Steel rail corrugation recognition method based on acoustic signals

Abstract

The invention discloses a steel rail corrugation recognition method based on acoustic signals, which comprises the following steps: acquiring a large number of acoustic signals in train operation, dividing the acoustic signals into a normal data set and a corrugation data set, and then cleaning the normal data set and the corrugation data set; slicing the acoustic signals in the normal data set and the corrugation data set after cleaning processing to obtain a plurality of normal time domain signals and a plurality of corrugation time domain signals; acquiring three-channel characteristics of the normal time domain signal and the corrugation time domain signal; the three-channel features are used as input, the time domain signal types corresponding to the three-channel features are used as labels, the built CNN-Transform network structure is trained, and a trained CNN-Transform model is obtained; and the real-time acoustic signal of the target train is recognized through the trained model, and a steel rail corrugation state recognition result corresponding to the real-time acoustic signal is obtained. The method can effectively improve the reliability and real-time performance of steel rail corrugation recognition and reduce the recognition cost.

Description

Technical Field

[0001] The invention relates to the technical field of rail corrugation recognition, and more particularly to a rail corrugation recognition method based on acoustic signals. Background Art

[0002] Rail corrugation refers to the wavy wear phenomenon formed on the rail surface along the longitudinal direction. Its existence is a great safety hazard for rail transportation systems. Corrugation will cause additional vibration and noise when the train is running, affecting the passenger experience, and increase the wear of track and train components, thereby increasing maintenance costs. Therefore, accurately identifying and evaluating the degree of rail corrugation is crucial to preventing accidents and ensuring driving safety.

[0003] Traditionally, the methods for detecting rail corrugation mainly include the chord measurement method and the inertial reference method. The chord measurement method uses photoelectric displacement sensors or other sensors to measure the changes in the track surface relative to the baseline to identify corrugation. However, this method is limited by its inherent limitations, such as the range of variation between the measured value and the theoretical value, and the data can only measure a maximum wavelength of 40 meters, which makes it difficult to fully reflect the corrugation status. The inertial reference method uses an acceleration sensor installed on the vehicle axle box to detect corrugation. This method relies on the vehicle body as a static reference point, but because the train body cannot remain static when the speed decreases or encounters long corrugation, this makes the inertial reference method invalid under certain conditions.

[0004] Given the limitations of traditional methods, researchers are exploring new methods to detect rail corrugation. With the development of artificial intelligence technology, neural networks, as a powerful pattern recognition tool, have been widely used in image recognition, natural language processing and other fields, and have demonstrated a strong ability to process complex data. The advancement of neural network technology, especially the maturity of deep learning frameworks, has enabled it to perform well in processing large amounts of complex data.

[0005] The invention patent application with publication number CN117828431A and titled Method and Device for Identifying Rail Corrugation of Railways discloses that the axle box acceleration data and line information of the specified railway are first obtained, and then the data are preprocessed to generate a time-frequency graph, and a sample data set is constructed based on this. The sample data set is used to train a neural network model, and finally a rail corrugation identification model is formed, which is used to detect the corrugation conditions of other railways.

[0006] The invention patent application with publication number CN115758289A, entitled A rail corrugation identification method based on multi-task learning neural network, discloses that by collecting vehicle body vibration acceleration and noise data and calculating the 1 / 3 octave vibration-noise spectrum as model input, wavelength classification and identification is completed using a multi-task learning rail corrugation detection neural network model.

[0007] The publication number is CN111623868A, and its name is "Convolutional Neural Network Construction Method for Rail Corrugation Identification". It discloses that vibration information of the vehicle during operation is collected by a vibration sensor installed on the vehicle, and the vibration time domain information is divided into segmented vibration time domain information corresponding to equal-length paths, and then converted into frequency domain information, which is used as a training sample for a one-dimensional convolutional neural network.

[0008] The publication number is CN115575508A, and the name is a rail transit rail corrugation identification method based on train vibration and sound composite characteristics. It discloses that the train vibration, noise signals and their related signals are collected, and these signals are processed and fused. The one-dimensional convolutional neural network is used to extract features from the fused vibration and sound data, which is suitable for the state identification and real-time monitoring of rail corrugation.

[0009] The above solutions are basically based on vibration data, and the invention patent application with publication number CN112215263A, entitled Rail Corrugation Detection Method Based on Deep Convolutional Neural Network, discloses taking images of the top surface of the rail and selecting and marking images of long corrugation, short corrugation and non-corrugation. A deep convolutional neural network is constructed using the DenseNet network model to train image data, thereby completing the recognition of rail corrugation.

[0010] It can be seen from the above that in the existing corrugation recognition technology, most methods rely on vibration data or image data. Although these methods have improved the accuracy of corrugation recognition to a certain extent, they also have some inherent defects. First, the methods based on vibration data, such as the invention patent applications with publication numbers CN117828431A, CN115758289A, CN111623868A and CN115575508A, although they use neural network models to improve the recognition accuracy, these methods still rely on data collected by acceleration sensors. In practical applications, the speed changes of trains, the complexity of track conditions and the high requirements for preprocessing of acceleration signals will affect the accuracy of the detection results. In addition, the installation and maintenance of sensors also require additional costs and technical support. Secondly, the methods based on image data, such as the invention patent application with publication number CN112215263A, although they provide a novel recognition approach, they rely on good lighting conditions and clear image quality. In a dynamic operating environment, there are many challenges in ensuring high-quality image acquisition, which directly affects the reliability and real-time performance of corrugation recognition.

[0011] Therefore, how to improve the reliability and real-time performance of rail corrugation identification and reduce the identification cost is an urgent problem to be solved by those skilled in the art. Summary of the invention

[0012] In view of the above problems, the present invention provides a rail corrugation identification method based on acoustic signals to at least solve some of the technical problems mentioned in the above background technology.

[0013] In order to achieve the above object, the present invention adopts the following technical solution:

[0014] The present invention provides a rail corrugation identification method based on acoustic signals, comprising the following steps:

[0015] Acquire a large number of acoustic signals during train operation and divide them into normal data sets and corrugation data sets;

[0016] Performing data cleaning processing on the normal data set and the corrugation data set;

[0017] Slicing the acoustic signals in the normal data set and the corrugation data set after cleaning respectively to obtain a plurality of normal time domain signals and a plurality of corrugation time domain signals;

[0018] Acquire three-channel characteristics of the normal time domain signal and the corrugation time domain signal;

[0019] Taking the three-channel features as input, taking the time domain signal type corresponding to the three-channel features as a label, training the constructed CNN-Transformer network structure to obtain a trained CNN-Transformer model;

[0020] The real-time acoustic signal of the target train is recognized through the trained CNN-Transformer model to obtain the rail corrugation state recognition result corresponding to the real-time acoustic signal.

[0021] Furthermore, the data cleaning process of the normal data set and the corrugation data set specifically includes:

[0022] removing corrugation acoustic signals from the normal data set;

[0023] The normal acoustic signal in the corrugation dataset is removed.

[0024] Furthermore, in the slicing process, a plurality of time domain signals are obtained by using a sliding window with a window length of 2 seconds and a window shift of 1 second.

[0025] Furthermore, the step of acquiring the three-channel features of the normal time domain signal and the corrugation time domain signal specifically includes:

[0026] Acquire Mel spectrogram features, MFCC features and CQT features of the normal time domain signal and the corrugation time domain signal respectively;

[0027] The Mel spectrum features, MFCC features and CQT features are combined to obtain three-channel features.

[0028] Further, obtaining Mel spectrogram features of the normal time domain signal and the corrugation time domain signal specifically includes:

[0029] (1) performing short-time Fourier transform processing on the normal time domain signal and the corrugation time domain signal respectively to obtain a corresponding first normal spectrum and first corrugation spectrum; which can be expressed by the formula:

[0030]

[0031] Wherein, m represents the frame number in the normal time domain signal or the corrugated time domain signal; k represents the frequency point of the first normal spectrum or the first corrugated spectrum; X(m,k) represents the first normal spectrum or the first corrugated spectrum of the mth frame at frequency k; x(n) represents the normal time domain signal or the corrugated time domain signal; w(n-mL) represents the window function; L represents the step size of the window function, that is, the step size of the window function moving on the time axis; N represents the total number of sampling points of the normal time domain signal or the corrugated time domain signal; n represents the sampling point index; e -j2πkn / N represents the basis function of discrete Fourier transform; j represents the imaginary unit;

[0032] (2) The first normal spectrum and the first corrugated spectrum are respectively converted into Mel frequency representation by a Mel filter group; the formula is expressed as follows:

[0033]

[0034] Among them, S mel (m,b) represents the energy of the mth frame on the bth Mel filter; H mel (b,k) represents the response value of the bth Mel filter at frequency k; |X(m,k) 2 represents the energy of the mth frame at frequency k; K represents the total number of frequency points;

[0035] (3) The energy S of the mth frame on the bth Mel filter mel (m,b) is logarithmically operated to obtain the Mel spectrum characteristics; it can be expressed by the formula:

[0036] MelSpectrogram(m,b)=10log 10 (S mel (m,b)+ε)

[0037] Among them, MelSpectrogram(m,b) represents the Mel spectrum feature; ε represents a constant to prevent the logarithm from being zero.

[0038] Further, obtaining MFCC features of the normal time domain signal and the corrugation time domain signal specifically includes:

[0039] The energy S of the mth frame on the bth Mel filter mel (m,b) performs logarithmic operation and discrete cosine transform to obtain MFCC features; it can be expressed by the formula:

[0040]

[0041] Where B represents the total number of Mel filters; c represents the MFCC coefficient index; MFCC(m,c) represents the cth MFCC coefficient of the mth frame.

[0042] Further, obtaining CQT features of the normal time domain signal and the corrugation time domain signal specifically includes:

[0043] (1) CQT processing is performed on the normal time domain signal and the corrugation time domain signal respectively to obtain a corresponding second normal spectrum and a second corrugation spectrum; the formula is expressed as follows:

[0044]

[0045] Where CQT(p) represents the CQT coefficient of the pth frequency band; x(n) represents the normal time domain signal or the corrugation time domain signal; N represents the total number of sampling points of the signal; n represents the sampling point index; h p * (n) represents the conjugate of the filter impulse response of the pth frequency band; f p represents the center frequency of the pth frequency band; f s Indicates the sampling frequency;

[0046] (2) Perform logarithmic transformation on the CQT coefficient CQT(p) of the pth frequency band to obtain the CQT feature; it can be expressed as follows:

[0047] CQT db (p) = 20log 10 (CQT(p)|+ε)

[0048] Among them, CQT db (p) represents the logarithmic CQT characteristic of the p-th frequency band; ε represents a constant that prevents the logarithm from being zero.

[0049] Furthermore, the CNN-Transformer network structure includes a one-dimensional convolutional layer, an attention module and a DeiT module;

[0050] The one-dimensional convolution layer uses a plurality of 3x3 convolution kernels to perform convolution processing on the input three-channel features, output a plurality of first feature maps, and serve as the input of the attention module;

[0051] The attention module is used to process each of the first feature maps and output multiple spatial attention maps as input of the DeiT module;

[0052] The DeiT module is used to process each of the spatial attention maps and output a rail corrugation state recognition result.

[0053] Furthermore, the attention module is composed of a first two-dimensional convolutional layer, a spatial attention module, a channel attention module and a second two-dimensional convolutional layer in sequence;

[0054] The first two-dimensional convolution layer uses a plurality of 3x3 convolution kernels to perform convolution processing on each input first feature map to generate a plurality of second feature maps;

[0055] The spatial attention module is used to pay attention to the spatial position of each of the second feature maps, perform enhancement processing on each of the second feature maps, and generate multiple spatial attention maps;

[0056] The channel attention module is used to adjust the channel weight of each of the spatial attention maps;

[0057] The second two-dimensional convolutional layer is used to reduce the number of channels of each of the spatial attention maps, and input the processed spatial attention maps into the DeiT module.

[0058] Furthermore, the DeiT module is based on the Transformer architecture and consists of a block embedding layer, a Transformer encoder layer, a first fully connected layer, and a second fully connected layer in sequence;

[0059] The block embedding layer is used to divide each input spatial attention map into multiple small blocks, map the multiple small blocks to a high-dimensional space, and output a high-dimensional feature vector;

[0060] The Transformer encoder layer includes a multi-head self-attention mechanism and a feed-forward network; the multi-head self-attention mechanism is used to deeply process each feature map in the high-dimensional feature vector to capture the long-distance dependency in each feature map;

[0061] The first fully connected layer is used to integrate the features extracted by the Transformer encoder layer and integrate them into a feature vector output;

[0062] The second fully connected layer is used to map the output of the first fully connected layer to the category of the time domain signal, and output the rail corrugation state recognition result.

[0063] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a rail corrugation identification method based on acoustic signals, which has the following beneficial effects:

[0064] The present invention identifies rail corrugation based on acoustic signals during train operation. Compared with vibration signals and images, acoustic detection is non-contact and does not require physical changes to the vehicle, simplifying the installation and maintenance of the equipment. Secondly, acoustic signal collection can be carried out under various operating conditions, is not easily affected by changes in train speed or track conditions, and has strong environmental adaptability.

[0065] The present invention proposes a corrugation recognition method based on a CNN-Transformer network, which can effectively improve the recognition accuracy of the rail corrugation state.

[0066] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0068] Figure 1 A schematic flow chart of a rail corrugation identification method based on acoustic signals provided in an embodiment of the present invention.

[0069] Figure 2 A schematic diagram of the CNN-transformer network structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] See also Figure 1 As shown, the embodiment of the present invention discloses a rail corrugation identification method based on acoustic signals, comprising the following steps:

[0072] S1. Obtain a large number of acoustic signals during train operation and divide them into a normal data set and a corrugation data set;

[0073] S2, performing data cleaning processing on the normal data set and the corrugation data set;

[0074] S3, slicing the acoustic signals in the normal data set and the corrugation data set after cleaning respectively to obtain a plurality of normal time domain signals and a plurality of corrugation time domain signals;

[0075] S4, obtaining three-channel characteristics of normal time domain signals and corrugation time domain signals;

[0076] S5. Take the three-channel features as input, take the time domain signal type corresponding to the three-channel features as a label, train the built CNN-Transformer network structure, and obtain a trained CNN-Transformer model;

[0077] S6. The real-time acoustic signal of the target train is identified through the trained CNN-Transformer model to obtain the rail corrugation state recognition result corresponding to the real-time acoustic signal.

[0078] Next, each of the above steps will be described in detail.

[0079] In the above step S1, a large number of acoustic signals in train operation are obtained and divided into a normal data set and a corrugation data set; specifically:

[0080] It can be obtained through the voiceprint collection terminal installed on the floor of the carriage; installing the voiceprint collection terminal on the floor of the carriage can effectively reduce the influence of airborne sound other than corrugation and mechanical noise from the air conditioning device in the car;

[0081] After obtaining a large number of acoustic signals, a sample data set can be established through vibration data comparison and expert manual annotation. The collected acoustic signals of trains during operation are divided into "normal" and "corrugation" by comparing with vibration data and expert manual annotation, corresponding to the normal data set and the corrugation data set.

[0082] S2. Performing data cleaning processing on the normal data set and the corrugation data set. Since the normal data set and the corrugation data set obtained above include mixed corrugation segments in the normal data set and similarly, the corrugation data set also includes normal segments, the data cleaning processing here specifically includes: removing the corrugation acoustic signal in the normal data set; and removing the normal acoustic signal in the corrugation data set.

[0083] S3, slicing the acoustic signals in the cleaned normal data set and the corrugated data set to obtain a plurality of normal time domain signals and a plurality of corrugated time domain signals; in the slicing process, a sliding window with a window length of 2 seconds and a window shift of 1 second is used to obtain a plurality of audio clips as time domain signals;

[0084] S4, obtaining three-channel features of the normal time domain signal and the corrugation time domain signal; specifically: obtaining the Mel spectrum features, MFCC features and CQT features of the normal time domain signal and the corrugation time domain signal respectively; combining the Mel spectrum features, MFCC features and CQT features to obtain the three-channel features;

[0085] In this step, the Mel Frequency Cepstral Coefficient (MFCC), Mel Spectrogram and Constant Q Transform (CQT) are calculated, and these features are combined into three-channel features to enrich the input information of the model. The combination of these three features can first enable the network to capture the characteristics of acoustic signals from different angles and provide more comprehensive information, which is particularly important for improving the model's ability to recognize rail corrugation. Secondly, each feature type has its unique advantages in expressing acoustic signals, such as the sensitivity of MFCC to human speech perception characteristics, the reflection of Mel spectrogram on sound frequency distribution, and the ability of CQT to capture pitch and timbre characteristics. These work together to enhance the network's ability to express acoustic signals. In addition, the combined use of these features can improve the accuracy of the network in identifying specific acoustic patterns, especially specific frequency components or change patterns generated by rail corrugation. At the same time, different features have different robustness to noise and environmental changes. By combining them, the adaptability and stability of the model in various environments can be enhanced. Finally, MFCC, Mel spectrogram, and CQT can capture complementary information in acoustic signals, which not only helps the network understand acoustic signals more comprehensively, but also improves the generalization ability of the model, making it perform better when facing unseen data;

[0086] The above three-channel feature extraction steps specifically include:

[0087] 1. Obtain the Mel spectrum characteristics of normal time domain signals and corrugation time domain signals; specifically including:

[0088] (1) Perform short-time Fourier transform (STFT) processing on the normal time domain signal and the corrugation time domain signal respectively to obtain the corresponding first normal spectrum and first corrugation spectrum; which can be expressed by the formula:

[0089]

[0090] Wherein, m represents the frame number in the normal time domain signal or the corrugated time domain signal; k represents the frequency point of the first normal spectrum or the first corrugated spectrum; X(m,k) represents the first normal spectrum or the first corrugated spectrum of the mth frame at frequency k; x(n) represents the normal time domain signal or the corrugated time domain signal; w(n-mL) represents the window function; L represents the step size of the window function, that is, the step size of the window function moving on the time axis; N represents the total number of sampling points of the normal time domain signal or the corrugated time domain signal; n represents the sampling point index; e -j2πkn / N represents the basis function of discrete Fourier transform; j represents the imaginary unit;

[0091] (2) The first normal spectrum and the first wave-grinding spectrum are respectively converted into Mel frequency representation by the Mel filter bank; the formula is expressed as follows:

[0092]

[0093] Among them, S mel (m,b) represents the energy of the mth frame on the bth Mel filter; H mel (b,k) represents the response value of the bth Mel filter at frequency k; |X(m,k) 2 represents the energy of the mth frame at frequency k; K represents the total number of frequency points;

[0094] (3) The energy S of the mth frame on the bth Mel filter mel (m,b) is logarithmically operated to obtain the Mel spectrum characteristics; it can be expressed by the formula:

[0095] MelSpectrogram(m,b)=10log 10 (S mel (m,b)+ε)

[0096] Among them, MelSpectrogram(m,b) represents the Mel spectrum feature; ε represents a constant to prevent the logarithm from being zero, usually taking a very small positive value, such as 1e-6 or similar values.

[0097] 2. Obtain the MFCC features of normal time domain signals and corrugation time domain signals, including:

[0098] The energy S of the mth frame on the bth Mel filter mel (m,b) performs logarithmic operation and discrete cosine transform (DCT) to obtain MFCC features; it can be expressed by the formula:

[0099]

[0100] Among them, B represents the total number of Mel filters; c represents the MFCC coefficient index, usually the first 12-13 coefficients; MFCC(m,c) represents the cth MFCC coefficient of the mth frame.

[0101] 3. Obtain the CQT characteristics of normal time domain signals and corrugation time domain signals, including:

[0102] (1) CQT processing is performed on the normal time domain signal and the corrugation time domain signal respectively to obtain the corresponding second normal spectrum and second corrugation spectrum; which can be expressed by the formula:

[0103]

[0104] Where CQT(p) represents the CQT coefficient of the pth frequency band; x(n) represents the normal time domain signal or the corrugation time domain signal; N represents the total number of sampling points of the signal; n represents the sampling point index; h p * (n) represents the conjugate of the filter impulse response of the pth frequency band; f p represents the center frequency of the pth frequency band; f s Indicates the sampling frequency;

[0105] (2) Perform logarithmic transformation on the CQT coefficient CQT(p) of the pth frequency band to obtain the CQT feature; it can be expressed as follows:

[0106] CQT db (p) = 20log 10 (CQT(p)|+ε)

[0107] Among them, CQT db (p) represents the logarithmic CQT characteristic of the p-th frequency band; ε represents a constant to prevent the logarithm from being zero, and usually takes a very small positive value, such as 1e-6 or a similar value.

[0108] In the above step S5, the three-channel features are used as input, and the time domain signal type corresponding to the three-channel features is used as a label to train the constructed CNN-Transformer network structure to obtain a trained CNN-Transformer model;

[0109] The present invention uses a neural network (CNN-Transformer network structure) to perform corrugation recognition. The neural network extracts deep-level pattern features by learning a large number of training samples, thereby achieving higher recognition accuracy. In addition, the neural network model can continuously optimize its performance through continuous training and adjustment, reduce the need for manual intervention, improve detection efficiency and accuracy, and adapt to different detection environments and conditions. Specifically, the CNN-Transformer network structure constructed by the present invention includes a one-dimensional convolutional layer, an attention module, and a DeiT module; see Figure 2 As shown:

[0110] 1. One-dimensional convolution layer:

[0111] The one-dimensional convolution layer uses multiple (for example, 16) 3x3 convolution kernels to perform convolution processing on the input three-channel features; the input three-channel feature size is [224, 224, 3]; after convolution processing, multiple (for example, 16) first feature maps are output, and the output size is [224, 224, 16]; each first feature map captures different local features, such as edges and textures, etc. These first feature maps provide rich local information for subsequent processing; finally, these first feature maps are used as input to the attention module.

[0112] 2. Attention Module:

[0113] The attention module is used to process each first feature map, output multiple spatial attention maps, and serve as the input of the DeiT module; specifically:

[0114] The attention module is composed of a first two-dimensional convolutional layer, a spatial attention module, a channel attention module and a second two-dimensional convolutional layer in sequence; wherein:

[0115] (1) The first two-dimensional convolution layer is located at the first layer of the attention module, which uses multiple (e.g., 16) 3x3 convolution kernels to perform convolution processing on each input first feature map to generate multiple (e.g., 16) second feature maps to capture more complex features or combined features;

[0116] (2) The spatial attention module enhances the model’s feature expression capability by focusing on the most important spatial locations in each second feature map and allows the model to focus on key areas in the image, thereby improving the feature’s ability to distinguish features.

[0117] (3) Channel attention module, which is used to adjust the channel weight of each spatial attention map, so that the model can pay more attention to the features that are more beneficial to the current task;

[0118] (4) The second 2D convolutional layer is located at the last layer of the attention module, which is used to reduce the number of channels of each spatial attention map (for example, from 16 to 3), and input the processed spatial attention map into the DeiT module.

[0119] 3. DeiT module

[0120] The DeiT module is used to process each spatial attention map and output the rail corrugation state recognition result; specifically:

[0121] The DeiT module is based on the Transformer architecture and consists of a block embedding layer, a Transformer encoder layer, a first fully connected layer, and a second fully connected layer in sequence;

[0122] (1) The patch embedding layer is used to split each input spatial attention map into multiple patches and map the patches into a high-dimensional space, outputting a high-dimensional feature vector of size [196,768];

[0123] (2) The Transformer encoder layer includes a multi-head self-attention mechanism and a feed-forward network. The multi-head self-attention mechanism is used to deeply process each feature map in the high-dimensional feature vector to capture the long-distance dependencies in each feature map. The input and output sizes of this layer are both [196, 768], which means that each patch still maintains a 768-dimensional feature representation after being processed by the Transformer encoder layer.

[0124] (3) The first fully connected layer is used to integrate the features extracted by the Transformer encoder layer and convert the output of the Transformer encoder layer into a more compact feature representation, that is, to integrate the features of 196 patches into a 512-dimensional vector and output the feature vector;

[0125] (4) The second fully connected layer serves as a classification head, which is used to map the output of the first fully connected layer to the category of the time domain signal and output the final prediction result of the model to realize the recognition of corrugation.

[0126] During the training process, accuracy, precision, recall and F1 score are used as evaluation indicators for corrugation recognition; among them:

[0127] Accuracy is expressed as:

[0128]

[0129] Precision is expressed as:

[0130]

[0131] The recall rate Recall is expressed as:

[0132]

[0133] The F1 score is expressed as:

[0134]

[0135] Among them, TP represents the number of samples correctly classified as "normal type"; TN represents the number of samples correctly classified as "corrugated type"; FP represents the number of samples incorrectly classified as "normal type"; FN represents the number of samples incorrectly classified as "corrugated type".

[0136] After the training is completed, the present invention collects the acoustic signals of the actual train running as the test set samples. These samples are subjected to the above-mentioned feature extraction method to ensure that they have the same format and feature scaling as the training data. During the test, the trained CNN-Transformer model was loaded, and finally an accuracy of 97.31% was achieved on the test set. In order to comprehensively evaluate the performance of the model, the precision, recall and F1 scores were calculated at the same time. The results show that the model shows extremely high precision and recall rates of 96.72% and 97.52% respectively when identifying the "normal" state. Similarly, when identifying the "corrugation" state, the precision and recall rates reached 97.03% and 96.08% respectively. The F1 score can be obtained by comprehensively considering the precision and recall rates: that is, the F1 score of the "normal" state is 97.12%, while the F1 score of the "corrugation" state is 96.56%, both of which are very close to 1, indicating that the classification effect of the model on both categories is very good. The above results fully demonstrate the effectiveness of the proposed rail corrugation recognition method based on acoustic signals, as well as the reliability of the constructed CNN-Transformer network in the rail corrugation recognition task.

[0137] In the above step S6, the real-time acoustic signal of the target train is recognized by the trained CNN-Transformer model to obtain the rail corrugation state recognition result corresponding to the real-time acoustic signal.

[0138] In summary, the embodiments of the present invention provide a rail corrugation identification method based on acoustic signals. Compared with the prior art, the present invention has many advantages in using sound data for corrugation identification. First, acoustic detection is non-contact and does not require physical changes to the vehicle, which simplifies the installation and maintenance of the equipment. Secondly, acoustic signal acquisition can be carried out under various operating conditions, is not easily affected by changes in train speed or track conditions, and has strong environmental adaptability. In addition, sound data can contain information from multiple sound sources at the same time, which can not only monitor rail corrugation, but also evaluate the working conditions of other parts of the train, such as wheels and bearings. Finally, acoustic detection methods are usually low-cost, easy to install, and easy to implement remote monitoring, which reduces the need for manual intervention and improves detection efficiency and accuracy. These advantages make corrugation identification based on sound data a more flexible and efficient solution.

[0139] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0140] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rail corrugation identification method based on acoustic signals, characterized in that: The steps include: Obtain a large number of acoustic signals during train operation and divide them into normal data sets and corrugation data sets; Performing data cleaning processing on the normal data set and the corrugation data set; Slicing the acoustic signals in the normal data set and the corrugation data set after cleaning respectively to obtain a plurality of normal time domain signals and a plurality of corrugation time domain signals; Acquire three-channel characteristics of the normal time domain signal and the corrugation time domain signal; Taking the three-channel features as input, taking the time domain signal type corresponding to the three-channel features as a label, training the constructed CNN-Transformer network structure to obtain a trained CNN-Transformer model; The real-time acoustic signal of the target train is recognized through the trained CNN-Transformer model to obtain the rail corrugation state recognition result corresponding to the real-time acoustic signal.

2. The rail corrugation identification method based on acoustic signals according to claim 1, characterized in that: The data cleaning process of the normal data set and the corrugation data set specifically includes: removing corrugation acoustic signals from the normal data set; The normal acoustic signal in the corrugation dataset is removed.

3. The rail corrugation identification method based on acoustic signals according to claim 1, characterized in that: In the slicing process, a plurality of time domain signals are obtained by using a sliding window with a window length of 2 seconds and a window shift of 1 second.

4. The rail corrugation identification method based on acoustic signals according to claim 1, characterized in that: The step of acquiring the three-channel features of the normal time domain signal and the corrugation time domain signal specifically includes: Acquire Mel spectrogram features, MFCC features and CQT features of the normal time domain signal and the corrugation time domain signal respectively; The Mel spectrum features, MFCC features and CQT features are combined to obtain three-channel features.

5. The rail corrugation identification method based on acoustic signals according to claim 4, characterized in that: Acquiring Mel spectrogram features of the normal time domain signal and the corrugation time domain signal specifically includes: (1) performing short-time Fourier transform processing on the normal time domain signal and the corrugation time domain signal respectively to obtain a corresponding first normal spectrum and a first corrugation spectrum; the formula is expressed as follows: Wherein, m represents the frame number in the normal time domain signal or the corrugated time domain signal; k represents the frequency point of the first normal spectrum or the first corrugated spectrum; X(m,k) represents the first normal spectrum or the first corrugated spectrum of the mth frame at frequency k; x(n) represents the normal time domain signal or the corrugated time domain signal; w(n-mL) represents the window function; L represents the step size of the window function, that is, the step size of the window function moving on the time axis; N represents the total number of sampling points of the normal time domain signal or the corrugated time domain signal; n represents the sampling point index; e -j2πkn / N represents the basis function of discrete Fourier transform; j represents the imaginary unit; (2) The first normal spectrum and the first corrugated spectrum are respectively converted into Mel frequency representation by a Mel filter group; the formula is expressed as follows: Among them, S mel (m,b) represents the energy of the mth frame on the bth Mel filter; H mel (b,k) represents the response value of the bth Mel filter at frequency k; | X(m,k) 2 represents the energy of the mth frame at frequency k; K represents the total number of frequency points; (3) The energy S of the mth frame on the bth Mel filter mel (m,b) is logarithmically operated to obtain the Mel spectrum characteristics; it can be expressed by the formula: MelSpectrogram(m,b)=10log 10 (S mel (m,b)+e) Among them, MelSpectrogram(m,b) represents the Mel spectrum feature; ε represents a constant to prevent the logarithm from being zero.

6. The rail corrugation identification method based on acoustic signals according to claim 5, characterized in that: Acquiring MFCC features of the normal time domain signal and the corrugation time domain signal specifically includes: The energy S of the mth frame on the bth Mel filter mel (m,b) performs logarithmic operation and discrete cosine transform to obtain MFCC features; it can be expressed by the formula: Where B represents the total number of Mel filters; c represents the MFCC coefficient index; MFCC(m,c) represents the cth MFCC coefficient of the mth frame.

7. The rail corrugation identification method based on acoustic signals according to claim 4, characterized in that: Acquiring CQT features of the normal time domain signal and the corrugation time domain signal specifically includes: (1) CQT processing is performed on the normal time domain signal and the corrugation time domain signal respectively to obtain a corresponding second normal spectrum and a second corrugation spectrum; the formula is expressed as follows: Where CQT(p) represents the CQT coefficient of the pth frequency band; x(n) represents the normal time domain signal or the corrugation time domain signal; N represents the total number of sampling points of the signal; n represents the sampling point index; h p * (n) represents the conjugate of the filter impulse response of the pth frequency band; f p represents the center frequency of the pth frequency band; f s Indicates the sampling frequency; (2) Perform logarithmic transformation on the CQT coefficient CQT(p) of the pth frequency band to obtain the CQT feature; it can be expressed as follows: CQT db (p)=20log 10 (CQT(p) | +ε) Among them, CQT db (p) represents the logarithmic CQT characteristic of the p-th frequency band; ε represents a constant that prevents the logarithm from being zero.

8. The rail corrugation identification method based on acoustic signals according to claim 1, characterized in that: The CNN-Transformer network structure includes a one-dimensional convolutional layer, an attention module and a DeiT module; The one-dimensional convolution layer uses a plurality of 3x3 convolution kernels to perform convolution processing on the input three-channel features, output a plurality of first feature maps, and serve as the input of the attention module; The attention module is used to process each of the first feature maps and output multiple spatial attention maps as input of the DeiT module; The DeiT module is used to process each of the spatial attention maps and output a rail corrugation state recognition result.

9. The rail corrugation identification method based on acoustic signals according to claim 8, characterized in that: The attention module is composed of a first two-dimensional convolutional layer, a spatial attention module, a channel attention module and a second two-dimensional convolutional layer in sequence; The first two-dimensional convolution layer uses a plurality of 3x3 convolution kernels to perform convolution processing on each input first feature map to generate a plurality of second feature maps; The spatial attention module is used to pay attention to the spatial position of each of the second feature maps, perform enhancement processing on each of the second feature maps, and generate multiple spatial attention maps; The channel attention module is used to adjust the channel weight of each of the spatial attention maps; The second two-dimensional convolutional layer is used to reduce the number of channels of each of the spatial attention maps, and input the processed spatial attention maps into the DeiT module.

10. The rail corrugation identification method based on acoustic signals according to claim 8, characterized in that: The DeiT module is based on the Transformer architecture and consists of a block embedding layer, a Transformer encoder layer, a first fully connected layer, and a second fully connected layer in sequence; The block embedding layer is used to divide each input spatial attention map into multiple small blocks, map the multiple small blocks to a high-dimensional space, and output a high-dimensional feature vector; The Transformer encoder layer includes a multi-head self-attention mechanism and a feed-forward network; the multi-head self-attention mechanism is used to deeply process each feature map in the high-dimensional feature vector to capture the long-distance dependency in each feature map; The first fully connected layer is used to integrate the features extracted by the Transformer encoder layer and integrate them into a feature vector output; The second fully connected layer is used to map the output of the first fully connected layer to the category of the time domain signal, and output the rail corrugation state recognition result.

Images