A Comparative Capsule Network Method for Intelligent Diagnosis of Wheel Pairs and Bearings under High Noise

Abstract

The invention discloses a contrast capsule network method for intelligent diagnosis of train wheel pair bearings under high noise, which relates to the field of intelligent fault diagnosis of train wheel pair bearings. The fault diagnosis method proposed by the invention overcomes the problem of poor network expression ability and difficult extraction of high-noise fault information. The capsule network based on the contrast capsule network can filter and screen features under high noise, extract features that can better reflect the essence of the data, and thus improve the fault recognition accuracy of the model. It can be used for fault diagnosis of train wheel pair bearings under high noise, and can realize "end-to-end" intelligent fault diagnosis from data acquisition to bearing health status categories; compared with traditional deep learning methods, this method combines the strategies of contrast learning and two-stage training models, enabling the invention to achieve higher diagnostic accuracy of train wheel pair bearings under high noise.

Description

Technical Field

[0001] The present invention relates to the intelligent fault diagnosis of train wheel set bearings, and particularly to a contrast capsule network method for intelligent diagnosis of train wheel set bearings under high noise. Background Art

[0002] The train wheel set is a large-scale complex electromechanical system, and the bearing is an extremely important mechanical component therein, which is widely used in train wheel sets. The operation state monitoring and fault diagnosis of the wheel set bearing are of great significance for ensuring the reliability of the equipment and avoiding safety accidents.

[0003] The train runs in a harsh environment for a long time. Once a fault occurs in the wheel set bearing and the abnormal condition is not successfully monitored in time, major accidents such as train derailment may eventually occur. However, since the fault feature information of the train wheel set bearing is often submerged by high background noise and other unstable components, extracting the feature information of the train wheel set bearing has become a difficult task. General fault diagnosis models for train wheel set bearings cannot effectively extract fault features and classify them correctly.

[0004] Currently, intelligent diagnosis of mechanical faults can be roughly divided into three types of methods. One is the method based on statistical analysis, which includes methods such as grey theory method, time series method, and multivariate statistical analysis method. The models of these methods are simple and the diagnostic accuracy is not good. The second is the method based on signal processing, which includes methods such as wavelet transform method, envelope analysis method, and spectral analysis method. These methods often require complex operations to find the fault feature frequency to identify the fault type. The third is the method based on artificial intelligence, which includes not only shallow networks such as support vector machines and extreme learning machines, but also deep learning networks such as deep belief networks, recursive neural networks, and convolutional neural networks. These methods can usually achieve end-to-end mapping, and using these methods to diagnose train wheel set bearing faults is a current research hotspot.

[0005] The difference between the capsule network and the traditional deep learning network is that each capsule in the capsule network is a vector, rather than the output of a neuron in the traditional neural network being a scalar, which enables the capsule network to extract more detailed features from the input data and has stronger feature expression ability. The capsule network updates the capsule layer parameters through a dynamic routing mechanism. Therefore, the capsule network is more suitable for processing highly non-linear structured data such as mechanical vibration signals.

[0006] Contrast learning focuses on learning the common features between the states of the same type of train wheel set bearings and distinguishing the differences between different category samples. While traditional deep learning can only learn the mapping from input to output and cannot learn the similarity between samples. Summary of the Invention

[0007] The present invention provides a contrast capsule network method for intelligent diagnosis of train wheel pair bearings under high noise, so as to realize effective intelligent diagnosis of train wheel pair bearings and solve the problems raised in the above background technology. The signal-to-noise ratio in the high-noise scenario here is generally less than 0 dB.

[0008] To achieve the above object, the technical solution of this method is a contrast capsule network method for intelligent diagnosis of train wheel pair bearings under high noise, including the following steps:

[0009] S1. Collect the mechanical equipment health status data collected by sensors according to the specific categories of train wheel pair bearings, preprocess the data, and establish a train wheel pair bearing health status database;

[0010] The sensor type is one or more of a vibration displacement sensor, a vibration velocity sensor, a vibration acceleration sensor, or a sound signal sensor; the data in the status database includes: normal status, single fault status, and compound fault status; the single fault status includes: rolling element fault, inner race fault, outer race fault; the compound fault status includes: outer race + inner race compound fault, inner race + rolling element compound fault, outer race + rolling element compound fault, inner race + outer race + rolling element compound fault; each data with the same health status has the same status label;

[0011] S2. Based on the train wheel pair bearing health status database, train the contrast capsule feature extraction model to obtain the trained contrast capsule feature extraction model; the loss function of the contrast capsule feature extraction model is a supervised contrast learning loss function, select an optimization algorithm to train the model until convergence, and the contrast capsule feature extraction model includes two parts: an improved time-domain convolutional network and a primary capsule layer;

[0012] The contrast capsule feature extraction model includes an improved time-domain convolutional network and a primary capsule layer connected in sequence. The time-domain convolutional network includes an input module, n multi-scale residual blocks, and an output module connected in sequence. The multi-scale residual block includes two inputs and two outputs. Among them, input 1 passes through two Inception units, a BN+ReLU+Dropout module, and a DDCID(1,1) layer in sequence. Here, DDCID is the abbreviation of causal dilated 1D convolution, where the parameter 1 refers to the kernel size and the parameter 2 refers to the causal dilation rate. The output of the DCCID(1,1) layer in the multi-scale residual block is added to the output of the input 2 of the multi-scale residual block that passes through a max-pooling layer and an Inception unit in sequence. One path serves as the output 2 of the multi-scale residual block, and the other path serves as the output 1 of the multi-scale residual block after passing through a BN+ReLU+Dropout module. The input 1 and input 2 of the first multi-scale residual block of the contrast capsule feature extraction model are both the input of the contrast capsule feature extraction model. The previous output 1 of the subsequent multi-scale residual blocks corresponds to the next input 1, and the previous output 2 corresponds to the next input 2. The output 1 of the last multi-scale residual block serves as the output of the contrast capsule feature extraction model.

[0013] After the input of the Inception unit, it is first divided into four paths. The first path includes a DCCID(1,1) layer, the second path includes a DCCID(3,3) layer and a DCCID(1,1) layer connected in sequence, the third path includes a DCCID(5,2) layer and a DCCID(1,1) layer connected in sequence, and the fourth path includes a DCCID(1,1) layer and a DCCID(3,2) layer connected in sequence. Then, the outputs of the four paths pass through an efficient channel attention mechanism module, which is the output of the Inception unit. The efficient channel attention mechanism first performs deep fusion on it, and secondly completes the cross-channel information interaction of the features of different channels.

[0014] S3. Freeze the weights of the above contrast capsule feature extraction model, and add a digital capsule layer and a Length layer after the contrast capsule feature extraction model to obtain the contrast capsule network model. The parameter update between the primary capsule layer and the digital capsule layer uses a dynamic routing mechanism, and the Length layer is used for the classification of the health status of the train wheel pair bearing. The loss function of the contrast capsule network model is the margin loss function. The contrast capsule network model is trained on the training set until convergence to obtain the contrast capsule network model.

[0015] S4. In the actual diagnosis process, obtain the health status data of the train wheel pair bearing, preprocess it, and input it into the contrast capsule network model for status judgment.

[0016] Further, the data preprocessing method in step S1 includes one of data normalization, data standardization, wavelet packet transform noise reduction, and ensemble empirical mode decomposition noise reduction;

[0017] Further, model training includes:

[0018] Adopt one of the Adam and SGD optimization algorithms as the model optimizer; adopt one of learning rate decay and cosine learning rate change as the model training strategy.

[0019] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0020] The fault diagnosis method proposed by the present invention overcomes the problem of poor network expression ability and difficult extraction of high-noise fault information. The capsule network based on the contrast capsule network can filter and screen features under high noise, extract features that can better reflect the essence of the data, and thus improve the fault recognition accuracy of the model. For the fault diagnosis of train wheel set bearings under high noise, it realizes the "end-to-end" intelligent fault diagnosis from data acquisition to the bearing health status category;

[0021] The present invention combines the ability of the improved time-domain convolutional network to extract fault features and the ability of the capsule network to mine information with vector input and vector output, and realizes the fault diagnosis of train wheel set bearings;

[0022] Compared with traditional deep learning methods, this method combines the strategies of contrast learning and two-stage training models, enabling the present invention to achieve high diagnostic accuracy for train wheel set bearings under high noise. Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0024] Figure 1 It is a flowchart of a contrast capsule network method for intelligent diagnosis of train wheel set bearings under high noise;

[0025] Figure 2 It is a working principle diagram of a contrast capsule network method for intelligent diagnosis of train wheel set bearings under high noise. Among them, ECA represents an efficient channel attention mechanism; causal dilation 1-dimensional convolution (Causal dilation 1dimensionalconvolution, DCC1D) is denoted as DCC1D(m, d) in the figure, where m represents the convolution kernel size and d refers to the number of intervals of the convolution kernel; GAP represents a global average pooling layer;

[0026] Figure 3 The improved time-domain convolutional network described in the embodiments;

[0027] Figure 4 The structural diagram of the comparative capsule network model described in the embodiments;

[0028] Figure 5 The graph of the accuracy of the classification training set and validation set of the comparative capsule network model changing with the number of iterations in the embodiments;

[0029] Figure 6 The classification confusion matrix graph of the comparative capsule network model in the embodiments. Detailed implementation manners

[0030] In order to better understand the technical content, features and advantages of the present invention, specific embodiments are provided below, and the present invention will be further described in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and do not limit the scope of the present invention.

[0031] See Figure 1-2 , the present invention proposes a comparative capsule network method for intelligent diagnosis of train wheel pair bearings under high noise, including the following steps:

[0032] S1. Collect the one-dimensional signal sequence data collected by the sensor according to the specific categories of train wheel pair bearings, preprocess the data, and establish a train wheel pair bearing health status database. The train wheel pair bearing health status database consists of several pairs of acceleration time series data and health status labels;

[0033] In the present invention, a certain electric locomotive wheel pair bearing is selected for experiments on a test bench. The model of the experimental bearing is 552732QT. The test bench includes a hydraulic motor, two support pads, a bearing to be detected with a hydraulic cylinder loaded on the outer ring, a hydraulic radial load application system, and a tachometer for measuring the axial speed. The bearing is installed in a mechanical system driven by a hydraulic motor. An accelerometer for measuring its vibration is installed on the load module near the outer ring of the test bearing.

[0034] To verify the effectiveness and applicability of the proposed contrast capsule network for the intelligent fault diagnosis method of train wheel pair bearings under high noise, experimental verification is carried out using the sample data of the health states of 8 types of faulty bearings in the train wheel pair bearing state database sampled by the above experimental device. The 8 types of health states are normal state, rolling element fault, inner race fault, outer race fault, outer race + inner race compound fault, inner race + rolling element compound fault, outer race + rolling element compound fault, and inner race + outer race + rolling element compound fault. The type of sensor signal collected in the experiment is vibration acceleration signal, the sampling frequency is 12 kHz, and here the experiment is carried out under the condition of a signal-to-noise ratio of -6 dB, and the length of the collected sample sequence is 1024. In the embodiment, the data preprocessing method is to normalize the collected data.

[0035] S2. Based on the train wheel pair bearing health state database, train the contrast capsule feature extraction model to obtain the trained contrast capsule feature extraction model. Use the data in the bearing health state database as the training set. The training set is used as the model input, the model loss function is the supervised contrast learning loss function, and the Adam optimization algorithm is selected to train the model. After the model converges, the contrast capsule feature extraction model can be obtained. The contrast capsule feature extraction model includes two parts: an improved time-domain convolutional network and a primary capsule layer;

[0036] The contrast capsule feature extraction model f selected in the embodiment θ See Figure 3 , the feature z obtained by the feature extraction network is z = f θ (S), I represents the number of batches in the training process, S represents the sample set corresponding to this batch, and s represents a certain sample in this set, which we call the anchor sample here. For an anchor sample s, the same-class samples and different-class samples are P s and N s . Here, the features extracted from the anchor sample, the same-class samples, and the different-class samples are z s , z r , z n , z p also represents the vector of the same-class samples as the anchor sample after feature extraction by the feature network. L csl means obtaining the similarity between z s and z p in the feature space. The temperature τ is used to adjust the concentration degree of the features in the feature space.

[0037]

[0038] Based on the above database, the number of samples in the training set of the embodiment is 640. Here, the number of capsules in the main capsule layer corresponds to 32, and the number of neurons in one capsule is 16. In the corresponding embodiment, for the structural diagram of the improved temporal convolutional network in step S2, see Figure 3 .

[0039] The improved temporal convolutional network is composed of multiple multi-scale residual blocks. The number of residual blocks in the embodiment is 4. Each residual block contains 3 Inception units and causal dilation 1-dimensional convolution (DCC1D). Figure 3 It is denoted as DCC1D(m, d) in Figure 3 . m represents the size of the convolutional kernel, and d refers to the number of intervals of the convolutional kernel. In

[0040] (c), convolutional kernels of multiple different sizes extract multi-scale features. Finally, feature depth fusion is performed on features of different scales. Features of different channels complete information interaction between channels through an efficient channel attention mechanism (ECA). The size of the convolutional kernel in the efficient channel attention mechanism module is adaptive, and the adaptation rate is Here, k represents the size of the convolutional kernel of the efficient channel attention mechanism, C is the number of channels after feature depth fusion, and b and r are hyperparameters.

[0041]

[0042] where, T k has two values, 0 and 1. If T k = 1, it means that the true category of the health state of the test sample wheel pair bearing is k. Conversely, if T k = 0; m + represents the margin upper limit, which is taken as 0.9 here; m - represents the margin lower limit, which is taken as 0.1 here; ||v k || 2 represents the norm of the output vector v k of the Length layer.

[0043] Here, the number of capsules in the digital capsule layer corresponds to the number of bearing health status categories (in other words, the number of capsules is 8), and the number of neurons in one capsule is 16. Here, the Adam optimization algorithm is selected as the optimization algorithm.

[0044] In the corresponding embodiment, the structure diagrams of the contrast capsule feature extraction network model and the contrast capsule network model in steps S2 and S3 are shown in Figure 4 .

[0045] S4. Using the train wheel pair bearing health status data obtained during the actual diagnosis process as input, the intelligent diagnosis result of the train wheel pair bearing fault can be obtained through the trained contrast capsule network.

[0046] Correspondingly, Figure 5 is the graph showing the change of the accuracy of the contrast capsule model in the training set and the validation set with the number of iterations in the embodiment. It can be seen from the results that the intelligent diagnosis accuracy of the train wheel pair bearing fault in the validation set reaches 85% in the last batch.

[0047] Correspondingly, Figure 6 is the confusion matrix of the validation set in the embodiment. The confusion matrix, also known as the error matrix, is a standard format for representing accuracy evaluation and is represented in the form of an 8-row and 8-column matrix. Each column of the confusion matrix represents the true category, and the total number of each column represents the true number of data instances of this category; each row represents the predicted category of the data, and the total number of data in each row represents the number of data instances predicted as this category.

[0048] Figure 6 The numbers 1 to 8 in [[ ]] respectively represent normal state, rolling element fault, inner ring fault, outer ring fault, outer ring + inner ring compound fault, inner ring + rolling element compound fault, outer ring + rolling element compound fault, inner ring + outer ring + rolling element compound fault. The confusion matrix shows that the fault diagnosis of each health state in the embodiment still achieves good results under high noise (signal-to-noise ratio -6dB), so the method of the present invention is effective.

Claims

1. A comparative capsule network method for intelligent diagnosis of train wheel pair bearings under high noise, comprising the following steps: S1. Collect the mechanical equipment health state data collected by sensors according to the specific categories of train wheel pair bearings, preprocess the data, and establish a train wheel pair bearing health state database; The sensor type is one or more of a vibration displacement sensor, a vibration velocity sensor, a vibration acceleration sensor, or a sound signal sensor; The data in the state database includes: normal state, single fault state, and compound fault state; the single fault state includes: rolling element fault, inner ring fault, outer ring fault; the compound fault state includes: outer ring + inner ring compound fault, inner ring + rolling element compound fault, outer ring + rolling element compound fault, inner ring + outer ring + rolling element compound fault; each piece of data with the same health state has the same state label; S2. Based on the train wheel pair bearing health state database, train a comparative capsule feature extraction model to obtain a trained comparative capsule feature extraction model; the loss function of the comparative capsule feature extraction model is a supervised contrast learning loss function, select an optimization algorithm to train the model until convergence, and the comparative capsule feature extraction model includes two parts: an improved time-domain convolutional network and a primary capsule layer; The comparative capsule feature extraction model includes an improved time-domain convolutional network and a primary capsule layer connected in sequence. The time-domain convolutional network includes an input module, n multi-scale residual blocks, and an output module connected in sequence; the multi-scale residual block includes two inputs and two outputs. Among them, input 1 passes through two Inception units, a BN+ReLU+Dropout module, and a DCCID(1,1) layer in sequence; here, DCCID is the abbreviation of causal dilated 1D convolution, where parameter 1 refers to the convolutional kernel size and parameter 2 refers to the causal dilation rate; the output of the DCCID(1,1) layer in the multi-scale residual block is added to the output of input 2 of the multi-scale residual block passing through a max pooling layer and an Inception unit in sequence. One path is used as the output 2 of the multi-scale residual block, and the other path passes through a BN+ReLU+Dropout module and is used as the output 1 of the multi-scale residual block; the input 1 and input 2 of the first multi-scale residual block of the comparative capsule feature extraction model are both the input of the comparative capsule feature extraction model. The previous output 1 of the subsequent multi-scale residual blocks corresponds to the next input 1, and the previous output 2 corresponds to the next input 2; the output 1 of the last multi-scale residual block is used as the output of the comparative capsule feature extraction model; After the input of the Inception unit, it is first divided into four paths. The first path includes a DCCID(1,1) layer, the second path includes a sequentially connected DCCID(3,3) layer and a DCCID(1,1) layer, the third path includes a sequentially connected DCCID(5,2) layer and a DCCID(1,1) layer, and the fourth path includes a sequentially connected DCCID(1,1) layer and a DCCID(3,2) layer. Then, the outputs of the four paths pass through an efficient channel attention mechanism module, which is the output of the Inception unit. The efficient channel attention mechanism first performs deep fusion on it, and then completes the cross-channel information interaction of the features of different channels. S3. Freeze the weights of the above-mentioned ratio capsule feature extraction model, and add a digital capsule layer and a Length layer after the ratio capsule feature extraction model to obtain a contrast capsule network model. The parameter update between the main capsule layer and the digital capsule layer uses a dynamic routing mechanism, and the Length layer is used for the classification of the health status of the train wheel pair bearing. The loss function of the contrast capsule network model is the margin loss function. The contrast capsule network model is trained on the training set until convergence to obtain the contrast capsule network model. S4. In the actual diagnosis process, obtain the health status data of the train wheel pair bearing, and input it into the contrast capsule network model for status judgment after preprocessing.

2. A contrast capsule network method for intelligent diagnosis of train wheel pair bearings under high noise as described in claim 1, characterized in that, the data preprocessing method in step S1 includes: one of data normalization, data standardization, wavelet packet transform noise reduction, and ensemble empirical mode decomposition noise reduction.

3. A contrast capsule network method for intelligent diagnosis of train wheel pair bearings under high noise as described in claim 1, characterized in that, the training method of the contrast capsule feature extraction model in step 2 includes: using one of the Adam and SGD optimization algorithms as the model optimizer; using one of learning rate decay and cosine learning rate change as the model training strategy.

Images