A locomotive wheel polygonal damage detection method based on lightweight neural network

Abstract

The present invention discloses a method for detecting polygonal damage in locomotive wheels based on a lightweight neural network. The method comprises the following steps: obtaining vertical vibration acceleration response signals of the axlebox of a heavy-loaded locomotive at different operating speeds and different combinations of wheel polygon wear levels; preprocessing the signals, obtaining the frequency domain characteristics of the signals through a fast Fourier transform; constructing a sample dataset based on the frequency domain characteristics; constructing a wheel polygonal damage detection network model; training the model using the sample dataset; and using the trained polygonal damage detection network model to identify wheel polygon wear amplitudes, thereby completing wheel polygon damage detection. This method can achieve precise quantitative detection of wheel polygon wear levels on heavy-loaded locomotives, and is characterized by accuracy, efficiency, robustness, and good interpretability.

Description

Technical Field

[0001] The present invention relates to the field of locomotive wheel detection, and in particular to a locomotive wheel polygonal damage detection method based on a lightweight neural network. Background Art

[0002] With the rapid growth of my country's railway freight volume, the scale of heavy-haul train formations has continued to expand, and heavy-haul trains are gradually moving towards higher loads and higher speeds. The resulting interaction between wheels and rails has become increasingly intense. Common locomotive wheel damage detection methods include static and dynamic testing. While static testing offers high measurement accuracy, it is labor-intensive and consumes train turnaround time, making its detection efficiency insufficient to meet the current operational needs of heavy-haul trains. Dynamic testing, on the other hand, allows for quantitative wheel identification during train operation, resulting in higher detection efficiency and is the future development direction for polygonal wheel detection in trains. Among these, the on-board dynamic detection method based on axlebox vibration acceleration is gaining popularity among railway personnel due to its simple structure, easy installation, high detection efficiency, and real-time performance. It is a simple and effective method that enables online and continuous monitoring of locomotive wheels.

[0003] However, the existing dynamic detection methods have the following technical problems: most of the existing technologies focus on qualitative identification of wheel polygons, which makes it difficult to accurately quantify the degree of wheel polygon wear and unable to guide the intelligent operation and maintenance of wheel polygons; the dynamic detection method of wheel polygons based on the trackside method is easily affected by factors such as bad weather and complex environment, and the detection accuracy is strongly negatively correlated with the locomotive operating speed. It is difficult to identify high-order polygons and cannot achieve accurate real-time status monitoring of wheel polygons.

[0004] In order to overcome the above problems, the present invention provides a locomotive wheel polygon damage detection method based on a lightweight neural network. The method realizes the intelligent quantitative detection of the polygon wear amplitude of heavy-load locomotive wheels through the vertical vibration acceleration signal of the locomotive axle box. The method has high detection accuracy and is not affected by the external environment. Summary of the Invention

[0005] The purpose of the present invention is to provide a locomotive wheel polygonal damage detection method based on a lightweight neural network.

[0006] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0007] The present invention provides a locomotive wheel polygonal damage detection method based on a lightweight neural network, comprising:

[0008] Obtain the vertical vibration acceleration response signal of the axle box of a heavy-load locomotive under different operating speeds and different wheel polygon wear degree combinations;

[0009] After preprocessing the signal, the frequency domain characteristics of the signal are obtained through fast Fourier transform, and a sample data set is constructed based on the frequency domain characteristics;

[0010] Constructing a wheel polygonal damage detection network model and training the model using a sample data set; the wheel polygonal damage detection network model includes a shallow spatial feature extraction module, a deep sequence feature extraction module, an advanced feature enhancement module, and a wear amplitude quantitative detection module;

[0011] The trained polygonal damage detection network model is used to identify the wheel polygon wear amplitude and complete the quantitative detection of wheel polygon damage.

[0012] Preferably, the method of preprocessing the signal includes abnormal data elimination and data expansion, specifically:

[0013] The collected signal is divided into 10-second / window lengths;

[0014] Based on the segmented signal, short-duration, high-amplitude abnormal pulses caused by sensor failure, electromagnetic interference or transient impact, and low-frequency baseline drift trend items caused by sensor drift or continuous speed changes are identified and eliminated;

[0015] The window size is set to 2048, the sliding step is set to 500, and the window slicing sampling method is used to expand the signal data.

[0016] Preferably, the method of obtaining the frequency domain characteristics of the signal by fast Fourier transform and constructing the sample data set based on the frequency domain characteristics is specifically as follows:

[0017] A 150 Hz low-pass filter was applied to the fast Fourier transformed signal to retain the excitation characteristics related to the polygon wear amplitude;

[0018] The frequency domain features under different speeds and different polygonal wheels are obtained, and a sample data set is constructed. The data set is divided into training set, validation set and test set in a ratio of 7:2:1, and normalized.

[0019] Preferably, each sample in the sample data set contains 1024 frequency domain features, and each sample corresponds to 3 label values ​​of polygon wear amplitudes.

[0020] Preferably, the shallow spatial feature extraction module is composed of a deep residual network and a deep separable convolutional network, wherein the deep separable convolutional network includes a deep convolutional network and a point-by-point convolutional network, specifically:

[0021] The expression of the deep residual network is:

[0022] (1);

[0023] Where, is the final output of the residual block, is the input feature map, represents the residual connection, is the output after the convolution layer, batch normalization layer and activation layer;

[0024] The expression of deep convolutional network is:

[0025] (2);

[0026] Where, The first layer in the deep convolutional network Channels at position The output, For input data, is the convolution kernel length, For the Channels at position The value of For the The first channel weights;

[0027] The expression of point-by-point convolutional network is:

[0028] (3);

[0029] Where, Point-by-point convolutional network Channels at position The output, is the number of input channels, For the Channels at position The value of For the The output channel pair The weight of the input channels is The convolution kernel.

[0030] Preferably, the deep residual network and the deep separable convolutional network are both three-layer network structures.

[0031] Preferably, the deep sequence feature extraction module is composed of a two-layer bidirectional long short-term memory network, which is used to model the front-end sequence of features and learn the long-range dependencies and dynamic change patterns of input signals.

[0032] Preferably, the advanced feature enhancement module is composed of a spatial attention mechanism and a channel attention mechanism, which are used to capture the long-range dependencies between features at different spatial positions, specifically:

[0033] The output of the spatial attention mechanism is:

[0034] (4);

[0035] Where, is the output of the spatial attention mechanism, are the learnable parameters of the spatial attention mechanism, is the total number of frequency points, and Frequency points and The eigenvector of Frequency point The original characteristics of is the input feature;

[0036] The output of the channel attention mechanism is:

[0037] (5);

[0038] Where, is the output of the channel attention mechanism, is the learnable parameter of the channel attention mechanism, is the number of input channels, and Channel and channel The eigenvector of

[0039] The final output of the dual attention mechanism is:

[0040] (6);

[0041] Where, For the final output, and are the outputs of the spatial attention mechanism and the channel attention mechanism, respectively.

[0042] Preferably, the wear amplitude quantitative detection module consists of a flattening layer and five fully connected layers.

[0043] Preferably, the method of training the model using the sample data set is specifically as follows:

[0044] The wheel polygon damage detection network model is trained using the training set, and the mean square error minimization is used as the objective function, where the mean square error is:

[0045] (7);

[0046] Where, is the sample size, and Respectively The true value and predicted value of the sample;

[0047] Back propagation is performed based on the error of the objective function to update the weight parameters and hyperparameter combinations of the wheel polygon damage detection network model.

[0048] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0049] (1) The wheel polygon damage detection network in the present invention includes deep learning models such as deep residual network and deep separable convolutional network, which effectively solves the problem of network model degradation. The network model computational efficiency is improved by reducing the number of model parameters, and the local lightweight private deployment of the network model and the effective extraction of key sensitive features are realized, thereby improving the accuracy of feature extraction.

[0050] (2) The present invention introduces a dual attention mechanism in the advanced feature enhancement module to mine sensitive features in both spatial and channel dimensions, retaining high-dimensional features closely related to polygon wear amplitude, and realizing accurate identification of polygon wear amplitude, which has the characteristics of strong generalization ability and robust performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The present invention provides a flowchart of a method for detecting polygonal damage of locomotive wheels based on a lightweight neural network.

[0052] Figure 2 Schematic diagram of the installation position of the axle box acceleration sensor in an embodiment of the present invention.

[0053] Figure 3 This is a diagram showing the wheel polygon test results in an embodiment of the present invention.

[0054] Figure 4 Schematic diagram of signals at a speed of 60 km / h in an embodiment of the present invention.

[0055] Figure 5 This is a frequency domain characteristic diagram at a speed of 60 km / h in an embodiment of the present invention.

[0056] Figure 6 This is a flow chart of the architecture of the wheel polygonal damage detection network model in an embodiment of the present invention.

[0057] Figure 7 This is a polygon wear amplitude identification result diagram at a speed of 40 km / h in an embodiment of the present invention.

[0058] Figure 8 This is a polygon wear amplitude identification result diagram at a speed of 60 km / h in an embodiment of the present invention.

[0059] Figure 9 This is a polygon wear amplitude identification result diagram at a speed of 80 km / h in an embodiment of the present invention.

[0060] Figure 10 This is an evaluation index diagram for quantitative detection of polygonal wear amplitude in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0062] Reference Figure 1 As shown, the present invention provides a locomotive wheel polygonal damage detection method based on a lightweight neural network, comprising:

[0063] The vertical vibration acceleration response signals of the axle box of a heavy-load locomotive are obtained under different combinations of running speeds and different wheel polygon wear degrees.

[0064] In this embodiment, Figure 2 The axle box acceleration sensor is installed at the position shown in the figure. v The vertical vibration response signal of the axle box is obtained in the locomotive speed direction. The orders of the locomotive wheel polygons are 17, 18 and 19, and the corresponding amplitudes are 0.0336mm, 0.0299mm and 0.0151mm respectively. Figure 3 Wheel polygon test results shown.

[0065] In this embodiment, the vertical vibration acceleration signals of the axle box of the heavy-load locomotive at the running speeds of 40 km / h, 60 km / h and 80 km / h are collected at a sampling frequency of 4096 Hz.

[0066] After preprocessing the signal, the frequency domain characteristics of the signal are obtained through fast Fourier transform, and a sample data set is constructed based on the frequency domain characteristics.

[0067] It should be explained that the excitation frequency under the wheel polygon excitation is closely related to the polygon harmonic order, and after the timing signal is subjected to fast Fourier transform, the polygon wear harmonic order can be determined:

[0068] (8);

[0069] Where, is the excitation frequency, is the polygon harmonic order, is the train running speed, is the wheel radius;

[0070] Therefore, using the frequency domain signal as the original input of the wheel polygonal damage detection network model can effectively alleviate the influence of harmonic order on the wear amplitude recognition accuracy.

[0071] Specifically, the method of preprocessing the signal includes abnormal data removal and data expansion, specifically:

[0072] The collected signal is divided into 10-second / window lengths;

[0073] Identify and eliminate short-duration, high-amplitude abnormal pulses caused by sensor failure, electromagnetic interference, or transient shocks, and eliminate low-frequency baseline drift trends caused by sensor drift or continuous speed changes;

[0074] The window size is set to 2048, the sliding step is set to 500, and the window slicing sampling method is used to expand the signal data.

[0075] Among them, eliminating short-term high-amplitude abnormal pulses can eliminate the interference of abnormal shocks on the data; eliminating low-frequency baseline drift trend items can ensure that the signal is better concentrated on the characteristic frequency components, ensuring data reliability.

[0076] Since the signals collected on site often contain a large amount of noise signals introduced by disturbances such as track irregularities and rail corrugation, the signals need to be filtered to avoid the influence of noise signals on the accuracy of quantitative diagnosis of wear amplitude.

[0077] Furthermore, the frequency domain features of the signal are obtained by fast Fourier transform, and a method of constructing a sample data set based on the frequency domain features is as follows:

[0078] A 150 Hz low-pass filter was applied to the fast Fourier transformed signal to retain the excitation characteristics related to the polygon wear amplitude;

[0079] The frequency domain features under different speeds and different polygonal wheels are obtained, and a sample data set is constructed. The data set is divided into training set, validation set and test set in a ratio of 7:2:1, and normalized.

[0080] Each sample in the sample data set contains 1024 frequency domain features, and each sample corresponds to the label value of 3 polygon wear amplitudes.

[0081] For example, after applying low-pass filtering, the vertical vibration acceleration response signal and frequency domain characteristics of the axle box under polygonal excitation at a speed of 60km / h are as follows: Figure 4 and Figure 5shown.

[0082] In this embodiment, the sample data set includes 700 training sets, 200 validation sets, and 100 test sets, and each feature in the data set is normalized:

[0083] (9);

[0084] Where, is the normalized feature data, is the feature in the original dataset, and are the maximum and minimum values ​​in the original features respectively.

[0085] A wheel polygonal damage detection network model is constructed and trained using a sample data set; the wheel polygonal damage detection network model includes a shallow spatial feature extraction module, a deep sequence feature extraction module, an advanced feature enhancement module and a wear amplitude quantitative detection module.

[0086] Specifically, the shallow spatial feature extraction module is composed of a deep residual network and a deep separable convolutional network, wherein the deep separable convolutional network includes a deep convolutional network and a point-by-point convolutional network, specifically:

[0087] The expression of the deep residual network is:

[0088] (1);

[0089] Where, is the final output of the residual block, is the input feature map, represents the residual connection, is the output after several convolutional layers, batch normalization layers, and activation layers;

[0090] The expression of deep convolutional network is:

[0091] (2);

[0092] Where, The first layer in the deep convolutional network Channels at position The output, For input data, is the convolution kernel length, For the Channels at position The value of For the The first channel weights;

[0093] The expression of point-by-point convolutional network is:

[0094] (3);

[0095] Where, Point-by-point convolutional network Channels at position The output, is the number of input channels, For the Channels at position The value of For the The output channel pair The weight of the input channels is The convolution kernel.

[0096] Specifically, the deep sequence feature extraction module consists of a two-layer bidirectional long short-term memory network, which is used to model the front-end and back-end sequences of features and learn the long-range dependencies and dynamic change patterns of input signals.

[0097] Specifically, the advanced feature enhancement module consists of a spatial attention mechanism and a channel attention mechanism to capture the long-range dependencies between features at different spatial locations, specifically:

[0098] The output of the spatial attention mechanism is:

[0099] (4);

[0100] Where, is the output of the spatial attention mechanism, are the learnable parameters of the spatial attention mechanism, is the total number of frequency points, and Frequency points and The eigenvector of Frequency point The original characteristics of is the input feature;

[0101] The output of the channel attention mechanism is:

[0102] (5);

[0103] Where, is the output of the channel attention mechanism, is the learnable parameter of the channel attention mechanism, is the number of input channels, and Channel and channel The eigenvector of

[0104] The final output of the dual attention mechanism is:

[0105] (6);

[0106] Where, For the final output, and are the outputs of the spatial attention mechanism and the channel attention mechanism, respectively.

[0107] Among them, the deep residual network solves the deep training problem through jump connections and can effectively extract key sensitive features. The deep separable convolutional network reduces the number of model parameters through decomposition operations, reduces computing resources, greatly improves computing efficiency, and facilitates lightweight deployment of network models.

[0108] The deep sequence feature extraction module uses a bidirectional long short-term memory network to perform forward and backward sequence modeling on the extracted sensitive features, and learns the long-range dependencies and dynamic change patterns of the input signals. Through this module, the characteristic components related to the polygon wear amplitude in the frequency domain signal can be effectively captured, laying the data feature foundation for improving the accuracy of polygon quantitative recognition.

[0109] The advanced feature enhancement module is implemented through a dual attention mechanism, which consists of a spatial attention mechanism and a channel attention mechanism. It captures the long-distance dependencies between key features in different spatial positions and the correlations between different channels, adaptively combines local features and global dependencies, suppresses noise signals, and improves the feature representation capabilities of important response data.

[0110] In this embodiment, the configuration parameters of each module of the polygonal damage detection network model are shown in Table 1:

[0111] Table 1 Module configuration parameters

[0112]

[0113] like Figure 6The figure shows the architectural flow chart of the wheel polygonal damage detection network model, in which the shallow spatial feature extraction module consists of a deep residual network and a deep separable convolutional network. The dataset with a shape of [32, 1024, 1] is input into the three-layer residual network module for primary feature extraction, and the extracted features are input into the three-layer deep separable convolutional network to further extract deep features, obtaining high-dimensional features with a shape of [32, 128, 128]. The features are input into the two-layer deep sequence feature extraction module, and the bidirectional long short-term memory network is used to realize high-dimensional feature data modeling and mine depth-sensitive features. The output dimension is [32, 128, 8]. The advanced feature enhancement module consists of a dual attention mechanism to realize the enhancement of high-dimensional features and the suppression of noise signals. Finally, the polygonal wear amplitude quantitative regression model composed of a fully connected network is used to realize high-precision quantitative detection of wear amplitude, and the final output dimension is [32, 3].

[0114] It should be understood that the polygonal damage detection network model uses a three-layer depth-separable convolutional network to replace the traditional convolutional network model, which can effectively improve the number of network model parameters. For example, the number of parameters of the common traditional convolutional network model is 3 16 32+3 32 64+3 64 128=32256, while the parameter number of the depthwise separable convolutional network model of the present invention is 3 16+16 32+3 32 64+3 64+64 128=11088, which is about 35% of the parameters of the traditional convolutional neural network model. This reduces the number of model parameters and lays a good foundation for lightweight local deployment.

[0115] The trained polygonal damage detection network model is used to identify the wheel polygon wear amplitude and complete the quantitative detection of wheel polygon damage.

[0116] Specifically, the method of training the model using the sample data set is as follows:

[0117] The wheel polygon damage detection network model is trained using the training set, and the mean square error minimization is used as the objective function, where the mean square error is:

[0118] (7);

[0119] Where, is the sample size, and Respectively The true value and predicted value of the sample;

[0120] Back propagation is performed based on the error of the objective function to update the weight parameters and hyperparameter combinations of the wheel polygon damage detection network model.

[0121] In this embodiment, the polygonal wheel wear amplitudes at 40km / h, 60km / h and 80km / h speed conditions are identified, and the identification results are as follows: Figure 7 、 Figure 8 and Figure 9 As shown, the root mean square error (RMSE) and goodness of fit (R 2 ) indicators such as Figure 10 As shown, they are 3.892E-4, 4.093E-4, 4.21E-4 and 0.99898, 0.99848, 0.99894 respectively.

[0122] Specifically, in this embodiment, the wheel polygon damage detection network model achieves high recognition accuracy and can become an effective means for online quantitative detection of polygon wear amplitude of heavy-load locomotive wheels.

[0123] The above content is merely an example and explanation of the structure of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the scope of protection of the present invention.

Claims

1. A locomotive wheel polygonal damage detection method based on lightweight neural network, characterized in that: The following steps are involved: Obtain the vertical vibration acceleration response signal of the axle box of a heavy-load locomotive under different operating speeds and different wheel polygon wear degree combinations; After preprocessing the signal, the frequency domain characteristics of the signal are obtained through fast Fourier transform, and a sample data set is constructed based on the frequency domain characteristics; Constructing a wheel polygonal damage detection network model and training the model using a sample data set; the wheel polygonal damage detection network model includes a shallow spatial feature extraction module, a deep sequence feature extraction module, an advanced feature enhancement module, and a wear amplitude quantitative detection module; The shallow spatial feature extraction module is composed of a deep residual network and a deep separable convolutional network, wherein the deep separable convolutional network includes a deep convolutional network and a point-by-point convolutional network; The advanced feature enhancement module consists of a spatial attention mechanism and a channel attention mechanism to capture the long-range dependencies between features at different spatial locations. Specifically: The output of the spatial attention mechanism is: （4）； Where, is the output of the spatial attention mechanism, are the learnable parameters of the spatial attention mechanism, is the total number of frequency points, and Frequency points and The eigenvector of Frequency point The original characteristics of is the input feature; The output of the channel attention mechanism is: （5）； Where, is the output of the channel attention mechanism, is the learnable parameter of the channel attention mechanism, is the number of input channels, and Channel and channel The eigenvector of The final output of the dual attention mechanism is: （6）； Where, For the final output, and are the outputs of the spatial attention mechanism and the channel attention mechanism respectively; The trained polygonal damage detection network model is used to identify the wheel polygon wear amplitude and complete the quantitative detection of wheel polygon damage.

2. The method for detecting polygonal damage of locomotive wheels based on a lightweight neural network according to claim 1, characterized in that: The method for preprocessing the signal includes abnormal data elimination and data expansion, specifically: The collected signal is divided into 10-second / window lengths; Based on the segmented signal, short-duration, high-amplitude abnormal pulses caused by sensor failure, electromagnetic interference or transient impact, and low-frequency baseline drift trend items caused by sensor drift or continuous speed changes are identified and eliminated; The window size is set to 2048, the sliding step is set to 500, and the window slicing sampling method is used to expand the signal data.

3. The method for detecting polygonal damage of locomotive wheels based on a lightweight neural network according to claim 1, characterized in that: The method of obtaining the frequency domain characteristics of the signal through fast Fourier transform and constructing a sample data set based on the frequency domain characteristics is specifically as follows: A 150 Hz low-pass filter was applied to the fast Fourier transformed signal to retain the excitation characteristics related to the polygon wear amplitude; The frequency domain features under different speeds and different polygonal wheels are obtained, and a sample data set is constructed. The data set is divided into training set, validation set and test set in a ratio of 7:2:1, and normalized.

4. The method for detecting polygonal damage of locomotive wheels based on a lightweight neural network according to claim 3, characterized in that: Each sample in the sample data set contains 1024 frequency domain features, and each sample corresponds to 3 label values ​​of polygon wear amplitude.

5. The method for detecting polygonal damage of locomotive wheels based on a lightweight neural network according to claim 1, characterized in that: The expression of the deep residual network is: （1）； Where, is the final output of the residual block, is the input feature map, represents the residual connection, is the output after the convolution layer, batch normalization layer and activation layer; The expression of deep convolutional network is: （2）； Where, The first layer in the deep convolutional network Channels at position The output, For input data, is the convolution kernel length, For the Channels at position The value of For the The first channel weights; The expression of point-by-point convolutional network is: （3）； Where, Point-by-point convolutional network Channels at position The output, is the number of input channels, For the Channels at position The value of For the The output channel pair The weight of the input channels is The convolution kernel.

6. The method for detecting polygonal damage of locomotive wheels based on a lightweight neural network according to claim 5, characterized in that: The deep residual network and the deep separable convolutional network both have three-layer network structures.

7. The method for detecting polygonal damage of locomotive wheels based on a lightweight neural network according to claim 1, characterized in that: The deep sequence feature extraction module consists of a two-layer bidirectional long short-term memory network, which is used to model the front-end and back-end sequences of features and learn the long-range dependencies and dynamic change patterns of input signals.

8. The method for detecting polygonal damage of locomotive wheels based on a lightweight neural network according to claim 1, characterized in that: The wear amplitude quantitative detection module consists of a flattening layer and five fully connected layers.

9. The method for detecting polygonal damage of locomotive wheels based on a lightweight neural network according to claim 1, characterized in that: The method of using the sample data set to train the model is specifically as follows: The wheel polygon damage detection network model is trained using the training set, and the mean square error minimization is used as the objective function, where the mean square error is: （7）； Where, is the sample size, and Respectively The true value and predicted value of the sample; Back propagation is performed based on the error of the objective function to update the weight parameters and hyperparameter combinations of the wheel polygon damage detection network model.

Images