIoT also provides new insights for the fault diagnosis of power equipment. The deterioration of power equipment, such as transformers and GIS, is a comprehensive, multi-physics process involving heat, electricity, and force, which gives rise to the complicated ageing mechanism. During the construction of the UPIoT, the whole operational process of equipment will be monitored and retained [27]. This solves the problems of the fault samples in terms of representativeness and comprehensiveness. This proves that artificial intelligence technology can be used for in-depth mining with the aim to digitally and intelligently diagnose power equipment failures.

THE PROPOSED METHOD

ShuffleNet lightweight convolutional neural network

As a lightweight CNN designed for mobile terminal devices, ShuffleNet was proposed by Face++ in 2017 [28]. The core of ShuffleNet consists of two operations, pointwise group convolution and channel shuffle, which can greatly reduce the calculation cost while maintaining accuracy. The pointwise group convolution can significantly reduce the calculation loss, and channel shuffle can effectively alleviate the insufficient information flow among the pointwise group convolution channel groups, thereby achieving state-of-the-art performance.

Figure 1 illustrates the pointwise group convolution process with channel shuffle. For a g group of convolutional layers, there are g × n output channels. First, reshape the group into a (g, n) shape, and then transpose it into an (n, g) shape. Finally, the g group is returned as the input of the next layer by flattening. This operation is differentiable, and the model can maintain end-to-end training. Suppose that four input channels and eight output channels are divided into two selected groups; each group will have two input channels and four output channels. When there is only one group, its calculation cost is $D_f^2\times D_k^2\times 2\times 4$; when there are two groups, the calculation cost is $D_f^2\times D_k^2\times 4\times 4$, where $D_f\times D_f$ is the size of the output feature map, and $D_k\times D_k$ is the size of the convolution kernel. The computational cost is reduced by at least half.

[IMAGE OMITTED. SEE PDF]

The performance of the model, however, can be affected by multiple factors, such as storage space, platform usage, and computational complexity. Taking all these factors into consideration, Face++ proposed network design guidelines in 2018 and named the network ShuffleNet-V2, which is one of the most satisfactory lightweight CNNs [29]. The basic design principles for optimising lightweight CNNs are (1) minimising memory access by using the same channel size; (2) increasing memory access cost through excessive use of group convolution; (3) reducing parallelism via network fragmentation; and (4) improving the recognition rate and speed of the model without ignoring element-level operations.

The current popular shuffle units are shown in Figure 2. Among them, Figure 2a, b are shuffle operations that only consider the computational complexity when designed, while Figure 2c, d are shuffle operations that consider the above four basic principles.

[IMAGE OMITTED. SEE PDF]

The difference between Figure 2a, b is that the latter reduces the size of the feature map. The comparison of Figure 2a, c shows that, first, Figure 2c adds a channel split operation at the beginning. This operation divides the channel of the input feature into c-c' and c', where c' is c/2. Then Figure 2c cancels the group operation in the 1 × 1 convolution layer, and the previous channel split was actually a disguised group operation. The operation of channel shuffle has been moved to the back of concat. The fact that the first 1 × 1 convolution layer does not have group operation also demonstrates that it is not necessary to follow it with a channel shuffle. Finally, Figure 2c replaces the element-wise add operation with concat. The comparison between Figure 2b, d is the same as that between Figure 2a, c, except that there is no channel split operation at the beginning of Figure 2d, so the number of feature map channels doubles after the final concat.

Comparative studies

Based on two basic types of shuffle units, this study constructs a comprehensive intelligent fault diagnosis network with reference to the work of Face++. The network structure in this study is presented in Table 1. During the model building process, the basic parameters of the constructed network are the same, while the number of iterations of the shuffle units is different. Moreover, the fault waveform picture is used as the input of the network with a size of 96 × 96. The original input is 224 × 224, but considering speed, the size of the input is reduced in this study. Because of the decrease of the input, this study increases the output channel of the first convolution layer to 36, which previously was 24, so as to increase the amount of information in the first layer. In stage 2, this study will not perform the channel shuffle operation according to previous research [29] because there are few channels in this layer that are not suitable for scrambling.

TABLE 1 The structure of the ShuffleNet network model

Model parameters	Shuffle1	Shuffle2	Shuffle3
Layer	Output size	Stride	KSize	Output channels	Repeat
Image	$96\times 96$	-	-	3	-	-	-
Conv1	$48\times 48$	2	3 × 3	32	1	1	1
MaxPool	$24\times 24$	2	3 × 3	32
Stage2	$12\times 12$	2	-	144	1	1	1
$12\times 12$	1	-	144	1	3	3
Stage3	$6\times 6$	2	-	288	1	1	1
$6\times 6$	1	-	288	1	3	7
Stage4	$3\times 3$	2	-	576	1	1	1
$3\times 3$	1	-	576	1	3	3
GlobalPool	$1\times 1$	-	7 × 7	576	-	-	-
FC	-	-	-	128	-	-	-

Table 1 presents the common 3 × 3 convolution and max pool layers that we started with. Then there are three stages; each stage is a stack of several repeated basic units of the SLCNN. For each stage, the first basic unit uses stride = 2, so that the width and height of the feature map are reduced by half, and the number of channels is doubled. The following basic units are stride = 1, and the feature map and the number of channels remain unchanged. As for the basic unit, the number of channels of the 3 × 3 convolution layer in the bottleneck layer is one-fourth of that of the output channels, which is the same as the design concept of the residual unit.

Comprehensive intelligent diagnosis framework

The process of comprehensive intelligent diagnosis of power equipment faults in the UPIoT context is shown in Figure 3. After receiving the intelligent sensing information, the UPIoT intelligent terminal recognises the device and the intelligent sensing information via the device fingerprint information embedded in the CIM4.0 model and the radiofrequency identification (RFID) technology. After the sensor information is identified, the mechanical or insulation defect mode is selected, and then the process moves into fault identification. During data processing, the original vibration signal is converted into a two-dimensional picture for processing through Gramain Angular Field method [30]. After a series of processing steps, the device perception information images perform fault diagnosis via the trained SLCNN. The biggest difference between the diagnosis of mechanical and insulation defects lies in the fact that the insulation information needs to be normalised before diagnosis. The main reason is that the amplitude varies in the defect information collected by the same sensor at different voltages and defect locations for the same defect of the same device. Normalisation can effectively help to avoid the errors caused by such problems.

[IMAGE OMITTED. SEE PDF]

The recognition of device fingerprint information in the UPIoT context mainly depends on RFID technology. The RFID system consists of four main components: the RFID reader, RFID antenna, RFID tag, and back-end UPIoT intelligent terminal for data collection [31]. The intelligent terminal of the UPIoT manipulates the RFID reader to perform a tag query through a serial port or network transmission. The reader sends out the modulated RFID signals to the tag through the antenna, and receives backscattered signals from the tag to read and decode the tag.

The reader finally returns the tag information to the background server, and the server further processes the tag information at the application layer. The overall framework of the RFID system is shown in Figure 4.

[IMAGE OMITTED. SEE PDF]

Of the whole process of the intelligent diagnosis of power equipment, the most important part is the fault diagnosis algorithm, namely the SLCNN proposed in this study. The feasibility of the SLCNN in the UPIoT context is, to a large extent, determined by the model's storage space, calculation cost, processing speed, and diagnostic accuracy. As typical representatives of insulation defects and mechanical defects, GIS insulation defects and rotating machinery vibration datasets are used to verify the performance of the SLCNN. Given the unique nature of the dataset, an image rotation transformation will not only fail to strengthen the data, but will introduce more errors. Therefore, we use conditional variational auto-encoders to generate data for strengthening [32].

EXPERIMENTS AND ANALYSIS

GIS PD diagnosis

Experiment description

As one of the main forms of insulation defect, PD further exacerbates equipment ageing [33]. A GIS insulation defect, as a typical representative of power equipment insulation defects, is of great significance to the research of power equipment insulation diagnosis. In view of the sample size requirements of deep learning models, this study uses the finite difference time domain (FDTD) method to simulate GIS PD, and obtains the PD signals via the XFdtd high-frequency electromagnetic field simulation software. A series of studies show that the PD signals obtained by FDTD simulation and by experiment are similar and universal [34–36]. During the simulation, four typical GIS PD defects are selected for data acquisition, namely free metal particle defects (M), metal tip defects (N), suspended electrode defects (O), and insulator air gap defects (P). To attain the actual GIS electromagnetic propagation characteristics, the simulation model shown in Figure 5 is established with reference to the actual GIS testing apparatus.

[IMAGE OMITTED. SEE PDF]

The outer diameter of the inner conductor of the GIS simulation model is 106 mm, the inner diameter of the shell is 320 mm, and the total length of the cavity is 7288 mm. Circular package covers are placed on the left and right sides of the T branch structure. Both the conductor and cover materials are aluminium. Two pot-type support insulators with 40 mm thickness and 520 mm outer diameter are installed at 533 and 2885 mm from the left side cover. The model uses an ideal coaxial structure similar to the actual GIS tank. The relative dielectric constant ε = 4 and the electrical conductivity is 0. The boundary is set as a perfect match layer to reduce the influence of the boundary structure on the simulation calculation and ignore the weak energy attenuation caused by the propagation structure itself during the simulation. The relative dielectric constant of the SF₆ filling the GIS cavity is 1.00205, and the density is set as 23.7273 kg/m³ under a pressure of 0.4 MPa-abs. In the software, the mesh size is 2 × 2 × 2 cm, with a total of 471 × 91 × 143 cells. The simulation step size is set as 38.52 ps, and the simulation time length is set as 150 ns.

The current pulses generated by the PD of different defects in GIS are not the same. In this study, the discharge signals and experimental conclusions measured in actual discharge tests are selected. Free metal particles, metal tips, suspended electrodes, the insulator gas gap, and ideal Gaussian pulses are used as excitation sources, and they are changed into src format based on their signal data to be imported into the XFdtd software. The setting position of the excitation source is shown in Figure 5. The receiving point in the same direction as the excitation source is placed at different locations along the conductor direction, as shown in Figure 5. The received load impedance is 50 Ω. Each measurement point is measured at intervals of $15°$ with respect to the defect distribution of $\sim 0°–180°$. The PD waveform signals of the four typical defects are presented in Figure 6.

[IMAGE OMITTED. SEE PDF]

Results

In this study, we simulate 2500 sets of data for each type of defect to form a GIS PD pattern recognition classification dataset, of which 80% is used for training (70% as a training set and 10% as a validation set), and 20% for model testing. We also train our model with Keras (TensorFlow back-end) on a machine that has a GeForce RTX 2060 GPU on traditional terminal and GeForce RTX 2080 GPU on intelligent terminal, 250 GB SSD*2 CPU. To verify the superiority of the SLCNN proposed in this study and find the optimal model, we construct six models that adopt the two different structures shown in Figure 2b, d, namely S-11, S-12, S-21, S-22, S-31, and S-32. In terms of these model names, the first number indicates the model structure introduced in Section 3.2, and the second number indicates whether the structure is that of Figure 2b or Figure 2d in the adopted Shuffle block (the S1 or S2 series, respectively). Three traditional CNNs, MobileNet, LeNet, AlexNet, and ResNet18, are used for PD pattern classification. The classification recognition results are shown in Table 2.

TABLE 2 Partial discharge diagnosis results

Model	Defects type	Overall (%)
M (%)	N (%)	O (%)	P (%)
S11	89.00	80.08	41.64	100.00	77.86
S12	97.20	94.30	68.36	100.00	89.97
S21	86.4	100.0	72.62	100.00	89.76
S22	100	96.7	88.36	98.00	95.77
S31	79.80	97.56	80.16	99.00	89.13
S32	93.80	100	89.84	96.00	94.91
MN	92.80	96.64	88.72	93.08	92.81
LN	76.42	84.26	68.36	88.54	79.40
AN	86.26	88.24	77.60	89.36	85.36
RN	86.80	98.98	80.82	99.00	91.40

Table 2 shows that the fault diagnosis accuracy of S22 is 95.77%, significantly higher than that of the other models. For the same series (S1 or S2), the recognition accuracy of the model significantly improves with increasing the network depth from 1 to 2. The minimum recognition accuracy of the model improves from 2 to 3, but the overall recognition rate declines. Since AN training starts from zero, a problem occurs in that the vanishing gradient and it cannot be trained, so it needs to be trained by a migration strategy. For the PD fault diagnosis of power equipment, both the minimum recognition rate of a single fault and the overall recognition rate should be considered. For example, both S22 and S32 have advantages in these two aspects. S32, however, demonstrates much more obvious disadvantages under other defect types with a slight improvement in the single recognition rate. In this case, S22 performs best in terms of accuracy.

Because the applicability of the model in the UPIoT context is determined by training time, storage space, and number of parameters, we present other model performance indicators in Table 3. Table 3 shows that the number of parameters and storage space of the SLCNN constructed in this study are greatly reduced compared with traditional models. The parameters and storage space of the S1 series model are the smallest, much smaller than in the S2 series model at the same level. The S2 series model costs much more to improve the recognition accuracy. From the perspective of time, the training time of the S1 series model is shorter than that of the S2 series model. In terms of testing time, there is not much difference between S1x and S2x, but there is a substantial increase to S3x. Considering the importance of the reliability of power equipment to the safe operation of the entire power grid, accuracy is the primary measurement indicator of the diagnostic model, with other indicators also measured. The S22 model is best for intelligent insulation diagnosis in the UPIoT context. Simultaneously, it can be seen that the training and testing time of the lightweight CNN is significantly longer than that of the traditional CNN. That is to say, the model accuracy is improved at the high cost of time.

TABLE 3 Other performance indicators of the model on intelligent terminal

Model	Parameters (Million)	Weight storage (MB)	Training time (minutes)	Testing time (seconds)
S11	0.34	1.57	2.47	9.36
S12	2.39	9.68	3.82	10.01
S21	0.45	2.14	3.63	10.37
S22	2.67	10.96	5.18	10.58
S31	0.97	4.37	5.32	13.13
S32	3.99	16.42	7.36	14.09
MN	3.21	12.83	7.02	13.59
LN	14.43	56.38	2.37	6.94
AN	24.72	96.59	3.32	7.49
RN	11.18	43.84	5.91	10.45

Table 4 gives the training and test results of the above models on traditional terminals. The comparison with Table 2 and Table 3 indicates that with the UPIoT intelligent terminal as the platform, the fault diagnosis recognition rate increases by more than 1 percentage on average. The main reason is that under the UPIoT intelligent terminal, the model parameters can be calculated more flexibly and efficiently. In other words, a large-capacity hardware platform is the necessary basis for the diagnosis method to achieve the best performance. Once the model is selected, its parameters and storage have been determined and will not change with the platform. We can also discover that with the same amount of calculation, the UPIoT intelligent terminal can shorten the training and testing time of the model through its high-performance hardware conditions, so as to achieve accurate and reliable diagnosis under UPIoT. Among them, the training time of the model is shortened more obviously, which can guarantee the update speed of the model. In terms of testing time, it is more difficult to reduce the testing time in seconds. Therefore, the saving of 1 s of testing time in this study is crucial to ensure the reliable diagnosis of faults.

TABLE 4 The training and test results of the above models on traditional terminals

Model	Diagnosis Accuracy (%)	Training time (minutes)	Testing time (seconds)
S11	76.36	3.24	10.04
S12	87.13	4.78	10.62
S21	88.24	4.51	10.83
S22	94.52	6.32	11.03
S31	88.62	7.91	13.68
S32	92.96	8.43	14.67
MN	91.76	8.24	14.03
LN	78.86	2.98	7.12
AN	83.44	4.32	7.98
RN	90.36	6.94	11.23

Mechanical fault diagnosis on Machinery Failure Prevention Technology dataset

Experiment description

Mechanical faults mainly occur in the bearings of rotating equipment such as wind turbines and hydraulic turbines, high-voltage circuit breaker operating mechanisms and equipment bodies, and isolation switches of GIS equipment. Among these locations, rotating equipment is frequently operated so that research on technology for the intelligent diagnosis of mechanical faults can be conducted with the massive amount of data. Since switch equipment rarely moves during service, with some switches operated only once or twice during several decades of service, this equipment is currently the bottleneck of research on fault diagnosis. Furthermore, the information on characteristics in the mechanical fault signals of the switch equipment in significantly higher level and less accessible than that in the signals of the rotating equipment. In this case, rotating equipment is much more suitable to study mechanical fault diagnosis. Transfer learning will help with performing mechanical diagnoses of other equipment [37].

As a widely used open set in current mechanical fault diagnosis research, the MFPT bearing fault set possesses typical representative significance for mechanical vibration signal fault diagnosis. The bearing failure datasets, provided by the American Society for Machinery Failure Prevention Technology (ASMFPT) and prepared by Dr. Eric Bechhoefer, chief engineer of NRG Systems on behalf of the ASMFPT, are adopted to facilitate research in bearing analysis [38, 39]. The bearing fault dataset is composed of data from the bearing testing stand: normal bearing data, outer ring fault data under different loads, inner ring fault data under different loads, and data from three real fault cases. Figure 7 shows a rolling element striking a local fault at the inner race.

[IMAGE OMITTED. SEE PDF]

To ensure the balance of the samples, three baseline conditions (NOR), seven inner race fault conditions (IR), and seven outer race fault conditions (OR) are selected to construct training samples to verify the performance of the SLCNN models. Thereby, the data consist of the following data points: NOR with 1,757,808 data points, IR with 1,025,388 data points, and OR with 1,025,388 data points, with the data sampling rate of 97,656. The total images produced from the dataset are as follows: NOR with 1800, In with 2100, and Out with 2100, with the time step of each picture at 0.01 s. The three types of sample pictures are shown in Figure 8.

[IMAGE OMITTED. SEE PDF]

Results

Like insulation diagnosis, 80% of the mechanical failure samples are used for training (70% as the training set and 10% as the validation set) and 20% are used for model testing. In the training process, we use transfer learning strategies. That is to say, the above PD diagnosis model parameters are fine-tuned to realize model training on the mechanical dataset. The recognition results of different models are shown in Table 5.

TABLE 5 Diagnosis results of MFPT mechanical failure

Model	Defects type	Overall (%)
NOR (%)	IR (%)	OR (%)
S11	100.00	90.95	65.00	77.86
S12	100.00	86.19	94.05	93.41
S21	100.00	69.52	78.33	89.76
S22	100.00	99.76	100.00	99.92
S31	100.00	78.33	81.90	86.74
S32	100.00	99.76	100.00	99.92
MN	100.00	99.76	100.00	99.92
LN	100.00	97.14	98.09	98.41
AN	100.00	99.05	100.00	99.68
RN	100.00	99.52	100	99.84

Table 5 shows that S22 and S32 demonstrate the highest recognition accuracy. The recognition accuracy of the S1 series models is very low, and these models find it particularly difficult to distinguish IR and OR faults. For the S2 series models, S22 almost reaches the highest recognition accuracy rate of almost 100% on the MFPT mechanical fault dataset. So further increasing the model depth will only end up increasing the cost on other performance indicators as well as the probability of overfitting. Trained directly, LN and AN demonstrate poor performance, AN in particular due to the vanishing gradient. Through transfer learning, the models also achieve good results for the fault diagnosis of mechanical datasets. The sample sizes of the mechanical dataset and the insulation dataset seem similar, and other performance indicators are similar to those of the insulation defects, so the results are not presented here.

Mechanical fault diagnosis on Case Western Reserve University fault dataset

Experiment description

As one of the most influential public datasets for mechanical fault diagnosis, Case Western Reserve University Bearing fault dataset is regarded as a reference for research and verification of the fault diagnosis performance of the proposed model [40]. The testing object is the driving end bearing, and the bearing type is a deep groove ball bearing SKF6205. The faulty bearing is made by electrical discharge machining (EDM), and the sampling frequency is 12 kHz. The bearing has a single-point fault set by the EDM process. The fault diameter includes four sizes, namely 0.178 mm, 0.356 mm, 0.533 mm, and 0.711 mm.

In this study, the driving end vibration signals are selected as the experimental data, including the vibration signals in four different states, namely Normal (NOR), Ball Fault (BF), Outer Race Fault (ORF), and Inner Race Fault (IRF). The signals collected in each state vary because of the different fault diameters and loads. The load size is respectively 0, 746, 1492, and 2238 W. The 12 k Drive End Bearing Fault Data is diagnosed by using 200 data points at a time. In total, 10 types of signal data are constructed for experiment, which cover the four signal states. With reference to [19], the 10 types of data are 0.178 mm BF (0), 0.178 mm IRF (1), 0.178 mm ORF (2), 0.356 BF (3), 0.711 mm IRF (4), 0.356 mm ORF (5), 0.533 mm BF (6), 0.533 mm IRF (7), 0.533 mm ORF (8), NOR (9).

Results

Like insulation diagnosis, 80% of the mechanical failure samples are used for training (70% as the training set and 10% as the validation set) and 20% are used for model testing. The recognition results of different models are shown in Table 6.

TABLE 6 Diagnosis results of Case Western Reserve University mechanical failure

Model	S11	S12	S21	S22	S31	S32	MN	LN	AN	RN
Result (%)	68.94	89.66	89.64	99.91	94.09	99.89	99.02	97.94	99.45	87.96

Table 6 shows that S22 demonstrates the highest recognition accuracy. The recognition accuracy of the S1 series models is very low. For the S2 series models, S22 almost reaches the highest recognition accuracy rate of almost 100% on the Case Western Reserve University mechanical fault dataset. So further increasing the model depth will only end up increasing the cost on other performance indicators as well as causing overfitting. Figure 9 shows the confusion matrix results of the S22, AN, LN and S31 model. From the confusion matrix, LN has the lowest diagnostic accuracy on data type (0), while the other models have low accuracy on type (3) and (4), but demonstrate high accuracy on type (6). Trained directly, LN and AN have suffered the overfitting and even vanishing gradient problem owing to the limited sample size. Through transfer learning, the models also achieve good results for the fault diagnosis of mechanical datasets. Other performance indicators have similar rules to the insulation defects, so the results are not presented here.

[IMAGE OMITTED. SEE PDF]

Analysis and discussion

The performance of the SLCNN models constructed in this study in the diagnosis of insulation defects or mechanical faults is verified in Sections 4.1, 4.2 and 4.3, respectively. Model S22 performs best according to comprehensive performance indicators, so it is the best model for the comprehensive intelligent diagnosis of insulation and mechanical faults in the context of the UPIoT.

To give an intuitional interpretation of the model predictions, t-distributed stochastic neighbour embedding (t-SNE) is applied to visualise the learned features in the last hidden, fully connected layer (FC). Figure 10 shows the visualisation results of the mechanical and insulation fault diagnoses. Figure 10 shows that in the planarised SLCNN layer, samples of the same fault type are clearly collected together and can even be separated, which indicates that the feature descriptors have good feature representation capabilities. After performing non-linear mapping in the classifier, except for the slight overlap of individual samples, the characteristics of different machine conditions are well separated in the last hidden, FC, which is consistent with the diagnostic accuracy results in Tables 2, 5 and 6. Furthermore, the aliasing degree and boundary of the insulation defect are not as obvious as those of the mechanical fault data, which indicates that PD has more significant randomness. Therefore, PD pattern recognition is more difficult than the diagnosis of mechanical faults.

[IMAGE OMITTED. SEE PDF]

CONCLUSION

In this study, we propose a comprehensive intelligent diagnosis method using the SLCNN for the mechanical and insulation fault diagnoses of power equipment in the context of the UPIoT. It is the first time that comprehensive intelligent diagnoses are conducted for the insulation and mechanical faults of power equipment, with multiple models constructed to seek an optimal model. The S22 model accurately diagnoses insulation and mechanic faults, with an accuracy of 95.77%, 99.92% and 99.91%, respectively. In addition, the model possesses obvious advantages over traditional methods in relation to multiple indicators such as number of parameters, storage space, training time, and testing time. The model is further visualised with t-SNE, which validates the outstanding ability of S22 in regard to identifying insulation and mechanical failure. Moreover, comparing the results of the insulation and mechanical diagnoses shows that the aliasing degree and boundary of the insulation defect are not as obvious as those of the mechanical failure, which indicates that PD has greater randomness. Therefore, insulation fault diagnosis is more difficult and complicated than mechanical fault diagnosis. In summary, the SLCNN proposed in this study is more capable of the reliable and accurate diagnosis of insulation and mechanical faults, and has broad application possibilities in the context of the UPIoT.

References

Janani, H., Kordi, B.: Towards automated statistical partial discharge source classification using pattern recognition techniques. High Volt. 3(3), 162–169 (2018)

Tang, J., Jin, M., Zeng, F., et al.: Assessment of PD severity in gas‐insulated switchgear with an SSAE. IET Sci. Meas. Technol. 11(4), 423–430 (2017)

Okabe, S., Ueta, G., Hama, H., et al.: New aspects of UHF PD diagnostics on gas‐insulated systems. IEEE Trans. Dielectr. Electr. Insul. 21(5), 2245–2258 (2014)

Wang, J., Yin, H.: Failure rate prediction model of substation equipment based on weibull distribution and time series analysis. IEEE Access. 7, 85298–85309 (2019)

Cheng, Y., Bi, J., Chang, W., et al.: Proposed methodology for online frequency response analysis based on magnetic coupling to detect winding deformations in transformers. High Volt. 5(3), 343–349 (2020)

Kluss, J., Chalaki, M., Whittington, W., et al.: Porcelain insulation – defining the underlying mechanism of failure. High Volt. 4(2), 81–88 (2019)

Zeng, F., Dong, Y., Tang, J.: Feature extraction and severity assessment of partial discharge under protrusion defect based on fuzzy comprehensive evaluation. IET Gener., Transm. Distrib. 9(16), 2493–2500 (2015)

Han, T., Liu, C., Yang, W., et al.: A novel adversarial learning framework in deep convolutional neural network for intelligent diagnosis of mechanical faults. Knowl. Base Syst. 165, 474–487 (2019)

Wang, X., Li, X., Rong, M., et al.: UHF signal processing and pattern recognition of partial discharge in gas‐insulated switchgear using chromatic methodology. Sensors. 17(1), 177 (2017)

Liu, R., Yang, B., Zio, E., et al.: Artificial intelligence for fault diagnosis of rotating machinery: A review. Mech. Syst. Signal Process. 108, 33–47 (2018)

Khan, Q., Refaat, S., Abu‐Rub, H., et al.: Partial discharge detection and diagnosis in gas insulated switchgear: State of the art. IEEE Elect. Insul. Mag. 35(4), 16–33 (2019)

Peng, X., Li, J., Wang, G., et al.: Random forest based optimal feature selection for partial discharge pattern recognition in HV cables. IEEE Trans. Power Deliv. 34(4), 1715–1724 (2019)

Budiman, F.N., Khan, Y., Malik, N.H., et al.: Utilization of artificial neural network for the estimation of size and position of metallic particle adhering to spacer in GIS. IEEE Trans. Dielectr. Electr. Insul. 20(6), 2143–2151 (2013)

Duan, L., Hu, J., Zhao, G., et al.: Method of inter‐turn fault detection for next‐generation smart transformers based on deep learning algorithm. High Volt. 4(4), 282–291 (2019)

Duan, L., Hu, J., Zhao, G., et al.: Identification of partial discharge defects based on deep learning method. IEEE Trans. Power Deliv. 34(4), 1557–1568 (2019)

Shao, H., Jiang, H., Zhang, H., et al.: Rolling bearing fault feature learning using improved convolutional deep belief network with compressed sensing. Mech. Syst. Signal Process. 100, 743–765 (2018)

Wang, G., Yang, F., Peng, X., et al.: Partial discharge pattern recognition of high voltage cables based on the stacked denoising autoencoder method. Proceedings of the 2018 International Conference on Power System Technology (POWERCON). Guangzhou, China, pp. 3778–3792. 6–8 November (2018)

Nguyen, M.T., Nguyen, V.H., Yun, S.J., et al.: Recurrent neural network for partial discharge diagnosis in gas‐insulated switchgear. Energies. 11(5), 1202 (2018)

Lee, D., Siu, V., Cruz, R., et al.: Convolutional neural net and bearing fault analysis. Proceedings of the International Conference on Data Mining Series (ICDM). Barcelona, San Diego, CA, USA, pp. 194–200. 12–15 December (2016)

Yu, H., Wang, K., Li, Y.: Multiscale representations fusion with joint multiple reconstructions autoencoder for intelligent fault diagnosis. IEEE Signal Proc. Let. 25(12), 1880–1884 (2018)

Li, G., Wang, X., Li , X., et al.: Partial discharge recognition with a multi‐resolution convolutional neural network. Sensors. 18(10), 3512 (2018)

Peng, X., Yang, F., Wang, G., et al.: A convolutional neural network‐based deep learning methodology for recognition of partial discharge patterns from high‐voltage cables. IEEE Trans. Power Del. 34(4), 1460–1469 (2019)

Wang, Y., Yan, J., Sun, Q., et al.: A mobilenets convolutional neural network for gis partial discharge pattern recognition in the ubiquitous power internet of things context: optimization, comparison, and application. IEEE Access. 7, 150226–150236 (2019)

Zhang, X., Zhou, X., Lin, M., et al: Shufflenet: An extremely efficient convolutional neural network for mobile devices. Proceedings of the International Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, pp. 6848–6856 (2018)

Hu, W., Yao, W., Hu, Y., et al.: Selection of cluster heads for wireless sensor network in ubiquitous power internet of things. Int. J. Comput. Commun. Contr. 14, 344–358 (2019)

Niu, X., Shao, S., Xin, C., et al.: Workload allocation mechanism for minimum service delay in edge computing‐based power internet of things. IEEE Access. 7, 83771–83784 (2019)

Suma, M., Madhumathy, P.: Acquisition and mining of agricultural data using ubiquitous sensors with internet of things. Proceedings of the International Conference on Computer Networks and Communication Technologies (ICCNCT). Springer, Singapore, pp. 249–261 (2016)

Lyu, J., Zhang, S., Qi, Y., et al.: AutoshuffleNet: Learning Permutation Matrices Via an Exact Lipschitz Continuous Penalty in Deep Convolutional Neural Networks. arXiv: 1901.08624 (2019)

Ma, N., Zhang, X., Zheng, H.‐T., et al.: Shufflenet V2: Practical guidelines for efficient CNN architecture design. Proceedings of the European Conference on Computer Vision (ECCV). Munich, Germany: Mar., pp. 116–131 (2018)

Wang, Z., Oates, T.: Encoding time series as images for visual inspection and classification using tiled convolutional neural networks. Proceedings of the Workshops at the Twenty‐Ninth AAAI Conference on Artificial Intelligence, pp. 40–46, Austin, TX, USA, 25–30 January (2015)

Fischer, M., Ferdik, M., Rack, L., et al.: An experimental study on the feasibility of a frequency diverse UHFf RFID system. IEEE Access. 7, 132311–132323 (2019)

Wang, Y., Yan, J., Yang, Z., et al.: Partial discharge pattern recognition of gas‐insulated switchgear via a light‐scale convolutional neural network. Energies. 12, 4674 (2019)

Gopinath, S., Sathiyasekar, K., Padmanaban, S., et al.: Insulation condition assessment of high‐voltage rotating machines using hybrid techniques. IET Gener., Transm. Distrib. 13(2), 171–180 (2019)

Li, X., Wang, X., Yang, A., et al.: Partial discharge source localization in gis based on image edge detection and support vector machine. IEEE Trans. Power Deliv. 34(4), 1795–1802 (2019)

Li, X., Wang, X., Yang, A., et al.: Propogation characteristics of PD‐induced UHF Signal in 126 KV GIS with three‐phase construction based on time–frequency analysis. IET Sci. Meas. Technol. 10(7), 805–812 (2016)

Behrmann, G., Franz, S., Smajic, J., et al.: UHF PD signal transmission in GIS: effects of 90° bends and an l‐shaped cigre step 1 test section. IEEE Trans. Dielectr. Electr. Insul. 26(4), 1293–1300 (2019)

Cao, P., Zhang, S., Tang, J.: Preprocessing‐free gear fault diagnosis using small datasets with deep convolutional neural network‐based transfer learning. IEEE Access. 6, 26241–26253 (2018)

Condition Based Maintenance Fault Database for Testing of Diagnostic and Prognostics Algorithms. https://mfpt.org/fault-data-sets/. Accessed October 2012.

Barbini, L., Ompusunggu, A., Hillis, A., et al.: Phase editing as a signal pre‐processing step for automated bearing fault detection. Mech. Syst. Signal Process. 91, 407–421 (2016)

Loparo, K.A.: Case Western Reserve University Bearing Data Center. http://csegroups.case.edu/bearingdatacen-ter/home (2012)