Train wheel axle fault acoustic emission detection method based on TCN network

Abstract

The application discloses a train wheel axle fault acoustic emission detection method based on a TCN network, and the method comprises the following steps: constructing a TCN network; inputting a large number of samples into the TCN network for network training, wherein the samples comprise acoustic emission signals, wheel axle rotating speeds and corresponding classification labels; the TCN network extracts features from the acoustic emission signals, considers the correlation between the acoustic emission signals and the wheel axle rotating speeds, combines the learned features with the wheel axle rotating speeds to further learn features, outputs a classification result, continuously performs network training, stops the training when a predetermined training frequency is reached or a loss function value of the network is the minimum, and obtains a trained TCN network model; inputting real-time detected acoustic emission signals and corresponding wheel axle rotating speeds into the trained TCN network model to obtain a train wheel axle fault detection result; and the application has the advantages that data information is fully utilized, and information loss caused by manual feature selection is avoided.

Description

Technical Field

[0001] This invention relates to the field of train wheel and axle fault detection, and more specifically to a method for detecting train wheel and axle fault acoustic emission based on a TCN network. Background Technology

[0002] Trains operate for extended periods, with their wheel axles subjected to constant mechanical rotation and immense loads, making them highly susceptible to wear and tear, and even malfunction. This can potentially lead to major accidents and significant economic losses. Therefore, timely detection of wheel axle wear and malfunctions, along with proper maintenance and replacement, can effectively prevent train accidents. Timely detection of train wheel axle malfunctions is of paramount importance.

[0003] Currently, train wheel and axle fault diagnosis methods are mainly divided into two categories. One is the contact-based fault diagnosis method based on vibration signals, and the other is the non-contact diagnostic method based on acoustic characteristics. Vibration signal-based wheel and axle fault diagnosis has a long history and is relatively mature, but its applicability is limited due to the low frequency range of vibration signals and the necessity of contact. Acoustic characteristic detection, being a non-contact method, has significant advantages in deployment cost and applicability. Acoustic wheel and axle fault diagnosis can be divided into acoustic emission diagnostic technology and acoustic signal diagnostic technology according to the sound emission mechanism. Acoustic signal diagnostic technology refers to using external force to strike the wheel and axle to diagnose the resulting acoustic emission. This method is difficult to detect early-stage faults and is not suitable for real-time online applications. Emission detection technology is a dynamic non-destructive testing method; its signal originates from the defect itself. Therefore, acoustic emission detection can determine the activity and severity of the defect. The literature "Tandon N, Nakra BC. Defect detection in rolling element bearings by acoustic emission method. 1990." proposes using the peak amplitude and acoustic emission count of the acoustic emission signal as two characteristic parameters for fault diagnosis. The paper "Wang Yanyan. Research on Fault Diagnosis of Rolling Bearings for Heavy-Duty Railway Freight Cars Based on Acoustic Emission Technology. Master's Thesis, Central South University, 2013" uses wavelet packet decomposition to extract fault feature information from acoustic emission signals. The paper "Li Hua, Xu Chunguang, Xiao Dingguo, et al. Reference Title for Papers at China Bearing Forum: Acoustic Emission Detection Technology of Rolling Bearings [J]. Bearings, 2002(7):4" only uses short-time root mean square parameters for feature identification of acoustic emission signals. There are many similar methods, but these methods mainly rely on manual methods to find fault features, which easily leads to information loss. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the existing methods for detecting acoustic emission of train wheel axle faults mainly rely on manual methods to find fault characteristics, which easily leads to information loss.

[0005] This invention solves the above-mentioned technical problems through the following technical means: a train wheel axle fault acoustic emission detection method based on TCN network, the method comprising:

[0006] Construct a TCN network;

[0007] A large number of samples are input into the TCN network for network training. The samples include acoustic emission signals, wheel and axle rotation speeds and corresponding classification labels.

[0008] The TCN network extracts features from acoustic emission signals and considers the correlation between acoustic emission signals and wheel rotation speed. It combines the learned features with wheel rotation speed for further feature learning, outputs classification results, and continuously trains the network. Training stops when a predetermined number of training iterations is reached or the network's loss function value is minimized, resulting in a well-trained TCN network model.

[0009] The real-time detected acoustic emission signals and the corresponding wheel and axle speeds are input into the trained TCN network model to obtain the train wheel and axle fault detection results.

[0010] This invention uses a TCN network to extract features from acoustic emission signals and considers the correlation between acoustic emission signals and wheel rotation speed. The learned features are combined with wheel rotation speed for further feature learning. Compared with traditional non-machine learning methods, this invention drives the TCN network to perform feature learning through a large number of samples, making full use of data information and avoiding information loss caused by manual feature selection.

[0011] Furthermore, the TCN network includes a first TCN module, a second TCN module, a feature fusion module, a third TCN module, a fully connected layer, and a softmax layer connected in sequence. The acoustic emission signal is input to the input terminal of the first TCN module, and the corresponding wheel axle rotation speed is input to the input terminal of the feature fusion module.

[0012] Furthermore, the first TCN module, the second TCN module, and the third TCN module have the same structure. The first TCN module includes three sequentially cascaded convolutional units, a 1×1 convolutional unit, and a ReLU activation function unit. The input of the first convolutional unit in the three sequentially cascaded convolutional units is connected to the input of the 1×1 convolutional unit and receives input data. The output of the third convolutional unit in the three sequentially cascaded convolutional units, the output of the 1×1 convolutional unit, and the input of the ReLU activation function unit are connected. The output of the ReLU activation function unit outputs data.

[0013] Furthermore, the 1×1 convolutional unit is a 1×1 convolutional layer, and the ReLU activation function unit is a ReLU activation function layer.

[0014] Furthermore, the convolutional unit includes a dilated convolutional layer, a batch normalization (BN) layer, a ReLU layer, and a dropout layer, which are connected sequentially.

[0015] Furthermore, the hollow convolutional layer is formed by stacking two causal convolutional layers, with a kernel size of 2.

[0016] Furthermore, the output sequence of the second TCN module is cross-combined with the wheel axle speed sequence to form a new feature sequence, which is used as the input of the feature fusion module. The cross-combination method is that the sequence input to the second TCN module is an odd sequence of the output sequence, and the wheel axle speed sequence is an even sequence of the output sequence.

[0017] Furthermore, the sampling rate of the wheel axle rotation speed data satisfies the requirement that the number of sampling points within time T is consistent with the length of the feature sequence output by the second TCN module.

[0018] Furthermore, the acoustic emission signal is input to the first TCN module before a data preprocessing step is included.

[0019] Furthermore, the data preprocessing steps include:

[0020] The audio signal is sampled at a sampling rate of 24kHz. N sampling points are extracted from the input audio data stream as a single frame of input data. The sampling time for the N sampling points is denoted as T, which is the recording time of one frame of data. The mean and variance of these N data points are calculated, and the data frame is standardized as follows:

[0021]

[0022] Where s is the input sequence, m is the mean, and σ is the standard deviation of the sequence.

[0023] The advantages of this invention are:

[0024] (1) This invention uses a TCN network to extract features from acoustic emission signals and considers the correlation between acoustic emission signals and wheel rotation speed. The learned features are combined with wheel rotation speed for further feature learning. Compared with traditional non-machine learning methods, this invention drives the TCN network to perform feature learning through a large number of samples, making full use of data information and avoiding information loss caused by manual feature selection.

[0025] (2) Existing wheel axle acoustic emission fault diagnosis using deep neural networks usually uses the time-frequency graph of the signal as the network input data. However, the time-frequency graph is a general algorithm and is not entirely driven by data to form features, which will also cause information loss. The method proposed in this invention directly extracts features from the acoustic emission time series, which is entirely driven by data to form features and will not cause information loss.

[0026] (3) This invention integrates the wheel axle speed information with the intermediate abstract features of the acoustic emission signal, and then performs the next step of feature extraction. It fully considers the relationship between wheel axle speed and acoustic emission signal, which is more conducive to the mining of fault features and can more effectively extract fault features and improve recognition efficiency. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the train wheel axle fault acoustic emission detection method based on TCN network disclosed in the embodiments of the present invention;

[0028] Figure 2 This is a diagram showing the overall structure of the TCN network in the TCN network-based acoustic emission detection method for train wheel and axle faults disclosed in this embodiment of the invention.

[0029] Figure 3 This is a structural diagram of the TCN module in the train wheel axle fault acoustic emission detection method based on TCN network disclosed in the embodiments of the present invention;

[0030] Figure 4 The diagram shows the void convolutional layer structure in the TCN network-based acoustic emission detection method for train wheel and axle faults disclosed in this embodiment of the invention.

[0031] Figure 5 This is a schematic diagram of the feature fusion module in the TCN network-based acoustic emission detection method for train wheel and axle faults disclosed in this embodiment of the invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] like Figure 1 As shown, the method for detecting acoustic emission of train wheel axle faults based on TCN network is divided into two parts: data preprocessing and TCN network, which will be described in detail below.

[0034] First, data preprocessing uses a 24kHz sampling rate to sample the audio signal. N sampling points are extracted from the input audio data stream as a single frame of input data. The sampling time for these N sampling points is denoted as T, which is the recording time for one frame of data. The mean and variance of these N data points are calculated, and the data frame is standardized as follows:

[0035]

[0036] Where s is the input time series, specifically referring to the data sequence of N sampling points with sampling time identifiers, m is the mean, and σ is the standard deviation of the sequence.

[0037] Then, the TCN network is constructed; the overall structure of the TCN network is as follows: Figure 2 As shown. The TCN network includes a first TCN module, a second TCN module, a feature fusion module, a third TCN module, a fully connected layer, and a softmax layer connected in sequence. The acoustic emission signal is input to the input terminal of the first TCN module, and the corresponding wheel axle rotation speed is input to the input terminal of the feature fusion module. Figure 2 The input audio signal is the output of the data preprocessing result. The sampling rate of the wheel axle speed data only needs to satisfy the requirement that the number of sampling points within time T is consistent with the length of the feature sequence output by the second TCN module. The first TCN module, the second TCN module, and the third TCN module correspond to... Figure 2 The TCN module 1, TCN module 2, and TCN module 3 are included.

[0038] like Figure 3 As shown, the first, second, and third TCN modules have the same structure. The first TCN module includes three sequentially cascaded convolutional units, a 1×1 convolutional unit, and a ReLU activation function unit. The input of the first convolutional unit is connected to the input of the 1×1 convolutional unit to receive input data. The output of the third convolutional unit is connected to the output of the 1×1 convolutional unit and the input of the ReLU activation function unit. The output of the ReLU activation function unit outputs data. The 1×1 convolutional unit is a 1×1 convolutional layer, and the ReLU activation function unit is a ReLU activation function layer. The convolutional unit includes a dilated convolutional layer, a batch normalization (BN) layer, a ReLU layer, and a dropout layer, which are sequentially connected. Figure 4As shown, the dilated convolutional layer is composed of two layers of causal convolution stacked together, with a kernel size of 2. The BN layer, ReLU layer, and Dropout layer are conventional batch normalization layer, ReLU activation function layer, and dropout layer, respectively. The 1x1 convolutional layer is directly connected to the ReLU activation function layer at the module output, forming a residual connection, which can prevent network degradation.

[0039] The structure of the feature fusion module is as follows: Figure 5 As shown, the output sequence of the second TCN module is cross-combined with the wheel axle speed sequence to form a new feature sequence, which is used as the input of the feature fusion module. The cross-combination method is that the sequence input to the second TCN module is an odd sequence of the output sequence, and the wheel axle speed sequence is an even sequence of the output sequence.

[0040] The specific execution process of the method of the present invention is as follows: A labeled dataset with a sampling rate of 24 kHz and a length of N points is normalized to obtain a large number of samples. These samples are then input into a TCN network for network training. The samples include acoustic emission signals, wheel and axle speeds, and corresponding classification labels. The TCN network extracts features from the acoustic emission signals and considers the correlation between the acoustic emission signals and wheel and axle speeds. The learned features are combined with the wheel and axle speeds for further feature learning, and classification results are output. The network is continuously trained until a predetermined number of training iterations or the minimum value of the network's loss function is reached, at which point training stops, resulting in a trained TCN network model. The real-time detected acoustic emission signals and corresponding wheel and axle speeds are then input into the trained TCN network model to obtain the train wheel and axle fault detection results.

[0041] Through the above technical solutions, this invention uses a TCN network to extract features from acoustic emission signals and considers the correlation between acoustic emission signals and wheel rotation speed. The learned features are combined with wheel rotation speed for further feature learning. Compared with traditional non-machine learning methods, this invention drives the TCN network to perform feature learning through a large number of samples, making full use of data information and avoiding information loss caused by manual feature selection.

[0042] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting acoustic emission of train wheel axle faults based on TCN networks, characterized in that, The method includes: A TCN network is constructed; the TCN network includes a first TCN module, a second TCN module, a feature fusion module, a third TCN module, a fully connected layer, and a softmax layer connected in sequence. The acoustic emission signal is input to the input terminal of the first TCN module, and the corresponding wheel axle rotation speed is input to the input terminal of the feature fusion module. The first, second, and third TCN modules have the same structure. The first TCN module includes three sequentially cascaded convolutional units, a 1×1 convolutional unit, and a ReLU activation function unit. The input terminal of the first convolutional unit in the three sequentially cascaded convolutional units is connected to the input terminal of the 1×1 convolutional unit and receives the input signal. According to the diagram, the output of the third convolutional unit in three sequentially cascaded convolutional units, the output of the 1×1 convolutional unit, and the input of the ReLU activation function unit are connected, with the output of the ReLU activation function unit outputting data. The output sequence of the second TCN module is cross-combined with the wheel axle speed sequence to form a new feature sequence, which serves as the input to the feature fusion module. The cross-combination method is that the sequence input to the second TCN module is an odd-numbered sequence of the output sequence, and the wheel axle speed sequence is an even-numbered sequence of the output sequence. The convolutional unit includes a dilated convolutional layer, a BN layer, a ReLU layer, and a Dropout layer, which are sequentially connected. A large number of samples are input into the TCN network for network training. The samples include acoustic emission signals, wheel and axle rotation speeds and corresponding classification labels. The TCN network extracts features from acoustic emission signals and considers the correlation between acoustic emission signals and wheel rotation speed. It combines the learned features with wheel rotation speed for further feature learning, outputs classification results, and continuously trains the network. Training stops when a predetermined number of training iterations is reached or the network's loss function value is minimized, resulting in a well-trained TCN network model. The real-time detected acoustic emission signals and the corresponding wheel and axle speeds are input into the trained TCN network model to obtain the train wheel and axle fault detection results.

2. The method for detecting acoustic emission of train wheel axle faults based on TCN network according to claim 1, characterized in that, The 1×1 convolutional unit is a 1×1 convolutional layer, and the ReLU activation function unit is a ReLU activation function layer.

3. The method for detecting acoustic emission of train wheel axle faults based on TCN network according to claim 1, characterized in that, The hollow convolutional layer is formed by stacking two causal convolutional layers, with a kernel size of 2.

4. The method for detecting acoustic emission of train wheel axle faults based on TCN network according to claim 1, characterized in that, The sampling rate of the wheel axle rotation speed data satisfies the requirement that the number of sampling points within time T is consistent with the length of the feature sequence output by the second TCN module.

5. The method for detecting acoustic emission of train wheel axle faults based on TCN network according to claim 1, characterized in that, The acoustic emission signal is input to the first TCN module before a data preprocessing step is included.

6. The method for detecting acoustic emission of train wheel axle faults based on TCN network according to claim 5, characterized in that, The data preprocessing steps include: The audio signal is sampled at a sampling rate of 24kHz. N sampling points are extracted from the input audio data stream as a single frame of input data. The sampling time for the N sampling points is denoted as T, which is the recording time of one frame of data. The mean and variance of these N data points are calculated, and the data frame is standardized as follows: Where s is the input sequence, and m is the mean. denoted as the standard deviation of the sequence.

Images