A method for acoustic emission signal recognition and a network model training method

Abstract

This invention discloses a method for acoustic emission signal recognition, comprising the following steps: collecting acoustic emission signal data generated during axle operation; preprocessing the collected signal data; importing the preprocessed signal data into a network recognition module, and using a pre-trained improved one-dimensional CNN network model in the network module to identify and analyze the imported signal data; and outputting the recognition result. This invention also discloses a model training method for training the improved one-dimensional CNN network model. This invention is applicable to the field of train fault diagnosis, enabling the identification of acoustic emission signals from train axles after acquisition and real-time monitoring during axle operation. It can extract higher-level signal features from the original acoustic emission signals through a deep network architecture and possesses good fault identification capabilities.

Description

Technical Field

[0001] This invention relates to the field of train fault diagnosis, and more specifically, to a method for identifying acoustic emission signals from train axles after acquisition, and a method for training the network model used therein. Background Technology

[0002] Railways, as a highly economical and massive mode of transportation, occupy a pivotal position in the national transportation system and are a vital national infrastructure for people's livelihood. Therefore, ensuring the safe and stable operation of trains is of paramount importance in railway transportation. Among all the components of a rail vehicle, the bogie is an extremely important part. The operating condition of the bogie is one of the key factors affecting the safe operation of the vehicle. Due to the complex working environment of the bogie, the probability of failure of its critical components is greatly increased. As a crucial component of the bogie, if the axle fails due to cracks or excessive wear, and this failure is not detected and repaired in a timely manner, it will lead to irreparable and serious consequences. Therefore, real-time monitoring of the axle's operating condition during train operation and accurate detection of axle faults during train shutdowns for maintenance are crucial to ensuring the normal operation of trains.

[0003] However, with the rapid increase in rail transit mileage and operating time, we often encounter faults with varying sizes, locations, and degrees of damage during actual inspections. Furthermore, the shapes of the inspected objects and the inspection environment also differ, posing significant challenges to traditional fault detection methods. Acoustic emission detection technology, combined with artificial neural networks, overcomes these obstacles with its unique detection principle. It enables real-time monitoring of train axle operation and comprehensive inspection and repair during maintenance, ensuring safe and stable train operation, extending axle service life, and reducing maintenance costs. Therefore, research on intelligent fault diagnosis for train axles is particularly important. Summary of the Invention

[0004] This invention provides an acoustic emission signal recognition method and a network model training method to solve the technical problem that traditional methods for monitoring fatigue cracks in train axles are difficult to implement in real time due to differences in the size, location, and degree of damage caused by the fault, as well as the shape of the object being tested and the testing environment.

[0005] To achieve the above objectives, the present invention provides a method for acoustic emission signal identification, the method comprising the following steps:

[0006] Step S1: Collect acoustic emission signal data generated during the operation of the axle;

[0007] Step S2: Perform preprocessing operations on the acquired signal data;

[0008] Step S3: Import the preprocessed signal data into the network recognition module, and use the improved one-dimensional CNN network model that has been trained in the network module to recognize and analyze the imported signal data.

[0009] Step S4: Output the recognition result. If the imported signal data is recognized as "axle fatigue crack signal feature", the output result is "0"; if the imported signal data is recognized as "interference signal feature", the output result is "1".

[0010] In the preferred embodiment of the acoustic emission signal identification method described above, step S2 involves preprocessing the acquired signal data, including dividing the signal data into segments of 1024 data points.

[0011] In the preferred embodiment of the acoustic emission signal recognition method described above, in step S3, the improved one-dimensional CNN network model includes a one-dimensional convolutional pooling module, a residual module, a global average pooling layer, and an output layer; the one-dimensional convolutional pooling module includes two convolutional layers and two pooling layers, one of which is a single-layer convolutional layer with eight 23*1 kernels, and the other is a double-layer convolutional layer with sixteen 12*1 kernels, and both pooling layers are 2*1 in size; the residual module consists of ten sequentially connected two-layer residual blocks.

[0012] In the preferred embodiment of the acoustic emission signal recognition method described above, step S3 involves using a pre-trained improved one-dimensional CNN network model in the network module to identify and analyze the imported signal data. This specifically includes the following steps:

[0013] Step S31: In one layer of the convolution kernel of the one-dimensional convolution pooling module, a one-dimensional convolution operation is performed on the imported signal data to obtain one layer of one-dimensional convolution feature data.

[0014] Step S32: In the first pooling layer of the one-dimensional convolutional pooling module, mean pooling is performed on the one-dimensional convolutional feature data of the layer, and the pooled data is output.

[0015] Step S33: In the second-layer convolution kernel of the one-dimensional convolution pooling module, perform another one-dimensional convolution operation on the imported data after the first-layer pooling to obtain the two-layer one-dimensional convolution feature data.

[0016] Step S34: In the second pooling layer of the one-dimensional convolutional pooling module, perform another mean pooling operation on the two-layer one-dimensional convolutional feature data, and output the data after the second pooling layer.

[0017] Step S35: Input the pooled data into the residual module. The ten sequentially connected two-layer residual blocks process the pooled data in sequence and output the overall output value.

[0018] Step S36: The overall output value enters the global average pooling layer for average pooling;

[0019] Step S37: The data after average pooling is output as the final result by the output layer;

[0020] In the preferred embodiment of the acoustic emission signal recognition method described above, the one-dimensional convolution operation in steps S31 and S33 includes the following steps:

[0021] Step S311: Input the signal data into the convolutional kernel of the first layer;

[0022] In step S312, a 3*3 one-dimensional convolutional kernel slides along the same length in one direction in the input one-dimensional signal data. After each slide, the values ​​of the overlapping part of the one-dimensional convolutional kernel and the one-dimensional signal data are multiplied and then added together to obtain the feature data.

[0023] The feature data is composed of multiple neurons, wherein the calculation formula for the j-th neuron in the i-th frame of the (l+1)-th layer is:

[0024]

[0025] in, X represents the weights and bias parameters of the i-th convolutional kernel in the l-th layer. (l) (j) represents the j-th local region in the l-th layer. This represents the j-th neuron in the i-th frame of layer l+1.

[0026] In the preferred embodiment of the acoustic emission signal identification method described above, the mean pooling operation in steps S32 and S34 is as follows:

[0027] The output feature data is subjected to mean pooling. A 2*1 pooling layer slides sequentially across the feature data in one direction with a specified step size. The output is the average of all elements in the overlapping portion of the pooling layer and the feature data, i.e., the pooled data. The calculation formula is as follows:

[0028]

[0029] In the formula Let t be the value of the t-th neuron in the i-th frame of the l-th layer, where t ∈ [(j-1)w+1, jw] and w is the width of the pooling layer's receptive region. This represents the value corresponding to the neuron in the (l+1)th layer of the pooling operation.

[0030] In the preferred embodiment of the acoustic emission signal identification method described above, step S35 specifically involves the following steps:

[0031] When the input and output data of the two residual blocks have the same dimension, the pooled data, after being input into the two residual blocks, will simultaneously pass through both the direct connection and the shortcut connection. The data in the direct connection will undergo convolution calculations by the convolutional layer and be activated by a linear rectification function. The shortcut connection will directly send the input data to the end of the residual block, adding it to the output of the direct connection to obtain the overall output value of the residual block. Its mapping function is expressed as:

[0032] H(x)=F(x)+x

[0033] Where H(x) represents the overall output value, F(x) represents the residual mapping function, and x represents the input value;

[0034] When F(x) = 0, the optimal mapping solution H(x) = x is obtained;

[0035] The mathematical expression for the linear rectifier function is:

[0036]

[0037] In the ten sequentially connected two-layer residual blocks of the residual module, the input, output, and parameter of the l-th two-layer residual block are x, respectively. l f(y) l ) and W l Then we get the formula:

[0038] y l =h(x l )+F(x l W l )

[0039] x l+1 =f(y l )

[0040] Where F(·) is the residual function, and f(·) represents the linear rectified function; y l This represents the output data of the l-th two-layer residual block before the linear rectified function is input; f(y) l h(x) represents the linear rectified function of the output of the l-th two-layer residual block, and is also the final output of the l-th two-layer residual block; l ) represents the mapping function of the input of the l-th two-layer residual block;

[0041] When h(x) l )=x l ,f(yl )=y l When, formula x l+1 =f(y l This can be transformed into:

[0042] x l+1 =x l +F(x l W l )

[0043] Then, the output of the Lth residual block can be derived using a recursive iteration method:

[0044]

[0045] Where xi represents the input of the i-th two-layer residual block, x l This represents the input to the l-th two-layer residual block;

[0046] x L+1 It represents the input of the (L+1)th two-layer residual block, and is also equal to the output value of the Lth two-layer residual block;

[0047] When the dimensions of the input and output data of the residual block are different, the pooled data is input into the two layers of residual blocks and passes through both direct connection and shortcut connection simultaneously. The data in the direct connection undergoes convolution calculation by the convolutional layer and is activated by the linear rectification function. The shortcut connection uses a 1*1 convolutional kernel to perform a dimensionality increase operation on the data, so that the data after the dimensionality increase operation has the same dimension as the output data of the residual block. The data after the dimensionality increase operation is added to the output result of the direct connection to obtain the overall output value of the residual block.

[0048] In the preferred embodiment of the acoustic emission signal recognition method described above, the ten sequentially connected two-layer residual blocks have different dimensions for input and output data in the first, third, sixth, and ninth residual blocks, and their shortcut connections use the 1*1 convolution kernel to perform dimensionality upscaling on the data; while the second, fourth, fifth, seventh, eighth, and tenth residual blocks have the same dimension for input and output data, and their shortcut connections directly transmit the input data to the tail of the residual block.

[0049] This invention also provides a network model training method, comprising the following steps:

[0050] Step S101: Label the two preprocessed signal datasets, with the axle fatigue crack signal set to "0" and the interference signal set to "1".

[0051] Step S102: Randomly shuffle all the set data, and randomly select 70% of the overall shuffled data as the training dataset and 30% of the overall shuffled data as the test dataset.

[0052] Step S103: Import the training dataset into the constructed improved one-dimensional CNN network model for network model training;

[0053] The training in step S103 specifically includes the following steps:

[0054] Step S1031: Selection of network optimizer. While ensuring that other network parameters are consistent, five optimizers, ASDG, Adagrad, RMSprop, Adadelat, and Adam, are selected for comparative experiments. The network optimizer is selected based on the accuracy of the network model in recognizing the training dataset as the number of network iterations increases. The network optimizer with the highest recognition accuracy is selected as the network optimizer of the improved one-dimensional CNN network model. Finally, Adam is selected as the network optimizer of the improved one-dimensional CNN network model.

[0055] Step S1032: Selection of batch size. Under the premise of ensuring that the parameters of other network models are consistent, comparative experiments are conducted from four batch sizes: 16, 32, 64, and 128. Based on the recognition rate and running time cost of the network model as the number of iterations increases, the batch size of the improved one-dimensional CNN network model is selected. Finally, the batch size of the improved one-dimensional CNN network model is determined to be 32.

[0056] Step S1033: Selection of learning rate. While ensuring that the parameters of other network models are consistent, network models with learning rates of 0.01, 0.001, and 0.0001 are compared. Based on the recognition rate of the network model as the number of iterations increases, the learning rate of the network model with the highest recognition rate is selected as the learning rate of the improved one-dimensional CNN network model. Finally, the learning rate of the improved one-dimensional CNN network model is determined to be 0.0001.

[0057] Step S1034: Select Adam, 32, and 0.0001 for the network optimizer, batch size, and learning rate of the improved one-dimensional CNN network model to obtain the best network performance, and save the trained network model.

[0058] Step S1035: Input the data from the test dataset into the trained improved one-dimensional CNN network model to identify and analyze the fatigue crack signal of the axle and verify the effectiveness of the model.

[0059] In the preferred embodiment of the above network model training method, the preprocessing operation in step S101 includes dividing the two sets of signal data into segments with a segmentation distance of 1024 data points.

[0060] This invention provides an acoustic emission signal recognition method and a network model training method, which can extract higher-level signal features from the original acoustic emission signal through a deep network architecture and has good fault identification capabilities. Attached Figure Description

[0061] Figure 1 This is a system hardware structure diagram of the data acquisition device in this invention;

[0062] Figure 2 This is a schematic diagram of data segmentation processing in this invention;

[0063] Figure 3 To improve the structure diagram of a one-dimensional CNN network model;

[0064] Figure 4 This is a schematic diagram of the convolution operation in a one-dimensional convolutional neural network.

[0065] Figure 5 This is a schematic diagram of the max pooling operation in a one-dimensional convolutional neural network;

[0066] Figure 6 This is a schematic diagram of the mean pooling operation in a one-dimensional convolutional neural network;

[0067] Figure 7 This is a schematic diagram of the residual block element;

[0068] Figure 8 This is a diagram illustrating the change in the dimension of the residual structure;

[0069] Figure 9 This is a diagram illustrating the recognition effects of different optimizers;

[0070] Figure 10 This is a schematic diagram illustrating the recognition effect for different batch sizes;

[0071] Figure 11 This is a diagram illustrating the recognition performance at different learning rates;

[0072] Figure 12 This is a structural diagram of the real-time monitoring module;

[0073] Figure 13 This is the flowchart of the network identification module;

[0074] Figure 14 This is a diagram of the network model selection interface;

[0075] Figure 15This is an overall flowchart of an acoustic emission signal recognition method according to the present invention;

[0076] Figure 16 It is a flowchart of an improved one-dimensional CNN network model for identifying and analyzing imported signal data;

[0077] Figure 17 This is a flowchart of the one-dimensional convolution operation in this invention;

[0078] Figure 18 This is a flowchart outlining the network model training method of the present invention.

[0079] Figure 19 This is a flowchart of importing the training dataset into the built improved one-dimensional CNN network model for training the network model.

[0080] In the diagram, 1 is a one-dimensional convolutional pooling module, 11 is a first-layer convolutional kernel, 12 is a second-layer convolutional kernel, 21 is the first pooling layer, 22 is the second pooling layer, 3 is the residual module, 31 is a two-layer residual block, 100 is the sensor, 200 is the signal amplifier, 300 is the power supply, and 400 is the data acquisition card. Detailed Implementation

[0081] This invention relates to the field of train fault diagnosis, specifically to the identification of acoustic emission signals collected from train axles and the real-time monitoring of axle operation.

[0082] The axle fatigue crack acoustic emission signal recognition system based on an improved one-dimensional convolutional neural network (CNN) mainly includes a real-time monitoring module and a network recognition module.

[0083] The acoustic emission signal recognition method is as follows: the real-time monitoring module of the recognition system collects acoustic emission signals through an external acoustic emission detection device during the axle's operation. After real-time preprocessing, the collected signal data is imported into the network recognition module. Then, the improved one-dimensional CNN network model trained in the network module is used to identify and analyze the collected acoustic emission signal data. The design process of the axle fatigue crack acoustic emission signal recognition system based on the improved one-dimensional CNN is as follows.

[0084] Data acquisition equipment design:

[0085] like Figure 1 As shown, the hardware components of the identification system mainly include a sensor 100, a signal amplifier 200, a power supply 300, and a data acquisition card 400. Figure 1The diagram shows the connection method of each component. Furthermore, the signal acquisition method and the signal transmission relationship between each component are as follows: Sensor 100 collects the vibration waves generated by the train axle and converts them into electrical signals, which are then transmitted to signal amplifier 200 via a transmission line. Signal amplifier 200 amplifies the incoming acoustic signals and transmits them through power supply 300 to acquisition card 400 via a transmission line. Power supply 300 is mainly responsible for powering sensor 100 and signal amplifier 200. The electrical signals transmitted to acquisition card 400 are converted into digital signals that can be recognized by a computer via an A / D converter.

[0086] A method for acoustic emission signal recognition based on an improved one-dimensional convolutional neural network:

[0087] This invention constructs an improved one-dimensional CNN network model in a Python environment and preprocesses the axle crack signal and interference signal datasets. Then, all preprocessed data is randomly shuffled and divided into training and testing datasets. Next, the training dataset is imported into the constructed improved one-dimensional CNN network model for training. During training, the network model parameters are adjusted based on the output training results to obtain optimal network performance, and the trained network model is saved. Finally, the test dataset data is input into the trained network model to identify and analyze axle fatigue crack signals.

[0088] Data preprocessing:

[0089] like Figure 2 As shown, the axle crack signal dataset used to build the improved one-dimensional CNN network model has 5 million data points, and the interference signal dataset has 1.43 million data points. Due to the large amount of data, directly inputting it into the network for training and testing would waste network resources and would not demonstrate good network recognition performance. Therefore, we segmented the signal data into segments of 1024 data points each. The 5 million axle crack signal data points were divided into 4942 signal segments, and the 1.43 million interference signal data points were divided into 1428 signal segments. Because the data size differs significantly between the two types of data after segmentation, to increase the dataset size and prevent excessive differences between positive and negative samples, the interference signal was segmented starting from the 256th, 512th, and 768th data points, with a segmentation distance of 1024 data points. These segmented signal segments were used as supplementary data of the same type. Figure 2 As shown in the figure, d represents the segmentation distance, a represents the original signal, and b represents the preprocessed sample signal.

[0090] Improved construction of one-dimensional CNN network models:

[0091] This invention is an identification system based on one-dimensional convolutional neural networks and residual networks. Because one-dimensional convolutional neural networks cannot extract deep-level data features when the amount of data is large, this invention improves one-dimensional convolutional neural networks using residual networks. Based on one-dimensional convolutional neural networks and residual networks, this invention proposes a method for identifying acoustic emission signals of axle fatigue cracks based on an improved one-dimensional convolutional neural network.

[0092] like Figure 3 The diagram shows the structure of the improved one-dimensional CNN network model. The network structure proposed in this invention mainly consists of a one-dimensional convolutional pooling module and a residual module. The one-dimensional convolutional pooling module 1 includes two convolutional layers. One convolutional layer has eight 23x1 convolutional kernels as the first convolutional kernel 11, and the other convolutional layer has sixteen 12x1 convolutional kernels as the second convolutional kernel 12. Both pooling layers are 2x1 in size, and are divided into a first pooling layer 21 and a second pooling layer 22. The residual module 3 is composed of ten two-layer residual blocks 31 connected end to end.

[0093] Figure 3 The residual blocks described herein, connected sequentially from top to bottom and left to right, require 1*1 convolutional kernels to perform dimensionality upscaling on the first, third, sixth, and ninth two-layer residual blocks 31. The network finally employs global average pooling and uses fully connected layers containing softmax functions and an output layer for final data output.

[0094] One-dimensional convolution operation:

[0095] One-dimensional convolution operations are as follows Figure 4 As shown, since it is a one-dimensional convolutional neural network, in the convolution calculation, the 3×3 one-dimensional convolution kernel 2 slides along the same length in one direction within the input one-dimensional data c. After each slide, the values ​​of the overlapping parts of the convolution kernel and the one-dimensional data are multiplied and then added together to obtain the final feature data e. In the convolution formula... X represents the weights and bias parameters of the i-th convolutional kernel in the l-th layer. (l) (j) represents the j-th local region in the l-th layer. This represents the input of the j-th neuron in the i-th frame of layer l+1.

[0096]

[0097] One-dimensional pooling operation:

[0098] like Figure 5 , Figure 6As shown, the pooling layer is a computational layer that performs downsampling operations, reduces the dimensionality of feature data, and decreases network operating parameters. In one-dimensional convolutional neural networks, the most commonly used pooling methods include max pooling and mean pooling. The operation diagrams of max pooling and mean pooling are shown below. Figure 5 As shown, the pooling regions (i.e., the first pooling layer 21 or the second pooling layer 22) slide sequentially across the feature data in one direction with a specified step size. Max pooling outputs the maximum value of the elements in the overlapping portion of the max pooling region and the feature data, while mean pooling outputs the average value of all elements in the overlapping portion of the pooling region and the feature data. The calculation formulas for both are:

[0099]

[0100]

[0101] In the formula Let t be the value of the t-th neuron in the i-th frame of the l-th layer, where t ∈ [(j-1)w+1, jw] and w is the width of the pooling layer's receptive region. This represents the value corresponding to the neuron in the (l+1)th layer of the pooling operation.

[0102] The average pooling method is used in this invention.

[0103] Residual network:

[0104] The residual module can learn deeper features of the input data, avoiding overfitting and improving the network's recognition performance. The residual module mainly consists of ten interconnected residual blocks, which are further divided into downsampling modules (for shortcut connections where data dimensionality needs to be increased) and identity mapping modules (for connections where data dimensionality does not need to be increased).

[0105] The residual block is the most basic building block of a residual neural network. The residual network of this invention is composed of basic residual block units, and the basic structural diagram of the residual block unit is shown below. Figure 7 As shown in the diagram, in the two-layer structure, x is the input data. After input, the data passes through both the direct connection and the shortcut connection. The data in the direct connection undergoes convolution calculation in the convolutional layer and is activated by the Rectified Linear Array (ReLU) function, with the final output being F(x). The x in the shortcut connection is directly sent to the tail of the residual block and added to the F(x) in the direct connection to obtain the overall output H(x) of the residual block.

[0106] like Figure 7 As shown in the figure, x is the input, H(x) is the output, and F(x) is the residual mapping function. By adding the identity mapping of "skip connection" on the convolutional layer, the actual mapping is expressed as H(x) = F(x) + x.

[0107] Where H(x) represents the overall output value, F(x) represents the residual mapping function, and x represents the input value;

[0108] When F(x) = 0, the optimal mapping solution H(x) = x is obtained;

[0109] The mathematical expression for the linear rectifier function is:

[0110]

[0111] In the ten sequentially connected two-layer residual blocks of the residual module, the input, output, and parameter of the l-th two-layer residual block are x, respectively. l f(y) l ) and W l Then we get the formula:

[0112] y l =h(x l )+F(x l W l )

[0113] x l+1 =f(y l )

[0114] Where F(·) is the residual function, and f(·) represents the linear rectified function; y l This represents the output data of the l-th two-layer residual block before the linear rectified function is input; f(y) l h(x) represents the linear rectified function of the output of the l-th two-layer residual block, and is also the final output of the l-th two-layer residual block; l ) represents the mapping function of the input of the l-th two-layer residual block;

[0115] When h(x) l )=x l ,f(y l )=y l When, formula x l+1 =f(y l This can be transformed into:

[0116] x l+1 =x l +F(x l W l )

[0117] Then, the output of the Lth residual block can be derived using a recursive iteration method:

[0118]

[0119] Where xi represents the input of the i-th two-layer residual block, x lThis represents the input to the l-th two-layer residual block;

[0120] x L+1 It represents the input of the (L+1)th two-layer residual block, and is also equal to the output value of the Lth two-layer residual block;

[0121] The input and output data of the above residual block structure have the same dimension, however Figure 3 In the sequence of connecting from top to bottom and left to right, the shortcut connections of the first, third, sixth, and ninth two-layer residual blocks 31 require a 1*1 convolutional kernel to perform dimensionality upscaling on the data. Therefore, when the dimensions of the input data and the output data of the residual blocks are different, it is necessary to use, as shown in the example... Figure 8 The residual structure shown modifies the data dimensionality. This structure simply adds a 1x1 convolutional kernel to increase the data dimensionality and adjusts the stride to downsampling the residual blocks. Ultimately, the input data x and the residual block output data F(x) have the same dimension, allowing direct addition of feature data to obtain the final output H(x) of the residual block. For example: Figure 8 As shown, when the input data is 56*56*64, the direct connection outputs 28*28*128 feature data after convolution calculation. Therefore, the shortcut connection needs to use 128 1*1*64 convolution kernels to increase the 64-dimensional data to 128 dimensions, so as to achieve the same dimension of the output data of the two lines.

[0122] The network model parameters of the improved one-dimensional convolutional neural network used in this invention are shown in Table 1. The first convolutional layer uses eight 23*1 convolutional kernels for convolution calculation, and the second convolutional layer uses sixteen 12*1 convolutional kernels for convolution calculation. Both the first and second pooling layers use pooling operations with a pooling region of 2*1. Ten residual blocks are extracted using 64 24*1, 128 12*1, 256 6*1, and 512 3*1 convolutional kernels, respectively, for deep data feature extraction. Subsequently, the data enters a global average pooling layer for average pooling, and the final result is output by the output layer.

[0123]

[0124]

[0125] Table 1. Parameters of the Improved One-Dimensional Convolutional Neural Network Structure

[0126] After the data preprocessing steps are completed, the two sets of signal datasets are labeled, with the axle fatigue crack signal set to "0" and the interference signal set to "1". All the labeled data are then randomly shuffled, and 70% of the shuffled data is randomly selected as the training dataset, while 30% is selected as the test dataset. The training dataset is then imported into the constructed improved one-dimensional CNN network for training the network model.

[0127] Network parameter selection:

[0128] During training, the network optimizer type, batch size, and learning rate of the network model are compared and selected by following the principle of the controlled variable method. Then, the optimal values ​​of each parameter are input into the network model to achieve the best recognition rate and obtain the best network performance.

[0129] Network optimizer selection:

[0130] While ensuring other network parameters remain consistent, this invention selected five optimizers—ASDG, Adagrad, RMSprop, Adadelat, and Adam—for comparative experiments to identify optimizer types that contribute to improved network performance. (See Table 2.) Figure 9 As shown, the accuracy of ASDG, Adadelat, and RSMprop optimizers in recognizing training dataset data remained around 90% with increasing network iterations, showing no significant improvement. However, the accuracy of Adam and Adagrad optimizers improved considerably with increasing network iterations, with Adam reaching 97.43%. Therefore, the optimizer selected for the network model in this invention is Adam.

[0131]

[0132] Table 2 Recognition rates of different optimizers

[0133] Selection of batch size:

[0134] While ensuring that other network model parameters remain consistent, this invention selects the optimal batch size from four batch sizes: 16, 32, 64, and 128. The relationship between the final recognition rate and the number of iterations for different batch sizes of the network model is as follows: Figure 10 As shown in Table 3. Figure 10As can be seen, when the batch size is 64 and 128, the recognition rate of the network model remains between 91% and 94% with the increase of the number of iterations. When the batch size is 16 and 32, the recognition rate of the network model can reach about 96.53% and 97.43% with the increase of the number of iterations. However, since the smaller the batch size, the higher the running time cost of the network model, the batch size selected for this model is 32.

[0135]

[0136]

[0137] Table 3 Recognition rates for different batch sizes

[0138] Selection of learning rate:

[0139] While keeping other network model parameters consistent, this invention compared network models with learning rates of 0.01, 0.001, and 0.0001, as shown in Table 4. Figure 11 As can be seen, when the learning rate is 0.01, the network model performs poorly, with a recognition rate of only about 45%, indicating that the network has not found the global optimum. When the learning rate is 0.001 and 0.0001, the recognition rate of the network model can reach over 90% with the increase of the number of iterations. When the number of iterations is 30, the recognition rate of the network model with a learning rate of 0.001 is about 91.07%, while the recognition rate of the network model with a learning rate of 0.0001 reaches 97.43%. Therefore, the learning rate of 0.0001 was selected for this model.

[0140]

[0141] Table 4 Recognition rates at different learning rates

[0142] The optimal network performance was obtained by selecting Adam, 32, and 0.0001 for the network optimizer, batch size, and learning rate of the improved one-dimensional CNN network model, and the trained network model was saved. Then, the test dataset was input into the trained network model to identify and analyze fatigue crack signals in the axle.

[0143] Setting up the real-time monitoring module:

[0144] The structure diagram of the real-time monitoring module is as follows: Figure 12 As shown, the real-time monitoring module works by collecting acoustic emission signals through an external acoustic emission detection device during the axle's operation. After preprocessing, the collected signals are then used by a network recognition module for real-time data recognition and the recognition results are output.

[0145] During axle operation, acoustic emission signals are collected using an external acoustic emission detection device. Since data acquisition is real-time, preprocessing is required before importing the collected data into the network model. This preprocessing involves dividing the data into segments of 1024 data points to meet the data input specifications of the network model. Furthermore, because the amount of real-time collected data is enormous, this invention utilizes a QTimer to adjust the data acquisition time and the inference and recognition time of the network model to ensure the periodic loading and operation of the monitoring module, thereby achieving real-time monitoring functionality.

[0146] When the start acquisition button is clicked, the system will drive the acquisition device to acquire acoustic emission signals in real time. The acquired signal data will be divided into data segments containing 1024 data points. With the participation of the timer, only the last data segment is used as the representative of this timing period for network identification in each timing cycle, and the image of the signal segment under test is output under the adjustment of the timer.

[0147] Import the preprocessed signal data into the network recognition module.

[0148] Building the network recognition module:

[0149] Since the signal data collected by the identification system is stored in the system's database, the data required for the network module to run needs to be imported from the system database.

[0150] The network identification module is the most important data processing, calculation, and identification module in the axle fatigue crack acoustic emission signal identification system. Its main function is to use a pre-trained network model to identify and analyze the collected acoustic emission signal data. The design and operation flowchart of the network identification module is as follows: Figure 13 As shown.

[0151] Import the pre-processed signal data from the real-time monitoring module into the network identification module.

[0152] The selection interface for the network identification module is as follows: Figure 14 As shown, the selection area of ​​the network recognition module includes four network model options: a classification model based on an improved one-dimensional CNN, a classification model based on PSO-DBN, a classification model based on wavelet transform + DBN, and a classification model based on VMD decomposition + CNN. When using the network recognition module for signal data recognition, the required network model needs to be selected from these four models and the selection confirmed. This invention requires the selection of the classification model based on an improved one-dimensional CNN.

[0153] Once a network model is selected, the imported acoustic emission signal data will be recognized by the selected network model, and the final recognition result will be output. Furthermore, since the network model was trained with labels "0" and "1" for the two signals respectively, the recognition result display interface of this module will label the axle fatigue crack signal feature and the interference signal feature as "0" and "1" respectively for output display. When the selected model recognizes the input data as one of these two features, the corresponding number will be output in the result graph.

[0154] like Figure 15 As shown, the present invention is a method for acoustic emission signal identification, the method comprising the following steps:

[0155] Step S1: Collect acoustic emission signal data generated during the operation of the axle;

[0156] Step S2: Perform preprocessing operations on the acquired signal data;

[0157] Step S3: Import the preprocessed signal data into the network recognition module, and use the improved one-dimensional CNN network model that has been trained in the network module to recognize and analyze the imported signal data.

[0158] Step S4: Output the recognition result. If the imported signal data is recognized as "axle fatigue crack signal feature", the output result is "0"; if the imported signal data is recognized as "interference signal feature", the output result is "1".

[0159] In step S2 of the above method, the preprocessing operation of the acquired signal data includes the step of dividing the signal data into segments with 1024 data points as a segmentation distance.

[0160] In step S3, the improved one-dimensional CNN network model includes a one-dimensional convolutional pooling module, a residual module, a global average pooling layer, and an output layer. The one-dimensional convolutional pooling module includes two convolutional layers and two pooling layers. One convolutional layer consists of eight 23*1-sized single-layer convolutional kernels, and the other convolutional layer consists of sixteen 12*1-sized two-layer convolutional kernels. Both pooling layers are 2*1 in size. The residual module consists of ten sequentially connected two-layer residual blocks.

[0161] like Figure 3 , Figure 16 As shown, in step S3, the improved one-dimensional CNN network model trained in the network module is used to identify and analyze the imported signal data, specifically including the following steps:

[0162] Step S31: In the convolution kernel 11 of the one-dimensional convolution pooling module 1, a one-dimensional convolution operation is performed on the imported signal data to obtain a layer of one-dimensional convolution feature data.

[0163] Step S32: In the first pooling layer 21 of the one-dimensional convolutional pooling module 1, mean pooling is performed on the one-dimensional convolutional feature data of the layer, and the pooled data is output.

[0164] Step S33: In the second-layer convolution kernel 12 of the one-dimensional convolution pooling module 1, the imported data after the first-layer pooling is subjected to another one-dimensional convolution operation to obtain the two-layer one-dimensional convolution feature data.

[0165] Step S34: In the second pooling layer 22 of the one-dimensional convolutional pooling module 1, the two-layer one-dimensional convolutional feature data is subjected to another mean pooling operation to output the data after the two-layer pooling.

[0166] Step S35: Input the pooled data into the residual module 3. The ten sequentially connected two-layer residual blocks 31 process the two-layer pooled data in sequence and output the overall output value.

[0167] Step S36: The overall output value enters the global average pooling layer for average pooling;

[0168] Step S37: The data after average pooling is output as the final result by the output layer;

[0169] like Figure 17 As shown, in the above-described acoustic emission signal recognition method, the one-dimensional convolution operation in steps S31 and S33 includes the following steps:

[0170] Step S311: Input the signal data into the convolutional kernel of the first layer;

[0171] Step S312, as follows Figure 4 The 3*3 one-dimensional convolution kernel 2 shown performs sliding calculations in the input one-dimensional signal data along one direction with the same length. After each sliding, the values ​​of the overlapping part of the one-dimensional convolution kernel and the one-dimensional signal data are multiplied and then added to obtain the feature data.

[0172] The feature data is composed of multiple neurons, wherein the calculation formula for the j-th neuron in the i-th frame of the (l+1)-th layer is:

[0173]

[0174] in, X represents the weights and bias parameters of the i-th convolutional kernel in the l-th layer. (l) (j) represents the j-th local region in the l-th layer. This represents the j-th neuron in the i-th frame of layer l+1.

[0175] The mean pooling operation described in steps S32 and S34 is as follows:

[0176] The output feature data is subjected to mean pooling. A 2*1 pooling layer slides sequentially across the feature data in one direction with a specified step size. The output is the average of all elements in the overlapping portion of the pooling layer and the feature data, i.e., the pooled data. The calculation formula is as follows:

[0177]

[0178] In the formula Let t be the value of the t-th neuron in the i-th frame of the l-th layer, where t ∈ [(j-1)w+1, jw] and w is the width of the pooling layer's receptive region. This represents the value corresponding to the neuron in the (l+1)th layer of the pooling operation.

[0179] The process of step S35 is as follows:

[0180] When the input and output data of the two residual blocks have the same dimension, the pooled data, after being input into the two residual blocks, will simultaneously pass through both the direct connection and the shortcut connection. The data in the direct connection will undergo convolution calculations by the convolutional layer and be activated by a linear rectification function. The shortcut connection will directly send the input data to the end of the residual block, adding it to the output of the direct connection to obtain the overall output value of the residual block. Its mapping function is expressed as:

[0181] H(x)=F(x)+x

[0182] Where H(x) represents the overall output value, F(x) represents the residual mapping function, and x represents the input value;

[0183] When F(x) = 0, the optimal mapping solution H(x) = x is obtained;

[0184] The mathematical expression for the linear rectifier function is:

[0185]

[0186] In the ten sequentially connected two-layer residual blocks of the residual module, the input, output, and parameter of the l-th two-layer residual block are x, respectively. l f(y) l ) and W l Then we get the formula:

[0187] y l =h(x l )+F(x l W l )

[0188] xl+1 =f(y l )

[0189] Where F(·) is the residual function, and f(·) represents the linear rectified function; y l This represents the output data of the l-th two-layer residual block before the linear rectified function is input; f(y) l h(x) represents the linear rectified function of the output of the l-th two-layer residual block, and is also the final output of the l-th two-layer residual block; l ) represents the mapping function of the input of the l-th two-layer residual block;

[0190] When h(x) l )=x l ,f(y l )=y l When, formula x l+1 =f(y l This can be transformed into:

[0191] x l+1 =x l +F(x l W l )

[0192] Then, the output of the Lth residual block can be derived using a recursive iteration method:

[0193]

[0194] Where xi represents the input of the i-th two-layer residual block, x l This represents the input to the l-th two-layer residual block;

[0195] x L+1 It represents the input of the (L+1)th two-layer residual block, and is also equal to the output value of the Lth two-layer residual block;

[0196] When the dimensions of the input and output data of the residual block are different, the pooled data is input into the two layers of residual blocks and passes through both direct connection and shortcut connection simultaneously. The data in the direct connection undergoes convolution calculation by the convolutional layer and is activated by the linear rectification function. The shortcut connection uses a 1*1 convolutional kernel to perform a dimensionality increase operation on the data, so that the data after the dimensionality increase operation has the same dimension as the output data of the residual block. The data after the dimensionality increase operation is added to the output result of the direct connection to obtain the overall output value of the residual block.

[0197] In the aforementioned acoustic emission signal recognition method, ten sequentially connected two-layer residual blocks are used. The input and output data dimensions of the first, third, sixth, and ninth residual blocks are different, and their shortcut connections use the 1*1 convolution kernel to perform dimensionality upscaling on the data. The input and output data dimensions of the second, fourth, fifth, seventh, eighth, and tenth residual blocks are the same, and their shortcut connections directly transmit the input data to the tail of the residual block.

[0198] like Figure 18 As shown, the present invention also includes a network model training method, comprising the following steps:

[0199] Step S101: Label the two preprocessed signal datasets, with the axle fatigue crack signal set to "0" and the interference signal set to "1".

[0200] Step S102: Randomly shuffle all the set data, and randomly select 70% of the overall shuffled data as the training dataset and 30% of the overall shuffled data as the test dataset.

[0201] Step S103: Import the training dataset into the constructed improved one-dimensional CNN network model for network model training;

[0202] like Figure 19 As shown, the training in step S103 specifically includes the following steps:

[0203] Step S1031: Selection of network optimizer. While ensuring that other network parameters are consistent, five optimizers, ASDG, Adagrad, RMSprop, Adadelat, and Adam, are selected for comparative experiments. The network optimizer is selected based on the accuracy of the network model in recognizing the training dataset as the number of network iterations increases. The network optimizer with the highest recognition accuracy is selected as the network optimizer of the improved one-dimensional CNN network model. Finally, Adam is selected as the network optimizer of the improved one-dimensional CNN network model.

[0204] Step S1032: Selection of batch size. Under the premise of ensuring that the parameters of other network models are consistent, comparative experiments are conducted from four batch sizes: 16, 32, 64, and 128. Based on the recognition rate and running time cost of the network model as the number of iterations increases, the batch size of the improved one-dimensional CNN network model is selected. Finally, the batch size of the improved one-dimensional CNN network model is determined to be 32.

[0205] Step S1033: Selection of learning rate. While ensuring that the parameters of other network models are consistent, network models with learning rates of 0.01, 0.001, and 0.0001 are compared. Based on the recognition rate of the network model as the number of iterations increases, the learning rate of the network model with the highest recognition rate is selected as the learning rate of the improved one-dimensional CNN network model. Finally, the learning rate of the improved one-dimensional CNN network model is determined to be 0.0001.

[0206] Step S1034: Select Adam, 32, and 0.0001 for the network optimizer, batch size, and learning rate of the improved one-dimensional CNN network model to obtain the best network performance, and save the trained network model.

[0207] Step S1035: Input the data from the test dataset into the trained improved one-dimensional CNN network model to identify and analyze the fatigue crack signal of the axle and verify the effectiveness of the model.

[0208] The specific steps to verify the effectiveness of the model are as follows:

[0209] 1. The relationship between the recognition rate, loss function, and number of iterations of the trained improved one-dimensional CNN network model was analyzed. It was ultimately determined that the network model reached its optimal performance after 70 iterations, with the recognition rate decreasing and the error value stabilizing, reaching 98.66%.

[0210] 2. After analyzing the confusion matrix output after recognizing the test dataset, it was finally determined that the network model correctly recognized 2057 data groups out of a total of 2085 data groups. Therefore, it can be concluded that the network model achieved a recognition rate of 98.66% for the test dataset.

[0211] 3. Visual analysis was performed on the data before entering the network model, the output data of the network model before training, and the output data of the network model after training. It was ultimately determined that only after the data passed through the trained network model did various data types essentially cluster in their respective regions, and the network model's recognition rate reached 98.66%. Through the above analysis of the performance of the improved one-dimensional CNN network model after training, it can be concluded that the trained network model has a good recognition effect on the acoustic emission signal of axle fatigue cracks, the model is effective, and training is complete.

[0212] In the above network model training method, step S101, the preprocessing operation includes dividing the two sets of signal data into segments with a segmentation distance of 1024 data points.

[0213] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying acoustic emission signals, characterized in that, The method includes the following steps: Step S1: Collect acoustic emission signal data generated during the operation of the axle; Step S2: Perform preprocessing operations on the acquired signal data; Step S3: Import the preprocessed signal data into the network recognition module, and use the improved one-dimensional CNN network model that has been trained in the network module to recognize and analyze the imported signal data. Step S4: Output the recognition result. If the imported signal data is recognized as "axle fatigue crack signal feature", the output result is "0"; if the imported signal data is recognized as "interference signal feature", the output result is "1". In step S2, the collected signal data is preprocessed, including the step of dividing the signal data into segments with 1024 data points as a segmentation distance. In step S3, the improved one-dimensional CNN network model includes a one-dimensional convolutional pooling module, a residual module, a global average pooling layer, and an output layer. The one-dimensional convolutional pooling module includes two convolutional layers and two pooling layers. One convolutional layer consists of eight 23*1 single-layer convolutional kernels, and the other convolutional layer consists of sixteen 12*1 double-layer convolutional kernels. Both pooling layers are 2*1 in size. The residual module consists of ten sequentially connected two-layer residual blocks. In step S3, the imported signal data is identified and analyzed using the improved one-dimensional CNN network model that has been trained in the network module. This specifically includes the following steps: Step S31: In the convolution kernel (11) of the one-dimensional convolution pooling module (1), perform a one-dimensional convolution operation on the imported signal data to obtain a layer of one-dimensional convolution feature data. Step S32: In the first pooling layer (21) of the one-dimensional convolutional pooling module (1), mean pooling is performed on the one-dimensional convolutional feature data of the layer, and the pooled data is output. Step S33: In the second-layer convolution kernel (12) of the one-dimensional convolution pooling module (1), the imported data after the first-layer pooling is subjected to another one-dimensional convolution operation to obtain the two-layer one-dimensional convolution feature data. Step S34: In the second pooling layer (22) of the one-dimensional convolutional pooling module (1), the two-layer one-dimensional convolutional feature data is subjected to mean pooling again, and the data after the two-layer pooling is output. Step S35: Input the pooled data into the residual module (3). The ten sequentially connected two-layer residual blocks (31) process the pooled data in sequence and output the overall output value. Step S36: The overall output value enters the global average pooling layer for average pooling; Step S37: The data after average pooling is output as the final result by the output layer.

2. The acoustic emission signal identification method according to claim 1, characterized in that, The one-dimensional convolution operation in step S31 and step S33 includes the following steps: Step S311: Input the signal data into the convolutional kernel of the first layer; Step S312: The 3*3 size one-dimensional convolution kernel (2) slides in the input one-dimensional signal data in one direction with the same length. After each slide, the value of the one-dimensional convolution kernel and the one-dimensional signal data overlaps and is then added to obtain the feature data. The feature data is composed of multiple neurons, wherein the first... The first in the layer The first frame The formula for calculating one neuron is: ； in, , Representing the first The first in the layer The weights and bias parameters of each convolutional kernel. Then it means the first The first in the layer A local area, Indicates the first The first in the layer The first frame One neuron.

3. The acoustic emission signal identification method according to claim 2, characterized in that, The mean pooling operation described in steps S32 and S34 is as follows: The output feature data is subjected to mean pooling. A 2*1 pooling layer slides sequentially across the feature data in one direction with a specified step size. The output is the average of all elements in the overlapping portion of the pooling layer and the feature data, i.e., the pooled data. The calculation formula is as follows: ； In the formula For the first Layer The first frame The value of each neuron, , For the width of the pooling layer's sensing region, For the pooling operation The corresponding values ​​of neurons in the layer.

4. The acoustic emission signal identification method according to claim 1, characterized in that, The process of step S35 is as follows: When the input and output data of the two residual blocks have the same dimension, the pooled data, after being input into the two residual blocks, will simultaneously pass through both the direct connection and the shortcut connection. The data in the direct connection will undergo convolution calculations by the convolutional layer and be activated by a linear rectification function. The shortcut connection will directly send the input data to the end of the residual block, adding it to the output of the direct connection to obtain the overall output value of the residual block. Its mapping function is expressed as: in, This represents the overall output value. Represents the residual mapping function. Indicates the input value; when When = 0, the optimal mapping solution is obtained. = ; The mathematical expression for the linear rectifier function is: ； In the ten sequentially connected two-layer residual blocks of the residual module, the first... The inputs, outputs, and parameters of the two residual blocks are respectively , Then we get the formula: ； ； in, For the residual function, Represents a linear rectified function; This represents the output data of the l-th two-layer residual block before the linear rectifier function is input; This represents the linear rectified function of the output of the l-th two-layer residual block, and is also the final output of the l-th two-layer residual block; The mapping function representing the input of the l-th two-layer residual block; when = , = When, formula It can be converted into: ； Then, the first step can be derived through recursive iteration. The output of the group residual block is: ； Where xi represents the input of the i-th two-layer residual block, This represents the input to the l-th two-layer residual block; This represents the input of the (L+1)th two-layer residual block, and is also equal to the output value of the Lth two-layer residual block; When the dimensions of the input and output data of the residual block are different, the pooled data is input into the two layers of residual blocks and passes through both direct connection and shortcut connection simultaneously. The data in the direct connection undergoes convolution calculation by the convolutional layer and is activated by the linear rectification function. The shortcut connection uses a 1*1 convolutional kernel to perform a dimensionality increase operation on the data, so that the data after the dimensionality increase operation has the same dimension as the output data of the residual block. The data after the dimensionality increase operation is added to the output result of the direct connection to obtain the overall output value of the residual block.

5. The acoustic emission signal identification method according to claim 4, characterized in that, The ten sequentially connected two-layer residual blocks have different dimensions for input and output data in the first, third, sixth, and ninth residual blocks, and their shortcut connections use the 1*1 convolution kernel to perform dimensionality upscaling on the data; while the second, fourth, fifth, seventh, eighth, and tenth residual blocks have the same dimension for input and output data, and their shortcut connections directly send the input data to the end of the residual block.

6. A method for training a network model, characterized in that, Includes the following steps: Step S101: Label the two preprocessed signal datasets, with the axle fatigue crack signal set to "0" and the interference signal set to "1". Step S102: Randomly shuffle all the set data, and randomly select 70% of the overall shuffled data as the training dataset and 30% of the overall shuffled data as the test dataset. Step S103: Import the training dataset into the constructed improved one-dimensional CNN network model for network model training; The training in step S103 specifically includes the following steps: Step S1031: Selection of network optimizer. While ensuring that other network parameters are consistent, five optimizers, ASDG, Adagrad, RMSprop, Adadelat, and Adam, are selected for comparative experiments. The network optimizer is selected based on the accuracy of the network model in recognizing the training dataset as the number of network iterations increases. The network optimizer with the highest recognition accuracy is selected as the network optimizer of the improved one-dimensional CNN network model. Finally, Adam is selected as the network optimizer of the improved one-dimensional CNN network model. Step S1032: Selection of batch size. Under the premise of ensuring that the parameters of other network models are consistent, comparative experiments are conducted from four batch sizes: 16, 32, 64, and 128. Based on the recognition rate and running time cost of the network model as the number of iterations increases, the batch size of the improved one-dimensional CNN network model is selected. Finally, the batch size of the improved one-dimensional CNN network model is determined to be 32. Step S1033: Selection of learning rate. While ensuring that the parameters of other network models are consistent, network models with learning rates of 0.01, 0.001, and 0.0001 are compared. Based on the recognition rate of the network model as the number of iterations increases, the learning rate of the network model with the highest recognition rate is selected as the learning rate of the improved one-dimensional CNN network model. Finally, the learning rate of the improved one-dimensional CNN network model is determined to be 0.0001. Step S1034: Select Adam, 32, and 0.0001 for the network optimizer, batch size, and learning rate of the improved one-dimensional CNN network model to obtain the best network performance, and save the trained network model. Step S1035: Input the data from the test dataset into the trained improved one-dimensional CNN network model to identify and analyze the fatigue crack signal of the axle and verify the effectiveness of the model.

7. The network model training method according to claim 6, characterized in that, In step S101, the preprocessing operation includes dividing the two sets of signal data into segments with a segmentation distance of 1024 data points.

Images