Track disease identification method based on multi-domain feature fusion

Abstract

The invention provides a multi-domain feature fusion track disease identification method, which belongs to the technical field of data processing, and comprises the following steps: collecting vibration acceleration signals of a track in a normal state and a plurality of disease states, carrying out multi-domain basic feature extraction on the collected vibration acceleration signals, and carrying out feature coding on the extracted multi-domain basic features to obtain a multi-domain feature fusion track disease identification result; obtaining a track disease detection data set; inputting the track disease detection data set into a CNN-LSTM multi-domain feature fusion network, performing multi-domain depth feature extraction, and fusing the extracted multi-domain depth features to obtain multi-domain fusion depth features; and taking the multi-domain fusion depth feature as input, and outputting an orbit state and a probability value thereof through a full connection layer and a Softmax function. Therefore, efficient and accurate recognition of various track diseases is realized through multi-domain feature fusion, and the problems of poor multi-disease recognition capability and insufficient feature extraction in track disease recognition in the prior art are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data processing, and in particular relates to a track disease identification method based on multi-domain feature fusion. BACKGROUND

[0002] Track diseases (such as rail corrugation, empty suspension, etc.) can seriously affect the safety and comfort of rail transit. Traditional track disease detection methods mainly rely on manual inspection and visual detection, which have the defects of high cost and being greatly affected by subjective factors and environmental factors. With the development of technology, track disease detection based on vehicle vibration response has become a research direction, but the existing technology has the following shortcomings: unable to identify multiple track diseases using vibration data; relying on single domain (such as only time domain or only frequency domain) data for feature extraction, which is insufficient in time series feature mining and has poor multi-disease identification effect; and the signal characteristics of some track diseases are not obvious in the vehicle vibration response, which affects the identification accuracy. SUMMARY

[0003] Therefore, the purpose of the present application is to provide a track disease identification method based on multi-domain feature fusion, which can integrate time domain, frequency domain and time-frequency domain three-dimensional information, and realize efficient and accurate identification of multiple track diseases through multi-domain feature fusion and Softmax classification.

[0004] In a first aspect, the present application provides a track disease identification method based on multi-domain feature fusion, which comprises the following steps: Collecting vibration acceleration signals under normal track state and multiple disease states, and performing multi-domain basic feature extraction on the collected vibration acceleration signals, and performing feature coding on the extracted multi-domain basic features to obtain a track disease detection dataset; Inputting the track disease detection dataset into a CNN-LSTM multi-domain feature fusion network to perform multi-domain deep feature extraction, and fusing the extracted multi-domain deep features to obtain multi-domain fusion deep features; Taking the multi-domain fusion deep features as input, outputting the track state and its probability value through a fully connected layer and a Softmax function.

[0005] In some embodiments, the multi-domain basic feature extraction on the collected vibration acceleration signals comprises the following steps: Pretreating the collected vibration acceleration signals; wherein the pretreatment includes removing outliers, filling missing values, and standardizing and normalizing processing; The pre-processed vibration acceleration signal is subjected to time domain basic feature extraction, frequency domain basic feature extraction and time-frequency domain basic feature extraction respectively; wherein, the time domain basic features include mean, variance and kurtosis factor; the frequency domain basic features are extracted through fast Fourier transform (FFT) and power spectral density analysis (PSD); and the time-frequency domain basic features are extracted through short-time Fourier transform (STFT) and wavelet transform.

[0006] In some embodiments, the extracted multi-domain basic features are subjected to feature coding, including the following steps: The extracted time domain basic features are sequentially sorted in a preset order to form a one-dimensional time domain vector; The extracted frequency domain basic features are sequentially sorted according to frequency correlation logic to form a one-dimensional frequency domain vector; The extracted time-frequency domain basic features are uniformly sized and gray-scale normalized to generate a two-dimensional time-frequency domain vector; wherein, the time-frequency domain basic features include STFT time-frequency map and wavelet transform time-frequency map, both in the form of two-dimensional matrix.

[0007] In some embodiments, the CNN-LSTM multi-domain feature fusion network includes a 1D-CNN module, an LSTM module and a 2D-CNN module, and the track disease detection data set is input into the CNN-LSTM multi-domain feature fusion network for multi-domain deep feature extraction, including the following steps: The one-dimensional frequency domain vector in the track disease detection data set is input into the 1D-CNN module, which preliminarily extracts the global features of the frequency domain through a wide convolution block, and then gradually extracts the local features of the frequency domain through three convolution blocks, and finally outputs the frequency domain deep features; The one-dimensional time domain vector in the track disease detection data set is input into the LSTM module including three layers of LSTM, which learns the long-term dependence of time series through the gating mechanism of input gate, forget gate, memory cell and output gate, and outputs the time domain deep features; The two-dimensional time-frequency domain vector in the track disease detection data set is input into the 2D-CNN module, which gradually excavates the spatial correlation features of time and frequency dimensions within the time-frequency domain through four residual blocks including convolution layer, batch normalization layer, ReLU activation layer and residual connection, and then compresses the dimensions through the AvgPool2d layer to output the time-frequency domain deep features.

[0008] In some embodiments, the extracted multi-domain deep features are fused, including the following steps: The frequency domain deep features, the time domain deep features and the time-frequency domain deep features are converted into feature vectors of uniform dimensions; The feature vectors of uniform dimensions are spliced into a fusion feature matrix, and the corresponding attention weights are obtained through an attention scoring network; The multi-domain fusion deep features are obtained by weighting and fusing the feature vectors in the unified dimension and the corresponding attention weights.

[0009] In some embodiments, the method further comprises the following steps: The multi-domain fusion deep features are input into a fully connected layer, and are mapped into a multi-dimensional score vector matching the number of disease categories through linear transformation. The multi-dimensional score vector is converted into probability values corresponding to each category of disease through a Softmax function, and the category corresponding to the maximum probability value is taken as the final recognition result.

[0010] In some embodiments, the method further comprises the following steps: The CNN-LSTM multi-domain feature fusion network is optimized and cyclically parameterized through three types of indexes of loss rate, confusion matrix and accuracy.

[0011] In a second aspect, the application further provides a track disease recognition device based on multi-domain feature fusion, which comprises: The acquisition module is configured to acquire vibration acceleration signals under normal track conditions and under various disease conditions, extract multi-domain basic features from the acquired vibration acceleration signals, encode the extracted multi-domain basic features, and obtain a track disease detection dataset. The feature fusion module is configured to input the track disease detection dataset into a CNN-LSTM multi-domain feature fusion network, extract multi-domain deep features, fuse the extracted multi-domain deep features, and obtain multi-domain fusion deep features. The recognition module is configured to input the multi-domain fusion deep features, output track conditions and their probability values through a fully connected layer and a Softmax function.

[0012] In a third aspect, the application further provides an electronic device, which comprises a processor, a memory and a bus, the memory stores machine readable instructions executable by the processor, and the processor and the memory communicate through the bus when the electronic device is running, and the machine readable instructions are executed by the processor to perform the steps of the track disease recognition method based on multi-domain feature fusion according to any one of the first aspect.

[0013] In a fourth aspect, the application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to perform the steps of the track disease recognition method based on multi-domain feature fusion according to any one of the first aspect.

[0014] This application describes a multi-domain feature fusion-based method for identifying track defects. It collects vibration acceleration signals under normal track conditions and various defect conditions, extracts multi-domain basic features from the collected vibration acceleration signals, and encodes these extracted features to obtain a track defect detection dataset. This dataset is then input into a CNN-LSTM multi-domain feature fusion network for multi-domain deep feature extraction. The extracted deep features are then fused to obtain multi-domain fused deep features. These multi-domain fused deep features are used as input, and the track state and its probability value are output through a fully connected layer and a Softmax function. This multi-domain feature fusion method achieves efficient and accurate identification of various track defects, effectively solving the problems of poor multi-defect identification capability and insufficient feature extraction in existing track defect identification technologies. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart of the track defect identification method based on multi-domain feature fusion described in an embodiment of this application is shown; Figure 2 An architecture diagram of the multi-domain feature fusion-based track defect identification method described in an embodiment of this application is shown; Figure 3 A schematic diagram of the structure of the CNN-LSTM multi-domain feature fusion network described in an embodiment of this application is shown; Figure 4 This paper shows a schematic diagram of the structure of the track defect identification device with multi-domain feature fusion according to an embodiment of this application; Figure 5 A schematic diagram of the structure of the electronic device described in an embodiment of this application is shown. Detailed Implementation

[0017] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. It should be understood that the drawings in the present application serve only the purpose of description and illustration, and do not serve to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the present application show the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowcharts can not be implemented in sequence, and the steps without logical context relationship can be reversed in sequence or implemented simultaneously. In addition, one or more other operations can be added to the flowcharts or one or more operations can be removed from the flowcharts under the guidance of the content of the present application.

[0018] In addition, the described embodiments are only some of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0019] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0020] In view of the technical problems presented in the background art, the present application provides a track disease identification method based on multi-domain feature fusion, which can integrate time domain, frequency domain, and time-frequency domain three-dimensional information, and realize efficient and accurate identification of multiple track diseases through multi-domain feature fusion and Softmax classification.

[0021] Referring to the drawings accompanying the Figure 1 and the drawings accompanying the Figure 2 , the present application provides a track disease identification method based on multi-domain feature fusion, which comprises the following steps: S1, collecting vibration acceleration signals under normal track state and under multiple disease states, and performing multi-domain basic feature extraction on the collected vibration acceleration signals, and performing feature coding on the extracted multi-domain basic features to obtain a track disease detection data set; S2, inputting the track disease detection data set into a CNN-LSTM multi-domain feature fusion network, performing multi-domain deep feature extraction, and fusing the extracted multi-domain deep features to obtain multi-domain fusion deep features; S3, taking the multi-domain fusion deep features as input, outputting track state and its probability value through a fully connected layer and a Softmax function.

[0022] Specifically, in step S1, the track vibration acceleration original signal collected by the track disease detection device (such as the acceleration sensor installed on the vehicle axle box) reflects the vibration response of the track under the action of the vehicle load, which contains the characteristic information of different diseases such as normal state, corrugation, and empty suspension. In order to provide a reliable data basis for subsequent multi-domain feature extraction and model training, the collected vibration acceleration signal also needs to be preprocessed before multi-domain basic feature extraction.

[0023] In an embodiment, the preprocessing of the collected vibration acceleration signal includes removing outliers, filling missing values, and standardizing and normalizing. For example, the three-sigma rule is used to identify and remove outliers; for missing values in the signal, linear interpolation is used for filling; the data is standardized to eliminate dimension differences, and the normalized data is mapped to the interval [0, 1].

[0024] Among them, the multi-domain basic feature extraction is to mine three types of basic features in time domain, frequency domain, and time-frequency domain from the preprocessed vibration acceleration signal. The three types of basic features reflect the signal rules of track diseases from different dimensions.

[0025] The time domain basic feature is directly extracted from the preprocessed vibration acceleration signal, which reflects the overall distribution, fluctuation degree and impact characteristics of the signal, including mean, variance and kurtosis factor. Among them, the mean is the average level of the time domain signal, the variance is the fluctuation range of the time domain signal deviating from the mean, and the kurtosis factor is the peak degree index of the time domain signal. The frequency domain basic feature is extracted from the frequency domain signal converted from the preprocessed vibration acceleration signal, which reflects the frequency distribution and energy concentration frequency band of the signal. Specifically, first, the time domain signal is converted into a frequency domain signal through fast Fourier transform (FFT), and the peak frequency and main band energy are calculated; then the energy characteristics are analyzed through power spectral density (PSD), and the PSD peak and gravity frequency are calculated. The time-frequency domain basic feature reflects the dynamic change of the signal frequency with time, wherein the preprocessed vibration acceleration signal is obtained through short-time Fourier transform (STFT) to obtain the STFT time-frequency diagram, and the preprocessed vibration acceleration signal is wavelet transformed to obtain the wavelet transform time-frequency diagram.

[0026] It should be noted that the corresponding calculation formula should be a technical means familiar to those skilled in the art, which is not described here.

[0027] Further, all time-domain basic features are arranged in a preset order (such as mean, variance, kurtosis factor in turn) to form a one-dimensional time-domain vector; all frequency-domain basic features are spliced in turn according to a frequency correlation logic (such as peak frequency, main band energy, PSD peak, PSD gravity frequency in turn, to ensure the physical meaning coherence between features), to form a one-dimensional frequency-domain vector; the STFT time-frequency diagram and the wavelet transform time-frequency diagram are adjusted to a preset uniform size and subjected to gray scale normalization, to generate a two-dimensional time-frequency domain vector, which is used to match the input requirements of a subsequent 2D-CNN module.

[0028] And the one-dimensional time-domain vector, the one-dimensional frequency-domain vector, and the two-dimensional time-frequency domain vector in the same track detection scene are packaged into a sample, while the disease label corresponding to the sample is associated, to form a sample unit with a label, and finally a track disease detection dataset is obtained.

[0029] In step S2, referring to the description accompanying drawings Figure 3 The CNN-LSTM multi-domain feature fusion network includes a 1D-CNN module, an LSTM module, and a 2D-CNN module, which are used to input the one-dimensional time-domain vector, the one-dimensional frequency-domain vector, and the two-dimensional time-frequency domain vector in the track disease detection dataset constructed in step S1 into the corresponding modules, respectively, for multi-domain feature deep processing.

[0030] CNN-LSTM is a hybrid model combining convolutional neural network CNN and long short-term memory network LSTM. CNN (Convolutional Neural Network) is a deep learning model specially designed for processing grid-shaped data (such as images, videos). Its core is to automatically scan the input data through the filter in the convolutional layer, efficiently extract local features from edge, texture to object components, and then compress key information through the pooling layer and enhance the model generalization ability. This local connection and weight sharing mechanism can significantly reduce the number of parameters and efficiently understand the spatial hierarchy of images. LSTM (Long Short-Term Memory Network) is a special recurrent neural network (RNN), whose core is the unique “gating mechanism” (including input gate, forget gate and output gate), which can precisely control the flow of information and effectively avoid the problem of gradient vanishing or explosion, which makes LSTM particularly good at processing long time series problems. CNN-LSTM combines the two, through CNN to effectively extract and understand two-dimensional features, through LSTM to process long time series data, while capturing the dynamic dependence between two-dimensional image features and long time series features, laying the foundation for multi-domain feature fusion. Multi-domain feature fusion refers to the integration of multiple features extracted from different domains of the same data object or phenomenon through certain algorithms and strategies, compensating for the limitations of single domain features, utilizing the complementarity between different domain features to enhance the robustness and discriminability of the model, thereby improving the accuracy of classification or recognition.

[0031] Specifically, the one-dimensional frequency domain vector input 1D-CNN module first expands the receptive field of the convolution kernel through a wide convolution block to preliminarily extract the global features of the frequency domain; then gradually extracts the local spatial features of the frequency domain through three convolution blocks (each convolution block contains a convolution layer, batch normalization, Leaky ReLU activation, and a pooling layer); and finally outputs the frequency domain deep features. The one-dimensional time domain vector input LSTM module includes three layers of LSTM, each layer of LSTM includes an input gate, a forget gate, a memory cell, and an output gate, which captures the long-term time sequence dependence of the time domain signal (such as the dynamic change rule of the track disease over time) through the gating mechanism, and finally outputs the time domain deep features. The two-dimensional time-frequency domain vector input 2D-CNN module first extracts the basic features by the residual block, extracts the basic features of the time-frequency graph through stacking four residual blocks, and then outputs the time-frequency domain deep features through convolution, normalization, activation, and residual connection in cooperation with AvgPool2d compression dimension.

[0032] Further, the obtained time domain deep features, frequency domain deep features, and time-frequency domain deep features are converted into vectors of the same dimension, then the feature vectors after uniform dimension are spliced into a fusion feature matrix, and the corresponding attention weights are obtained through the attention scoring network; the uniform dimension feature vectors of each domain are multiplied by the corresponding attention weights to obtain weighted features, and the three weighted feature vectors are added element by element to obtain the final multi-domain fusion deep features.

[0033] In step S3, the multi-domain fusion deep features are flattened and input into a fully connected layer, and then the probability distribution of different track states (such as normal, corrugation, and empty suspension) is output through the Softmax function, and the class corresponding to the maximum probability value is taken as the final recognition result.

[0034] Thus, based on the CNN-LSTM multi-domain feature fusion network, the 1D-CNN module is used to capture the frequency domain local features, the LSTM module is used to capture the time sequence dependence of the time domain, and the 2D-CNN module is used to extract the two-dimensional features of the time-frequency domain, and then the three types of features are integrated through the attention mechanism, thereby solving the problem of insufficient recognition ability of single domain features and realizing accurate multi-class recognition of track diseases.

[0035] Further, in other embodiments, the constructed track disease detection dataset can be divided into a training set and a test set, and according to the evaluation results of the loss rate, the confusion matrix and the accuracy, the performance iteration improvement can be realized by adjusting the CNN-LSTM multi-domain feature fusion network parameters and the training strategy. Specifically, the model composed of the steps S2 and S3, that is, the "1D-CNN module, LSTM, 2D-CNN module, attention fusion, Softmax function" model is iteratively trained using the training set. After each round of training, the classification results of the test set are used to calculate the loss rate (cross-entropy loss), the confusion matrix and the accuracy, and the model parameters (such as the 1D-CNN convolution kernel size, the LSTM hidden layer neuron number and the attention weight initialization method) are adjusted according to the evaluation indexes; when the fault diagnosis accuracy of the model on the test set reaches the peak value and there is no improvement for continuous multiple rounds, the training is stopped, the model parameters at this time are saved, and the final track disease recognition model is obtained. For example, when the accuracy of the test set reaches more than 95% and the recall rate of each type of disease reaches more than 90%, all the parameters (including the CNN convolution kernel weight, the LSTM gate weight and the attention mechanism weight) of the model are saved, and the final deployable track disease recognition model is generated.

[0036] The track disease recognition method provided in the application integrates time domain, frequency domain and time-frequency domain three-dimensional information, avoids the limitations of single domain features, and comprehensively captures the feature performance of track diseases. The CNN-LSTM multi-domain feature fusion network is used to improve the feature extraction depth, and the multi-domain feature fusion and Softmax classification are used to realize efficient and accurate recognition of multiple track diseases, thereby providing strong support for track safety maintenance.

[0037] Based on the same inventive concept, the application also provides a track disease recognition device based on multi-domain feature fusion. Since the principle of the device in the application solves the problem similarly to the track disease recognition method based on multi-domain feature fusion described above, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described here.

[0038] As shown in the accompanying drawings Figure 4 The application also provides a track disease recognition device based on multi-domain feature fusion, which comprises: The acquisition module 401 is configured to acquire vibration acceleration signals under normal track state and under multiple disease states, perform multi-domain basic feature extraction on the acquired vibration acceleration signals, perform feature encoding on the extracted multi-domain basic features, and obtain a track disease detection dataset; The feature fusion module 402 is configured to input the track disease detection dataset into a CNN-LSTM multi-domain feature fusion network, perform multi-domain deep feature extraction, fuse the extracted multi-domain deep features, and obtain multi-domain fusion deep features. The identification module 403 is configured to input the multi-domain fusion deep features as input, and output track states and probability values thereof through a full connection layer and a Softmax function.

[0039] In some embodiments, the collection module 401 performs multi-domain basic feature extraction on the collected vibration acceleration signals, including: pre-processing the collected vibration acceleration signals; wherein the pre-processing includes removing outliers, filling missing values, and standardizing and normalizing; performing time-domain basic feature extraction, frequency-domain basic feature extraction, and time-frequency domain basic feature extraction on the pre-processed vibration acceleration signals respectively; wherein the time-domain basic features include mean, variance, and kurtosis factor; the frequency-domain basic features are extracted through fast Fourier transform (FFT) and power spectral density (PSD) analysis; and the time-frequency domain basic features are extracted through short-time Fourier transform (STFT) and wavelet transform.

[0040] In some embodiments, the collection module 401 performs feature encoding on the extracted multi-domain basic features, including: sequentially ordering the extracted time-domain basic features in a preset order to form a one-dimensional time-domain vector; sequentially ordering the extracted frequency-domain basic features according to frequency correlation logic to form a one-dimensional frequency-domain vector; and generating a two-dimensional time-frequency domain vector after uniformly sizing and grayscale normalizing the extracted time-frequency domain basic features; wherein the time-frequency domain basic features include STFT time-frequency maps and wavelet transform time-frequency maps, both of which are two-dimensional matrix forms.

[0041] In some embodiments, the CNN-LSTM multi-domain feature fusion network includes a 1D-CNN module, an LSTM module, and a 2D-CNN module, the feature fusion module 402 inputs the track disease detection dataset into the CNN-LSTM multi-domain feature fusion network to perform multi-domain deep feature extraction, including: inputting the one-dimensional frequency-domain vector in the track disease detection dataset into the 1D-CNN module, preliminarily extracting global features of the frequency domain through a wide convolution block, gradually extracting local features of the frequency domain through three convolution blocks, and finally outputting frequency domain deep features; inputting the one-dimensional time-domain vector in the track disease detection dataset into the LSTM module including three layers of LSTM, learning long-term dependencies of time series through a gating mechanism of an input gate, a forget gate, a memory cell, and an output gate, and outputting time domain deep features; and inputting the two-dimensional time-frequency domain vector in the track disease detection dataset into the 2D-CNN module, gradually mining spatial correlation features of time and frequency dimensions within the time-frequency domain through four residual blocks including a convolution layer, a batch normalization layer, a ReLU activation layer, and a residual connection, and outputting time-frequency domain deep features through an AvgPool2d layer to compress the dimensions.

[0042] In some embodiments, the feature fusion module 402 fuses the extracted multi-domain deep features, including: converting the frequency domain deep features, the time domain deep features and the time-frequency domain deep features into feature vectors of uniform dimensions; splicing the feature vectors after uniform dimension into a fusion feature matrix, and obtaining the corresponding attention weights through an attention scoring network; performing weighted fusion based on the feature vectors after uniform dimension and the corresponding attention weights to obtain multi-domain fusion deep features.

[0043] In some embodiments, the identification module 403 takes the multi-domain fusion deep features as input, outputs the track state and its probability value through a fully connected layer and a Softmax function, including: inputting the multi-domain fusion deep features into a fully connected layer, mapping them into a multi-dimensional score vector matching the number of disease categories through linear transformation; converting the multi-dimensional score vector into the probability value corresponding to each category of disease through a Softmax function, and taking the category corresponding to the maximum probability value as the final identification result.

[0044] In some embodiments, the device further comprises: An optimization module for optimizing the CNN-LSTM multi-domain feature fusion network through loss rate, confusion matrix and accuracy rate three types of indexes.

[0045] The track disease identification device of multi-domain feature fusion described in the present application, through the acquisition module, acquires vibration acceleration signals under normal track state and under various disease states, and performs multi-domain basic feature extraction on the acquired vibration acceleration signals, and feature encoding on the extracted multi-domain basic features to obtain a track disease detection dataset; through the feature fusion module, the track disease detection dataset is input into a CNN-LSTM multi-domain feature fusion network, multi-domain deep feature extraction is performed, and the extracted multi-domain deep features are fused to obtain multi-domain fusion deep features; through the identification module, the multi-domain fusion deep features are taken as input, and the track state and its probability value are output through a fully connected layer and a Softmax function. Thus, through multi-domain feature fusion, efficient and accurate identification of various track diseases is realized, effectively solving the problems of poor multi-disease identification capability and insufficient feature extraction in the prior art track disease identification.

[0046] Based on the same concept of the present application, as described in the specification Figure 5As shown, the electronic device 500 provided by the embodiment of the present application includes at least one processor 501, at least one network interface 504 or other user interface 503, a memory 505, and at least one communication bus 502. The communication bus 502 is used to realize the connection and communication between the components. The electronic device 500 can optionally include a user interface 503, including a display (for example, a touch screen, an LCD, a CRT, holographic imaging (Holographic) or a projector, etc.), a keyboard or a clicking device (for example, a mouse, a trackball, a touchpad or a touch screen, etc.).

[0047] The memory 505 can include read-only memory and random access memory, and provide instructions and data for the processor 501. A part of the memory 505 can also include non-volatile random access memory (NVRAM).

[0048] In some embodiments, the memory 505 stores the following elements, executable modules or data structures, or a subset thereof, or an extended set thereof: An operating system 5051 includes various system programs, used to realize various basic services and process hardware-based tasks; Application modules 5052 include various application programs, such as a desktop (launcher), a media player (MediaPlayer), a browser (Browser), etc., used to realize various application services.

[0049] In the embodiment of the present application, the processor 501 is used to execute the steps of the track disease identification method by calling the programs or instructions stored in the memory 505.

[0050] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to perform the steps of the track disease identification method.

[0051] Specifically, the storage medium can be a general storage medium, such as a mobile disk, a hard disk, etc., and the computer program stored on the storage medium can integrate time domain, frequency domain, and three-dimensional information of time-frequency domain when executed, and realize efficient and accurate identification of various track diseases through multi-domain feature fusion and Softmax classification.

[0052] In the embodiments of the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. The embodiments described above are merely exemplary, for example, the division of units is only a logical function division, and there can be another division manner in actual implementation; for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between units can be indirect coupling or communication connection through some interfaces, and can be electrical, mechanical or other forms.

[0053] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0054] In addition, each functional unit in the embodiments of the present application can be integrated into a processing unit, or each unit can be a physically independent unit, or two or more units can be integrated into one unit.

[0055] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in the various embodiments of the present application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various program codes that can store program codes.

[0056] Finally, it should be noted that the above examples are merely specific embodiments of the present application, and are used to illustrate the technical solutions of the present application, but are not limiting thereof, and the protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing examples, it should be understood by those skilled in the art that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing examples, or make equivalent replacements to some of the technical features, within the technical range disclosed by the present application, and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application. All should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for identifying track defects through multi-domain feature fusion, characterized in that, The method includes the following steps: Vibration acceleration signals under normal track conditions and various defect conditions are collected, and multi-domain basic features are extracted from the collected vibration acceleration signals. The extracted multi-domain basic features are then feature-encoded to obtain a track defect detection dataset. The track defect detection dataset is input into a CNN-LSTM multi-domain feature fusion network to extract multi-domain deep features, and the extracted multi-domain deep features are fused to obtain multi-domain fused deep features. The multi-domain fusion deep features are used as input, and the orbital state and its probability value are output through a fully connected layer and a Softmax function.

2. The method for identifying track defects by multi-domain feature fusion according to claim 1, characterized in that, The multi-domain basic feature extraction of the acquired vibration acceleration signal includes the following steps: The acquired vibration acceleration signals are preprocessed, including outlier removal, missing value filling, and standardization / normalization. The preprocessed vibration acceleration signal was subjected to time-domain, frequency-domain, and time-frequency-domain basic feature extraction respectively. The time-domain basic features included mean, variance, and kurtosis factor. The frequency-domain basic features were extracted by Fast Fourier Transform (FFT) and Power Spectral Density Analysis (PSD). The time-frequency-domain basic features were extracted by Short-Time Fourier Transform (STFT) and Wavelet Transform.

3. The method for identifying track defects by multi-domain feature fusion according to claim 2, characterized in that, The feature encoding of the extracted multi-domain basic features includes the following steps: The extracted time-domain basic features are sorted sequentially according to a preset order to form a one-dimensional time-domain vector; The extracted frequency domain basic features are sorted sequentially according to the frequency correlation logic to form a one-dimensional frequency domain vector; The extracted time-frequency domain basic features are uniformly sized and grayscale normalized to generate a two-dimensional time-frequency domain vector; wherein, the time-frequency domain basic features include STFT time-frequency maps and wavelet transform time-frequency maps, both of which are in two-dimensional matrix form.

4. The method for identifying track defects by multi-domain feature fusion according to claim 3, characterized in that, The CNN-LSTM multi-domain feature fusion network includes a 1D-CNN module, an LSTM module, and a 2D-CNN module. The step of inputting the track defect detection dataset into the CNN-LSTM multi-domain feature fusion network for multi-domain deep feature extraction includes the following steps: The one-dimensional frequency domain vector in the track defect detection dataset is input into the 1D-CNN module. First, the global features in the frequency domain are initially extracted through a wide convolutional block, and then the local features in the frequency domain are gradually extracted through three convolutional blocks. Finally, the frequency domain depth features are output. The one-dimensional time-domain vector from the track defect detection dataset is input into an LSTM module consisting of three LSTM layers. The long-term dependencies of the time series are learned through the gating mechanism of input gate, forget gate, memory cell and output gate, and the time-domain deep features are output. The two-dimensional time-frequency domain vectors in the track defect detection dataset are input into the 2D-CNN module. First, the spatial correlation features of time and frequency dimensions in the time-frequency domain are gradually mined through four residual blocks including convolutional layers, batch normalization layers, ReLU activation layers and residual connections. Then, the dimensions are compressed through the AvgPool2d layer to output the time-frequency domain depth features.

5. The method for identifying track defects by multi-domain feature fusion according to claim 4, characterized in that, The process of fusing the extracted multi-domain deep features includes the following steps: Convert the frequency domain deep features, time domain deep features, and time-frequency domain deep features into a feature vector of a uniform dimension; The feature vectors with unified dimensions are concatenated into a fused feature matrix, and the corresponding attention weights are obtained through an attention scoring network. We obtain multi-domain fusion deep features by weighting and fusing the feature vectors with a unified dimension and the corresponding attention weights.

6. The method for identifying track defects by multi-domain feature fusion according to claim 4, characterized in that, The process of taking the multi-domain fused deep features as input and outputting the orbital state and its probability value through a fully connected layer and a Softmax function includes the following steps: The multi-domain fusion deep features are input into a fully connected layer and mapped into a multi-dimensional score vector that matches the number of disease categories through a linear transformation. The multidimensional score vector is converted into probability values ​​for each disease category using the Softmax function, and the category corresponding to the highest probability value is taken as the final identification result.

7. The method for identifying track defects by multi-domain feature fusion according to claim 6, characterized in that, The method further includes the following steps: The CNN-LSTM multi-domain feature fusion network was optimized and its parameters were tuned cyclically using three metrics: loss rate, confusion matrix, and accuracy.

8. A track defect identification device based on multi-domain feature fusion, characterized in that, The device includes: The acquisition module is used to acquire vibration acceleration signals under normal track conditions and various defect conditions, and to extract multi-domain basic features from the acquired vibration acceleration signals and perform feature encoding on the extracted multi-domain basic features to obtain a track defect detection dataset. The feature fusion module is used to input the track defect detection dataset into the CNN-LSTM multi-domain feature fusion network, extract multi-domain deep features, and fuse the extracted multi-domain deep features to obtain multi-domain fused deep features. The recognition module is used to take the multi-domain fusion deep features as input and output the orbital state and its probability value through a fully connected layer and a Softmax function.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of a track defect identification method based on multi-domain feature fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a track defect identification method based on multi-domain feature fusion as described in any one of claims 1 to 7.

Images