Track defect detection method based on LSTM-BiGRU network

Abstract

The present invention provides a track defect detection method based on an LSTM-BiGRU network. The method comprises: acquiring a multimodal vibration signal of a track and converting the multimodal vibration signal into an electrical signal; performing tearing separation on the electrical signal; separating single and double components from the modulated sample signal, applying a long short-term memory network to extract features from the single and double components respectively, and performing weighted processing on the feature signals; calculating information entropy on the weighted feature signals according to an improved ECER entropy change extreme value feature recognition method, and calculating the feature weight of each weighted feature signal based on the information entropy; fusing the weighted feature signals according to the feature weights, inputting the fused features into a BiGRU network for fault detection, and outputting a track defect detection result based on the final representation vector. The embodiments of the present invention can achieve efficient and accurate track defect diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of track detection technology, and in particular to a track defect detection method based on an LSTM-BiGRU network. Background Art

[0002] With the rapid expansion of high-speed railway networks, track defect monitoring has become a key link in ensuring the safety of train operations.

[0003] However, existing methods for detecting track defects face many challenges. Traditional manual inspections and periodic maintenance are not only time-consuming and labor-intensive, but also difficult to maintain consistency in detection conditions and accuracy in actual environments. In addition, these methods are highly dependent on the operator's experience and technical level, resulting in a high subjective error rate. In recent years, although some intelligent monitoring systems based on sensors and preliminary data processing technologies have been applied to track defect detection, they still have significant technical defects. For example, early automatic detection systems usually rely on fixed thresholds to judge the track status. This method cannot adapt to changes in different environmental conditions and is prone to false alarms or missed alarms. In addition, these systems have limited capabilities in extracting track defect characteristics, especially in complex and changing environments, making it difficult to achieve efficient and accurate defect identification. Summary of the Invention

[0004] The purpose of the present invention is to provide a track defect detection method based on LSTM-BiGRU network, aiming to solve the problem that traditional technologies are prone to false positives or missed positives when performing track defect detection, making it difficult to achieve efficient and accurate defect identification.

[0005] In a first aspect, the present invention provides a track defect detection method based on an LSTM-BiGRU network, the method comprising:

[0006] Acquiring a multimodal vibration signal of the track, and converting the multimodal vibration signal into an electrical signal to obtain an electrical signal;

[0007] performing tearing and separation on the electrical signal to combine the electrical signal with the angular frequency and modulation phase of the radio frequency sampling clock to obtain a modulated sample signal;

[0008] Separating single and double components from the modulated sample signal, and applying a long short-term memory network to perform feature extraction on the single and double components respectively to obtain feature signals at each moment corresponding to the single and double components respectively, and performing weighted processing on the feature signals to obtain weighted feature signals;

[0009] Performing information entropy calculation on the weighted feature signals according to the improved ECER entropy change extreme value feature recognition method, and calculating the feature weight of each weighted feature signal according to the information entropy;

[0010] The weighted feature signals are fused according to the feature weights to obtain fused features, and the fused features are input into the BiGRU network for fault detection to obtain a final characterization vector, and a track disease detection result is output according to the final characterization vector.

[0011] In a second aspect, the present invention provides a track defect detection system based on an LSTM-BiGRU network, the system comprising:

[0012] a signal conversion module, configured to obtain a multimodal vibration signal of the track and convert the multimodal vibration signal into an electrical signal to obtain an electrical signal;

[0013] a signal modulation module, configured to perform tearing and separation on the electrical signal, so as to combine the electrical signal with the angular frequency and modulation phase of the radio frequency sampling clock to obtain a modulated sample signal;

[0014] a signal separation module, configured to separate single and double components from the modulated sample signal, and apply a long short-term memory network to perform feature extraction on the single and double components respectively to obtain feature signals at each moment corresponding to the single and double components respectively, and perform weighted processing on the feature signals to obtain weighted feature signals;

[0015] An information entropy acquisition module is used to calculate the information entropy of the weighted feature signal according to the improved ECER entropy change extreme value feature recognition method, and calculate the feature weight of each weighted feature signal according to the information entropy;

[0016] The feature fusion module is used to fuse the weighted feature signals according to the feature weights to obtain fused features, input the fused features into the BiGRU network for fault detection, obtain a final characterization vector, and output the track disease detection result based on the final characterization vector.

[0017] In a third aspect, the present invention provides a storage medium storing one or more programs, which, when executed by a processor, implement the above-mentioned track defect detection method based on the LSTM-BiGRU network.

[0018] In a fourth aspect, the present invention provides an electronic device, comprising a memory and a processor, wherein:

[0019] The memory is used to store computer programs;

[0020] When the processor is used to execute the computer program stored in the memory, it implements the above-mentioned track defect detection method based on the LSTM-BiGRU network.

[0021] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0022] 1. Based on the aforementioned LSTM-BiGRU network-based track defect detection method, the LSTM-BiGRU network, fused with an attention mechanism, achieves accurate monitoring of track defects. This method uses sophisticated preprocessing techniques to remove noise and highlight key fault characteristics. A membership function is introduced to weight the extracted features, increasing the weight of important features. This allows the LSTM network to perform deep feature extraction on the signal, enabling bidirectional information mining and capturing long-term dependencies of track defects. Furthermore, the ECER entropy change extreme value feature recognition method is introduced to address feature uncertainty, and the attention mechanism is used to further optimize feature weights, ensuring the model focuses on key information. Finally, fault classification is performed through fully connected layers and softmax layers, achieving efficient and accurate fault diagnosis. This method is suitable for complex and changing environments, effectively overcoming the shortcomings of earlier systems in adapting to environmental changes and feature extraction, and significantly improving the accuracy and robustness of defect identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a track defect detection method based on an LSTM-BiGRU network proposed in one embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the structure of a track defect detection system based on an LSTM-BiGRU network proposed in one embodiment of the present invention.

[0025] The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are 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 making creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be the common meanings understood by people with ordinary skills in the field to which the invention belongs. The words "including" and similar words used in this article mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0027] like Figure 1As shown, an embodiment of the present invention proposes a track defect detection method based on an LSTM-BiGRU network, the method comprising steps S101 to S105, wherein:

[0028] Step S101: acquiring a multimodal vibration signal of a track, and converting the multimodal vibration signal into an electrical signal to obtain an electrical signal;

[0029] It should be noted that this embodiment uses an IEPE vibration sensor for signal acquisition, covering multimodal vibration signals in different frequency ranges and multiple modes. The multimodal vibration signals can be acceleration, velocity, displacement and other signals.

[0030] Furthermore, in some embodiments, in order to facilitate the transmission of collected signals to a computer for subsequent processing and analysis, a new hyper-dimensional multivariate signal electrical domain conversion technology is used, specifically performing electrical signal conversion according to the following formula:

[0031] ;

[0032] in, is the converted electrical signal, is the frequency response function, is the gain coefficient, is the sensitivity coefficient of the signal acquisition sensor, and are the amplitudes of the i-th and j-th multimodal vibration signals, respectively, and are the angular frequencies of the i-th and j-th multimodal vibration signals, respectively, and are the jth and ith phase angles respectively, is the number of modes, i.e. the intrinsic characteristics of vibration, such as vertical vibration, lateral vibration and other different modes, is the number of types of multimodal vibration signals, that is, the external types of vibration signals, such as vibration caused by external factors such as train running, track defects and environmental factors (such as wind or earthquake), For the The gain coefficient of the multimodal vibration signal, is the sensitivity coefficient of the i-th multimodal vibration signal, is the wave function of the i-th multimodal vibration signal, is the attenuation coefficient of the i-th multimodal vibration signal, t is the current time, is the integration variable.

[0033] Step S102: performing tearing and separation on the electrical signal to combine the electrical signal with the angular frequency and modulation phase of the radio frequency sampling clock to obtain a modulated sample signal;

[0034] It should be pointed out that in this step, a refined preprocessing method based on single and double tearing separation is used for the input electrical signal, which can remove noise and interference and highlight the fault characteristics.

[0035] In some embodiments, tear separation is first performed according to the following formula:

[0036] ;

[0037] In order to analyze the characteristics of the signal and obtain the spectrum information of the modulated signal, the modulated sample signal is expanded. The formula is as follows:

[0038] ;

[0039] in, is the RF sampling clock angular frequency, is the modulation phase, is the modulated sample signal, is the imaginary unit, is the harmonic order, .

[0040] Step S103: separating single and double components from the modulated sample signal, and applying a long short-term memory network to perform feature extraction on the single and double components respectively to obtain feature signals at each moment corresponding to the single and double components respectively, and performing weighted processing on the feature signals to obtain weighted feature signals;

[0041] It should be pointed out that in order to further highlight the fault characteristics, a digital filter is used to analyze the single harmonic and double harmonic components in the spectrum, and the modulated sample signal is filtered to separate the single and double components of the modulated sample signal.

[0042] In some embodiments, a single component is separated according to the following formula:

[0043] ;

[0044] The two components are separated according to the following formula:

[0045] ;

[0046] In addition, the long short-term memory network (LSTM network) is applied to extract features of the modulated single and double component signals respectively, capture long-term dependencies and weight the features through membership functions, amplify the weights of important features, and propose the ECER entropy change extreme value feature recognition method to deal with feature uncertainty. The feature fusion method is used to integrate the processed single and double component features to form a comprehensive feature signal, which provides a basis for model diagnosis.

[0047] In some embodiments, the single-component and dual-component data samples are each input into the LSTM network as independent data samples, with each input matrix being a 1×1024 matrix. The input data is then fed sequentially into the LSTM layers, with a time step length of 1×32, for a total of 32 time steps. This model has three LSTM layers, each with 128 hidden nodes. These three LSTM layers enable in-depth exploration of the deep features of the multimodal rail vibration signal.

[0048] Specifically, feature extraction is performed according to the following formula:

[0049] ;

[0050] in, is the input signal of the LSTM network, including single component and double component, is the characteristic signal output by the LSTM network at the t-1th moment, 、 、 、 is the weight matrix for each gate, 、 、 、 is the bias matrix for each gate, 、 、 is the output value of each gate at the tth moment of the i-th layer network, is the new candidate value of information, is the memory unit at the tth moment, is the memory unit at the t-1th moment, For the The characteristic signal output by the LSTM network at the t-th moment of the layer network, 、 Both are activation functions.

[0051] Step S104: performing information entropy calculation on the weighted feature signals according to the improved ECER entropy change extreme value feature recognition method, and calculating the feature weight of each weighted feature signal according to the information entropy;

[0052] In this step, the feature signal extracted by the LSTM network may contain redundant or noisy features. Therefore, a dynamic Gaussian membership function is introduced. By adjusting parameters such as the number of Gaussian kernels, contribution, and scaling factor, the feature membership values are dynamically calculated. This membership value is used to weight features, amplifying the weights of important features. This emphasizes important features, reduces the influence of irrelevant features, and improves the accuracy of fault diagnosis.

[0053] In some embodiments, the membership value of each characteristic signal is calculated according to:

[0054] ;

[0055] in, is the membership value of each characteristic signal, m is the total number of dynamic Gaussian kernels, For the The contribution of the Gaussian components, For the The scaling factor of the Gaussian kernel, For the The mean of the Gaussian kernels, For the The standard deviation of the Gaussian kernel, is the dynamic adjustment coefficient, For the The weight parameters of the Gaussian components, For the The bias term of the Gaussian components adjusts the activation threshold.

[0056] Then the weighted characteristic signal Its membership value Multiply them together to get the weighted feature signal.

[0057] In addition, due to the uncertainty of the weighted feature signal, an improved ECER entropy change extreme value feature identification method is proposed. This identification method identifies the extreme points of entropy change by calculating the information entropy change of the weighted feature, and further adjusts the weights or selects features for the feature signals corresponding to these extreme points. Information entropy is used to quantify the uncertainty of the feature set (the fault information contained varies in richness) to evaluate the contribution of the feature to fault diagnosis. By calculating the information entropy of the weighted feature signal, the change trend of the entire feature set can be analyzed, thereby determining the extreme points of entropy change that can reflect the significant change in the uncertainty of the feature. Usually, these extreme points of entropy change correspond to feature signals with higher information value, that is, these features contribute more to the classification or recognition task in the feature set.

[0058] In some embodiments, first, all weighted characteristic signals obtained from the single component are aggregated into a first characteristic signal set, and all weighted characteristic signals obtained from the dual component are aggregated into a second characteristic signal set;

[0059] Then the information entropy is calculated according to the following formula:

[0060] ;

[0061] in, is the information entropy of the i-th feature signal in the first feature signal set or the second feature signal set, is the i-th feature signal in the first feature signal set or the second feature signal set, is the probability of the i-th characteristic signal appearing at the j-th moment, is the smoothing factor, h is the total number of moments of the characteristic signal;

[0062] By calculating the information entropy of the weighted feature signal, we can analyze the changing trend of the entire feature set, thereby determining the extreme entropy change points that can reflect the significant change in the uncertainty of the feature. Usually, these extreme entropy change points correspond to feature signals with higher information value, that is, these features contribute more to the classification or recognition task in the feature set.

[0063] Based on the information entropy and entropy change extreme points of the characteristic signals, the richness of the fault information contained in each characteristic signal can be evaluated. Given the differences in the information content of the characteristic signals, it is necessary to adjust the weights of these characteristic signals to optimize the quality of the feature set.

[0064] The feature weight is calculated according to the following formula:

[0065] ;

[0066] in, is the feature weight of the i-th weighted feature signal, is the information entropy of the i-th weighted feature signal, is the extreme value of entropy change, is the adjustment coefficient, is the adjustment coefficient for the interaction between characteristic signals, is the relationship adjustment coefficient between characteristic signals, is the total number of characteristic signals in the first characteristic signal set or the second characteristic signal set, is the covariance between the i-th feature signal and the j-th feature signal, where the two feature signals belong to the same feature signal set, is the correlation between the i-th feature signal and the j-th feature signal, where the two feature signals belong to the same feature signal set. is the jth feature signal in the first feature signal set or the second feature signal set.

[0067] Step S105: The weighted feature signals are fused according to the feature weights to obtain fused features, and the fused features are input into the BiGRU network for fault detection to obtain a final characterization vector, and the track disease detection result is output according to the final characterization vector.

[0068] In this step, by giving priority to assigning more weight to features with higher information value and reducing the weight of features with less information, the effect of eliminating redundant features and highlighting important features is achieved, while reducing the interference of noise features.

[0069] In some embodiments, first, all first target feature weights greater than a first preset threshold are screened out from all feature weights corresponding to the first feature signal set; and all second target feature weights greater than the first preset threshold are screened out from all feature weights corresponding to the second feature signal set;

[0070] Then, the corresponding characteristic signal is adjusted according to the target characteristic weight:

[0071] ;

[0072] ;

[0073] in, is the adjusted characteristic signal set corresponding to a single component, is the adjusted characteristic signal set corresponding to the dual components, 、 、 The first, second, and The first target feature weight, 、 、 The first, second, and The feature signal corresponding to the first target feature weight, For the target feature weights, 、 、 The first, second, and The second target feature weight, 、 、 The first, second, and The characteristic signal corresponding to the second target feature weight, For the target feature weights, 、 are the total number of the first target feature weight and the second target weight respectively.

[0074] In addition, after weighting the weighted feature signals, the single-component and dual-component signals obtained may contain complementary information. In order to fully utilize this information, they need to be fused into a comprehensive feature representation to provide a basis for the next step of model diagnosis. Specifically, the adjusted feature signal set is fused according to the following formula:

[0075] ;

[0076] in, To fusion features, is the global weight parameter, is the correlation coefficient between the kth adjusted feature signal and the jth adjusted feature signal, is the sum of the first target feature weight and the second target weight, is a regularization term, It is a global weight parameter used to balance the contributions of single-component and dual-component features in the final fusion feature. It can be obtained by gradually approaching the optimal value through sampling and updating the posterior distribution using a conventional Bayesian optimization algorithm (such as a Gaussian process). It can also be obtained by using a conventional random search method to randomly select several values to train the model and verify which parameter is the best.

[0077] The BiGRU network is then used to perform deeper information mining on the fused features from both positive and negative directions. The attention mechanism is introduced to strengthen the model's focus on key information, ignore some information irrelevant to the results, and reduce the risk of information loss to improve accuracy.

[0078] The fused features are fed into a BiGRU network. A Bidirectional Gated Recurrent Unit (BiGRU) network is a combination of two GRU models: a forward and a reverse GRU. Because the input is the entire time series, when mining information, the features at a given moment rely on both previous and future information.

[0079] In some embodiments, the calculation formula of the BiGRU network is as follows:

[0080] ;

[0081] in, and are the forward and backward hidden states, respectively, 、 、 are weight matrices for linear transformations of input gates, reset gates, and candidate hidden states, is the long sequence after processing, 、 、 are all normalized weight matrices, is a candidate hidden state, 、 are the outputs of the update gate and reset gate respectively, To update the part that does not pass through the gate, is the hidden state at the t-1th moment, is the hidden state at the t+1th moment, is the input signal at the tth moment, that is , is the output signal at the tth moment, and the arrow is the propagation direction.

[0082] Since not all feature vectors output by the BiGRU network play a decisive role in track disease diagnosis, different BiGRU hidden states are normalized and assigned different weights to achieve the effect of focusing on important information. The expression is as follows:

[0083] ;

[0084] ;

[0085] in, and are bias and weight respectively, The tth weight assigned to the attention mechanism, , is the hidden state at the tth moment, T is the total time step, S is the final representation vector, is a function used to process the output signal of the BiGRU network.

[0086] In addition, in the output of the BiGRU network, since the features at different time steps may have different contributions to the final disease diagnosis, the attention mechanism is introduced to strengthen the model's focus on key information, ignore some information irrelevant to the results, and reduce the risk of information loss to improve accuracy.

[0087] Finally, the final representation vector is sent to the fault classification layer for integration. The classification layer contains the fully connected Softmax Layer, each neuron in the fully connected layer is connected to the neurons in the upper layer, the extracted features are mapped, the Softmax layer performs fault classification, and the results are compared with the fault values set in advance. The one with a larger proportion can be identified as a fault, thus completing the fault diagnosis. Specifically, the final representation vector is input into the fully connected layer and Layers are generated to obtain the probability of each defect category. The maximum defect probability is selected from the probabilities of all defect categories, and a determination is made as to whether the maximum defect probability is greater than a second preset threshold. If the maximum defect probability is greater than the second preset threshold, a track defect is determined. Examples of track defects include: track surface wear due to long-term train operation; track material corrosion due to environmental factors (such as humidity and salt); and cracks due to fatigue, impact, or material defects, which may lead to fracture in severe cases.

[0088] In summary, the aforementioned track defect detection method based on the LSTM-BiGRU network achieves accurate monitoring of track defects by employing an LSTM-BiGRU network fused with an attention mechanism. This method uses sophisticated preprocessing techniques to remove noise and highlight key fault characteristics. A membership function is introduced to weight the extracted features, increasing the weight of important features. This allows deep feature extraction from the signal using the LSTM network, enabling bidirectional information mining and capturing long-term dependencies of track defects. Furthermore, the ECER entropy change extreme value feature recognition method is introduced to address feature uncertainty, and the attention mechanism is used to further optimize feature weights, ensuring the model focuses on key information. Finally, fault classification is performed using fully connected layers and softmax layers, achieving efficient and accurate fault diagnosis. This method is suitable for complex and changing environments, effectively overcoming the shortcomings of earlier systems in adapting to environmental changes and feature extraction, and significantly improving the accuracy and robustness of defect identification.

[0089] like Figure 2 As shown, an embodiment of the present invention further proposes a track defect detection system based on an LSTM-BiGRU network, the system comprising:

[0090] The signal conversion module 10 is used to obtain a multi-modal vibration signal of the track and convert the multi-modal vibration signal into an electrical signal to obtain an electrical signal;

[0091] a signal modulation module 20 for performing tearing and separation on the electrical signal to combine the electrical signal with the angular frequency and modulation phase of the radio frequency sampling clock to obtain a modulated sample signal;

[0092] a signal separation module 30 for separating single and double components from the modulated sample signal, and applying a long short-term memory network to perform feature extraction on the single and double components respectively to obtain feature signals at each moment corresponding to the single and double components respectively, and performing weighted processing on the feature signals to obtain weighted feature signals;

[0093] An information entropy acquisition module 40 is configured to calculate the information entropy of the weighted feature signals according to an improved ECER entropy change extreme value feature recognition method, and calculate the feature weight of each weighted feature signal according to the information entropy;

[0094] The feature fusion module 50 is used to fuse the weighted feature signals according to the feature weights to obtain fused features, input the fused features into the BiGRU network for fault detection, obtain a final characterization vector, and output the track disease detection result based on the final characterization vector.

[0095] On the other hand, the present invention further proposes a storage medium having one or more programs stored thereon, which, when executed by a processor, implements the above-mentioned track defect detection method based on the LSTM-BiGRU network.

[0096] On the other hand, the present invention also proposes an electronic device, including a memory and a processor, wherein the memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the above-mentioned track defect detection method based on the LSTM-BiGRU network.

[0097] Those skilled in the art will appreciate that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device), or in conjunction with such instruction execution system, apparatus, or device. For purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by an instruction execution system, apparatus, or device, or in conjunction with such instruction execution system, apparatus, or device.

[0098] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting, or processing it in another suitable manner as necessary, and then storing it in a computer memory.

[0099] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the aforementioned embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or a combination of the following technologies known in the art may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0100] While the embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations of these embodiments are possible. However, it should be understood that such modifications and variations are within the scope and spirit of the present invention as set forth in the claims. Furthermore, the invention described herein is susceptible to other embodiments and may be practiced or implemented in a variety of ways.

Claims

1. A track defect detection method based on LSTM-BiGRU network, characterized in that: The method comprises: Acquiring a multimodal vibration signal of the track, and converting the multimodal vibration signal into an electrical signal to obtain an electrical signal; performing tearing and separation on the electrical signal to combine the electrical signal with the angular frequency and modulation phase of the radio frequency sampling clock to obtain a modulated sample signal; The tear separation is performed according to the following formula: ； Expand the modulated sample signal: ； in, is the RF sampling clock angular frequency, is the modulation phase, is the modulated sample signal, is the imaginary unit, is the harmonic order, ; Separating single and double components from the modulated sample signal, and applying a long short-term memory network to perform feature extraction on the single and double components respectively to obtain feature signals at each moment corresponding to the single and double components respectively, and performing weighted processing on the feature signals to obtain weighted feature signals; Performing information entropy calculation on the weighted feature signals according to the improved ECER entropy change extreme value feature recognition method, and calculating the feature weight of each weighted feature signal according to the information entropy; Aggregating all weighted characteristic signals obtained from the single component into a first characteristic signal set, and aggregating all weighted characteristic signals obtained from the dual component into a second characteristic signal set; The information entropy is calculated according to the following formula: ； in, is the information entropy of the i-th feature signal in the first feature signal set or the second feature signal set, is the i-th feature signal in the first feature signal set or the second feature signal set, is the probability of the i-th characteristic signal appearing at the j-th moment, is the smoothing factor, h is the total number of moments of the characteristic signal; The feature weight is calculated according to the following formula: ； in, is the feature weight of the i-th weighted feature signal, is the information entropy of the i-th weighted feature signal, is the extreme value of entropy change, is the adjustment coefficient, is the adjustment coefficient for the interaction between characteristic signals, is the relationship adjustment coefficient between characteristic signals, is the total number of characteristic signals in the first characteristic signal set or the second characteristic signal set, is the covariance between the i-th feature signal and the j-th feature signal, where the two feature signals belong to the same feature signal set, is the correlation between the i-th feature signal and the j-th feature signal, where the two feature signals belong to the same feature signal set. is the j-th characteristic signal in the first characteristic signal set or the second characteristic signal set; The weighted feature signals are fused according to the feature weights to obtain fused features, and the fused features are input into the BiGRU network for fault detection to obtain a final characterization vector, and a track disease detection result is output according to the final characterization vector.

2. The track defect detection method based on LSTM-BiGRU network according to claim 1 is characterized in that: The step of obtaining a multimodal vibration signal of the track and converting the multimodal vibration signal into an electrical signal to obtain an electrical signal includes: The electrical signal conversion is performed according to the following formula: ； in, is the converted electrical signal, is the frequency response function, is the gain coefficient, is the sensitivity coefficient of the signal acquisition sensor, and are the amplitudes of the i-th and j-th multimodal vibration signals, respectively, and are the angular frequencies of the i-th and j-th multimodal vibration signals, respectively, and are the jth and ith phase angles respectively, is the number of modes, is the number of types of multimodal vibration signals, For the The gain coefficient of the multimodal vibration signal, is the sensitivity coefficient of the i-th multimodal vibration signal, is the wave function of the i-th multimodal vibration signal, is the attenuation coefficient of the i-th multimodal vibration signal, t is the current time, is the integration variable.

3. The track defect detection method based on LSTM-BiGRU network according to claim 2 is characterized in that: The steps of separating single and double components from the modulated sample signal, and applying a long short-term memory network to extract features of the single and double components respectively to obtain feature signals at each moment corresponding to the single and double components respectively include: Separate the individual components according to the following formula: ； The two components are separated according to the following formula: ； The single component and the double component are each input into the LSTM network as independent data samples, and feature extraction is performed according to the following formula: ； in, is the input signal of the LSTM network, including single component and double component, is the characteristic signal output by the LSTM network at the t-1th moment, 、 、 、 is the weight matrix for each gate, 、 、 、 is the bias matrix for each gate, 、 、 is the output value of each gate at the tth moment of the i-th layer network, is the new candidate value of information, is the memory unit at the tth moment, is the memory unit at the t-1th moment, is the characteristic signal output by the LSTM network at the t-th moment of the layer network, 、 Both are activation functions.

4. The track defect detection method based on LSTM-BiGRU network according to claim 3 is characterized in that: The step of performing weighted processing on the characteristic signal to obtain a weighted characteristic signal comprises: The dynamic Gaussian membership function is introduced to calculate the membership value of each characteristic signal: ； in, is the membership value of each characteristic signal, m is the total number of dynamic Gaussian kernels, For the The contribution of the Gaussian components, For the The scaling factor of the Gaussian kernel, For the The mean of the Gaussian kernels, For the The standard deviation of the Gaussian kernel, is the dynamic adjustment coefficient, For the The weight parameters of the Gaussian components, For the The bias term of the Gaussian components; The characteristic signal Its membership value Multiply them together to get the weighted feature signal.

5. The track defect detection method based on LSTM-BiGRU network according to claim 4 is characterized in that: The step of fusing the weighted feature signals according to the feature weights to obtain fused features comprises: Filter out all first target feature weights greater than a first preset threshold from all feature weights corresponding to the first feature signal set; Filter out all second target feature weights greater than a first preset threshold from all feature weights corresponding to the second feature signal set; According to the target feature weight, the corresponding feature signal is adjusted: ； ； in, is the adjusted characteristic signal set corresponding to a single component, is the adjusted characteristic signal set corresponding to the dual components, 、 、 The first, second, and The first target feature weight, 、 、 The first, second, and The feature signal corresponding to the first target feature weight, For the target feature weights, 、 、 The first, second, and The second target feature weight, 、 、 The first, second, and The characteristic signal corresponding to the second target feature weight, For the target feature weights, 、 are the total number of the first target feature weight and the second target weight respectively; The adjusted feature signal set is fused according to the following formula: ； in, To fusion features, is the global weight parameter, is the correlation coefficient between the kth adjusted feature signal and the jth adjusted feature signal, is the sum of the first target feature weight and the second target weight, is a small regularization term.

6. The track defect detection method based on LSTM-BiGRU network according to claim 5 is characterized in that: The step of inputting the fused features into the BiGRU network for fault detection to obtain the final representation vector includes: The calculation formula of the BiGRU network is as follows: ； in, and are the forward and backward hidden states, respectively, 、 、 are weight matrices for linear transformations of input gates, reset gates, and candidate hidden states, is the long sequence after processing, 、 、 are all normalized weight matrices, is a candidate hidden state, 、 are the outputs of the update gate and reset gate respectively, To update the part that does not pass through the gate, is the hidden state at the t-1th moment, is the hidden state at the t+1th moment, is the input signal at the tth moment, that is , is the output signal at the tth moment, and the arrow is the propagation direction; Normalize the output signal of the BiGRU network and assign different weights: ； ； in, and are bias and weight respectively, The tth weight assigned to the attention mechanism, , is the hidden state at the tth moment, T is the total time step, S is the final representation vector, is a function used to process the output signal of the BiGRU network.

7. The track defect detection method based on LSTM-BiGRU network according to claim 6, characterized in that: The step of outputting the track defect detection result according to the final characterization vector includes: The final representation vector is input into the fully connected layer and Layer, get the probability of each disease category: Screening out a maximum disease probability from the probabilities of all disease categories, and determining whether the maximum disease probability is greater than a second preset threshold; If the maximum probability of damage is greater than a second preset threshold, it is determined that there is damage on the track.

Images