Multi-level track disease identification system based on vehicle body vibration data

Abstract

The invention discloses a multi-level track disease recognition system based on vehicle body vibration data, and particularly relates to the field of blasting auxiliary devices. Comprising the steps that a data acquisition module acquires heterogeneity data of a track environment; the environment adaptive module analyzes the correlation between the vehicle body vibration data and the heterogeneity data, constructs a dual-view similarity matrix, and realizes adaptive classification of the track environment through a spectral clustering method; the label enhancement module generates high-quality pseudo labels and forms a completely labeled sample set; the disease recognition module adopts an attention-guided multi-task cascade CNN architecture, and outputs a multi-level recognition result containing disease positions and severity. By the adoption of the technical scheme, the problem that an existing track state evaluation method cannot evaluate the real state of track diseases in a complex environment is solved, and good performance is achieved in the aspect of geometric and structural disease recognition.

Description

Technical Field

[0001] The invention relates to the field of blasting auxiliary devices, and in particular to a multi-level track defect recognition system based on vehicle body vibration data. Background Art

[0002] Timely detection and maintenance of track defects are crucial to ensure the safety of railway operations. With the rapid development of urban rail transit, track defect detection faces higher timeliness and accuracy requirements. At present, track defects mainly include two levels: geometric defects and structural defects. The geometric state of the track is mainly represented by the following geometric parameters: longitudinal level, alignment, cross-level, gauge, twist, etc. All countries in the world have stipulated the deviation management values ​​of track geometric parameters, and graded the degree of deviation according to management needs. The state exceeding these management values ​​is track geometric defects. Structural defects mainly include surface defects of rails (such as side wear, vertical wear, etc.). Similar to geometric defects, countries around the world have formulated corresponding detection standards and management systems. The detection unit needs to grade and evaluate the defects according to the specific damage type, location and degree. These defects will directly affect the safety of vehicle operation and the comfort of passengers.

[0003] Traditional track inspection mainly relies on inspection vehicles and manual inspections. Inspection vehicles can combine multiple detection technologies such as ultrasound and eddy current to accurately evaluate the track status. However, these detection methods can usually only be carried out during the maintenance window, the frequency of detection is limited, and the detection cost is high. Since there is a significant correlation between vehicle body vibration and track status, high-frequency track status monitoring can be achieved by collecting vibration through portable on-board equipment. In order to improve detection efficiency, track status assessment methods based on on-board vibration data have gradually attracted attention.

[0004] At present, the track condition assessment method based on vehicle-mounted vibration data still faces some technical difficulties in practical application: the track infrastructure is widely distributed, and parameters such as line profile, track composition and running speed vary significantly in space, forming a complex and diverse track environment. The heterogeneity of this environment makes it difficult for traditional disease identification methods to maintain stable detection results under different line conditions. In addition, the promotion and application of portable detection equipment makes it possible to obtain vibration data of the entire line, but due to the limitations of manpower and time costs, on-site verification and labeling work are still mainly concentrated in key sections. This unbalanced data acquisition method leads to significant distribution differences between vibration data and structural disease labels.

[0005] Current research mainly focuses on single heterogeneous factors. The difference in support stiffness of different types of roadbed structures will lead to frequency-dependent damping effects, which will affect the mapping relationship between vehicle body vibration and track state. In the study of curved sections, it was found that the wheel-rail contact interface will produce stronger self-excited tangential forces, forming unique vibration characteristics. Although these single-factor studies have revealed the basic laws of the vibration-disease relationship, it is difficult to fully reflect the impact of complex environments. Therefore, it is difficult to combine relevant parameters to evaluate the true state of track diseases. Summary of the invention

[0006] The present invention aims to provide a multi-level track defect identification system based on vehicle body vibration data, which solves the problem that the existing track condition assessment method cannot assess the true condition of track defects in a complex environment.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows: a multi-level track defect identification system based on vehicle body vibration data, comprising: Data acquisition module, used to collect heterogeneous data of the track environment; The environment adaptation module is used to analyze the correlation between vehicle vibration data and heterogeneous data, construct a dual-view similarity matrix with differentiated weights, and realize adaptive classification of the track environment through spectral clustering method; The label enhancement module calculates the multi-dimensional uncertainty index of the sample through the self-training method of the spectral normalized Gaussian process model, and gradually generates high-quality pseudo-labels through a dynamic threshold mechanism, ultimately forming a complete annotated sample set; Disease recognition module, using attention-guided multi-task cascade The architecture is to detect the disease location through the binary classification layer, filter the detection results of the binary classification layer with the help of the mask layer, evaluate the disease level through the multi-classification layer, and finally output a multi-level recognition result including the disease location and severity.

[0008] Furthermore, the basic rules for dataset construction are as follows: ; in, This is a vehicle body vibration data set. Each sampling point contains 6 vibration parameters. , specifically: represents the longitudinal acceleration, represents the lateral acceleration, is the vertical acceleration, represents the longitudinal angular velocity, represents the lateral angular velocity, Represents the vertical angular velocity; B is the heterogeneous factor data set, each sampling point contains Environmental parameters ; In order to eliminate the dimension differences between different parameters, the [-1, 1] interval normalization process is adopted; For non-numerical ballast type parameters, numerical mapping is performed according to the following rules: the integral ballast is assigned a value of -1, the rubber vibration isolation pad ballast is assigned a value of -0.33, the double-layer nonlinear fastener ballast is assigned a value of 0.33, and the steel spring floating plate ballast is assigned a value of 1; is the disease label set, and the label information of each sampling point is It indicates that, Corresponding to Different types of track disease label values; In order to obtain statistically representative training samples, the method of constructing grid samples is as follows: ; ; in, Indicates floor operation. Representative grid samples, is the total number of grids, which is determined by the total number of sampling points The number of sampling points of a single grid Sure: , and Respectively represent In the grid The vibration characteristics and heterogeneity factor information of each sampling point.

[0009] Furthermore, the process of the spectral clustering algorithm is as follows: the correlation strength between the vehicle body vibration and the data heterogeneity data is analyzed to determine the feature weights, and based on this, an improved dual-view similarity matrix is ​​designed to make full use of the differential information of the weights, and then spectral clustering is performed to convert the line grid sample data set into Divided into different categories.

[0010] Furthermore, the dual-view similarity matrix is ​​constructed as follows: the feature space is divided into high-weight viewpoints based on the median of the feature weights and low-weight perspective , high-weight perspectives use larger connection parameters and structural protection parameters , to allow more samples to establish connections and protect local structural characteristics; low-weight perspectives use smaller connection parameters and structural protection parameters , limiting the connection between samples and reducing the dependence on local structure, characterizing the local structure association through local density and reconstruction relationship within each perspective: Local density through the sample The overlap of neighbor sets is measured: ; in, Representation sample of Number of neighbors , each perspective quilt adjust: ,in Take 10; The reconstruction relationship is calculated by normalizing the distance between samples: ; Gaussian kernel function in the design perspective based on local indicators: ; in, is the feature weight vector, is the local adaptive bandwidth, taking samples The average distance of the nearest neighbors, It is a local structure protection term, which consists of local density and reconstruction relationship.

[0011] Furthermore, the similarity matrices are fused to balance the contribution of the two perspectives to the final similarity: ; in, and Represent the similarity matrix of two perspectives respectively, is the fusion coefficient, which is used to control the dominance of the high-weight perspective; based on the fused similarity matrix Creation Matrix : ； Among them, all non-diagonal elements are 0, and the Laplace matrix is ​​calculated from the similarity matrix and the degree matrix , using the symmetric normalization method: ; in, The symmetric normalization method can maintain the geometric structure of high-dimensional data while having the advantages of high computational stability and efficient processing of data of different scales. Before The eigenvector corresponding to the smallest eigenvalue , and use it as a column vector to construct the eigenvector matrix ,Will Each row of is normalized to unit length to obtain the matrix : ; Will Each row in the dimensional points are clustered, and the samples The category is the cluster to which it finally belongs.

[0012] Furthermore, the pseudo-label generation method includes a spectral normalized Gaussian process model and an adaptive threshold training mechanism: Spectrally normalized Gaussian process model: The Gaussian kernel function is used to measure the similarity between two points in the input space, which decays exponentially as the distance between the points increases: ; in is the length scale parameter, and its initial value is set to 0.1; RFF is mapped by the following function To approximate the kernel function: ; in, From the normal distribution The sampled random projection matrix, is a random bias vector sampled from a uniform distribution , is the feature dimension (512); through this mapping, any two input and The inner product of satisfies: ; The predicted value is obtained by linear combination of the mapped features, and the uncertainty can be calculated by the kernel matrix in the feature space; Multi-dimensional uncertainty assessment and adaptive threshold training mechanism: Design of multidimensional uncertainty assessment value based on prediction value and uncertainty value output by spectral normalized Gaussian process model , considering four metrics: is entropy, which represents the uncertainty of distribution; is the marginal difference; is the confidence level, taking the highest prediction probability; To predict the variance, characterize the model stability; ; ; ; ; ; in, denotes the diagonal elements of the kernel matrix, Indicates that the sample belongs to The predicted probability of each category, represents the highest predicted probability of the sample, Represents the second highest predicted probability of the sample; , , and are the weights of entropy, marginal difference, confidence and prediction variance respectively; By determining the multidimensional uncertainty assessment value Whether it exceeds the dynamic threshold to select the training set for the next iteration, the specific method is: set the dynamic low threshold , Lower than Indicates that the sample is predicted to be a known type of sample with a high degree of certainty, that is, the certainty of the real disease is high, then the sample prediction value is set as a pseudo label; set a dynamic high threshold , Higher than It indicates that the sample may have characteristics different from those of known labeled samples and is more likely to belong to a new category, that is, if there is no disease in reality, the sample prediction value is set to the new category '0'.

[0013] Furthermore, the initial threshold is set as follows: ; ; in, and are the mean and standard deviation of the model uncertainty estimate; 3 The interval setting provides a statistically reliable boundary. Considering that the model's prediction ability gradually improves with training iterations, the threshold needs to be adjusted accordingly to adapt to changes in model performance: ; ; in, , is the threshold of the previous iteration, , ; is the average confidence of the newly labeled samples, and 0.95 is the benchmark confidence; after reaching the predetermined number of iterations, the mature model is used to predict the categories of the remaining samples, and the samples of the new category in the remaining samples are screened according to the last threshold.

[0014] Furthermore, the binary classification layer adopts a weighted binary cross entropy loss function by introducing a penalty factor Increase the penalty for false negative samples: ; in, is the true label of each sample, that is =1 means there is disease, =0 means there is no disease. is the predicted value, between 0 and 1; is the penalty factor value, when =1, the penalty term will be activated. The closer the predicted value is to 0, the larger the loss function value will be.

[0015] Furthermore, the mask layer includes a filtering layer and a multi-classification input layer. The filtering layer applies a non-zero mask to the output of the binary classification layer, retains the positive predictions in the output, that is, the prediction value>0.5, and sets all other prediction values ​​to zero; the filtered binary classification layer output is element-wise multiplied with the original input to filter out samples that the binary classification layer believes to have diseases; the filtered samples form the multi-classification input layer for training the multi-classification layer to evaluate the disease level.

[0016] Furthermore, the channel attention mechanism is used to weight the output features of the convolutional block. The first-stage channel attention weights are nonlinearly transformed through feature remapping, and then go through the dimension adjustment and Activation function; through The function generates normalized guidance weights, which are element-wise multiplied with the channel attention outputs of the multi-classification stage; the weighted features are transformed through a multi-layer fully connected network, and each layer uses Prevent overfitting; finally, according to the number of levels of different diseases, The function outputs the multi-classification prediction results. This stage uses the standard cross entropy loss function, which is weighted and added to the loss function of the binary classification layer to form the overall optimization goal of the model: ; in, is the binary classification loss, is the multi-classification loss, and is the corresponding weight coefficient.

[0017] Compared with the prior art, this solution has the following beneficial effects: 1. This paper proposes a correlation-driven dual-view spectral clustering algorithm to achieve adaptive classification of the track environment. Compared with the traditional method of only analyzing a single heterogeneous factor, this algorithm analyzes the correlation between vibration and multiple heterogeneous factors, and combines the dual-view design with differentiated weights to significantly improve the classification effect in complex track environments. The silhouette coefficient is significantly improved compared with existing methods.

[0018] 2. The present invention proposes a label enhancement method based on the adaptive threshold self-training method of spectral normalized Gaussian process, which solves the problem of sparse labels of track structure defects. This method achieves high-quality labeling of unlabeled samples by introducing spectral normalized Gaussian process and designing a dynamic threshold strategy, and all evaluation indicators are significantly better than existing methods.

[0019] 3. This invention proposes a multi-task cascade based on attention guidance The model has realized multi-level parallel identification of track geometry and structural defects for the first time. Through the cascaded architecture and cross-level attention guidance mechanism, the model ensures the accuracy of defect detection while achieving accurate assessment of severity. It has achieved good performance in both geometric and structural defect identification, providing more comprehensive technical support for track condition assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of a multi-level track defect identification system based on vehicle body vibration data of the present invention; Figure 2 It is a method for generating pseudo labels of structural defects based on an adaptive threshold self-training method of a spectral normalized Gaussian process in this embodiment; Figure 3 This is the multi-task cascade based on attention guidance in this embodiment Used for multi-level track damage classification maps; Figure 4 is the feature-vibration correlation heat map in this embodiment; Figure 5 is a silhouette coefficient analysis diagram for determining the optimal number of clusters in this embodiment; Figure 6 is the clustering result of grid samples and sampling points in this embodiment Visualization diagram; Figure 7 is a performance comparison chart of K-means clustering, spectral clustering and density-based clustering methods in this embodiment; Figure 8 is the pseudo-label generation evolution diagram in cluster 0 and cluster 1 in this embodiment; Fig. 9 The training accuracy and loss function evolution curves of the same type of track defects in this embodiment; Fig.10 Evolution of attention weights in 100 rounds of training in this example. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below through specific embodiments: Example 1

[0022] like Figures 1 to 4 As shown in the figure, the multi-level track defect identification system based on vehicle body vibration data includes: Data acquisition module, the data acquisition module is used to collect heterogeneous data of the rail environment; in this embodiment, the data acquisition module uses a portable detector independently developed by the State Key Laboratory of Rail Traffic Control and Safety of Beijing Jiaotong University, and the size of the detector host is 270×170×91mm. Figure 1As shown in (ab), the body vibration data is collected by a portable detector on an ordinary passenger subway train. The body vibration data includes the longitudinal, lateral and vertical acceleration and angular velocity of the body, as well as the speed and mileage position information. The heterogeneous data of the track environment includes static and dynamic data. The static heterogeneous data is extracted from the line foundation design drawing, including the roadbed type, line curvature, and slope; the dynamic heterogeneous data is the train running speed.

[0023] The basic rules for dataset construction are as follows: (1); in, This is a vehicle body vibration data set. Each sampling point contains 6 vibration parameters. , specifically: represents the longitudinal acceleration, represents the lateral acceleration, is the vertical acceleration, represents the longitudinal angular velocity, represents the lateral angular velocity, Represents the vertical angular velocity. B is a heterogeneous factor data set, each sampling point contains Environmental parameters In order to eliminate the dimension differences between different parameters, the [-1, 1] interval normalization process is adopted; for non-numerical ballast type parameters, numerical mapping is performed according to the following rules: the integral ballast is assigned a value of -1, the rubber vibration isolation pad ballast is assigned a value of -0.33, the double-layer nonlinear fastener ballast is assigned a value of 0.33, and the steel spring floating plate ballast is assigned a value of 1. is the disease label set, and the label information of each sampling point is It indicates that, Corresponding to Different types of track disease label values; In order to obtain statistically representative training samples, the method of constructing grid samples is as follows: (2); (3); in, Indicates floor operation. Representative grid samples, is the total number of grids, which is determined by the total number of sampling points The number of sampling points of a single grid Sure: , and Respectively represent In the grid Through the above technical solution, the present invention realizes the effective fusion of vibration data, heterogeneous factors and disease labels, laying a data foundation for subsequent disease identification.

[0024] Environmental Adaptation Module,The environmental adaptation module is used to analyze the correlation between vehicle vibration data and heterogeneous data,,construct a dual-view similarity matrix with differentiated weights, and realize the adaptive classification of the track environment,through the spectral clustering method.

[0025] The process of the spectral clustering algorithm is as follows: first, the correlation strength between vehicle body vibration and heterogeneous factors is analyzed to determine the feature weights, and based on this, an improved dual-view similarity matrix is ​​designed to make full use of the differential information of the weights, and then spectral clustering is performed to convert the line grid sample data set into The feature weights are divided into different categories. The feature weight calculation process is as follows: Considering the possible nonlinear relationship between the heterogeneous factors and the vehicle body vibration, the Spearman rank correlation coefficient and the Pearson correlation coefficient are used for calculation respectively.

[0026] The Spearman rank correlation coefficient calculation formula is: (4); in, Indicates Heterogeneous factors, Indicates The vehicle body vibrates; and Respectively express , Sort by Indicates that the data has a sample.

[0027] Pearson correlation coefficient calculation formula: (5); in, ,Right now Column mean; ,Right now Column mean; Indicates that the data has The larger value of formula (4) and (5) is taken as the first The heterogeneity factor and The correlation between the vibrations of the vehicle body.

[0028] (6); The initial weight value calculation is shown in the following equations (7) and (8). These two weights will be set as the basic weight values ​​for calculating the weight values ​​of each heterogeneous factor according to the designed feature weight algorithm. Since there are 6 vehicle body vibration characteristics, The value is 6.

[0029] (7); (8); The input of the algorithm includes heterogeneity data, vehicle body vibration data, and the correlation matrix between features and vehicle body vibration. , and predefined initial weight values and The output is a list of heterogeneity factor weights in descending order of weight. The core idea of ​​the algorithm is to According to the strength of correlation, they are divided into two groups, and The specific steps are as follows: First, the correlation value Sort in descending order and group by n into related groups , and the rest are classified into the correlation group Then the weight of each feature is accumulated. In the weight calculation stage, all , if in In the middle, add ,exist In the middle, add Finally, the accumulated weights of each heterogeneous factor are sorted to obtain the final feature weight list .

[0030] The label enhancement module calculates the multidimensional uncertainty index of the sample through the self-training method of spectral normalized Gaussian process, and gradually generates high-quality pseudo-labels through a dynamic threshold mechanism, finally forming a complete annotated sample set.

[0031] The method for constructing the dual-view similarity matrix is ​​as follows: the feature space is divided into high-weight viewpoints based on the median of the feature weights. and low-weight perspective For these two perspectives, different connection parameters are designed and structural protection parameters : High-weight perspectives use larger connection parameters and structural protection parameters , to allow more samples to establish connections and protect local structural characteristics; low-weight perspectives use smaller connection parameters and structural protection parameters , limiting the connection between samples and reducing the reliance on local structure. This differentiated parameter design enables the two perspectives to characterize the intrinsic structure of the data from different levels. Within each perspective, the local structure association is characterized by local density and reconstruction relationship: Local density through the sample The overlap of neighbor sets is measured: (9); in, Representation sample of Number of neighbors , each perspective quilt adjust: ,in Take 10; The reconstruction relationship is calculated by normalizing the distance between samples: (10); The Gaussian kernel function within the perspective is designed based on the local indicator, as follows: (11); in, is the feature weight vector, is the local adaptive bandwidth, taking samples The average distance of the nearest neighbors, It is a local structure protection term, which consists of local density and reconstruction relationship.

[0032] In order to balance the contribution of the two perspectives to the final similarity, the similarity matrix is ​​fused as follows: (12); in, and Represent the similarity matrix of two perspectives respectively, is the fusion coefficient, which is used to control the dominance of the high-weight perspective, so that it can maintain the dominant role of the high-weight feature while retaining the auxiliary information provided by the low-weight feature. Creation Matrix : (13); Among them, all non-diagonal elements are 0, and the Laplace matrix is ​​calculated from the similarity matrix and the degree matrix , using the symmetric normalization method: (14); in, The symmetric normalization method can maintain the geometric structure of high-dimensional data while having the advantages of high computational stability and efficient processing of data of different scales. Before Eigenvectors corresponding to (expected number of clusters) minimum eigenvalues , and use it as a column vector to construct the eigenvector matrix ,Will Each row of is normalized to unit length to obtain the matrix : (15); Will Each row in the dimensional points are clustered, and the samples The category is the cluster to which it finally belongs. The above spectral clustering process classifies all line grid units, and the sampling points in the grid units are also divided into different categories, laying the foundation for the subsequent disease identification tasks of different categories.

[0033] The pseudo-label generation method includes a spectral normalized Gaussian process model and an adaptive threshold training mechanism: like Figure 2 As shown in Figure 1, the spectral normalized Gaussian process model (SNGP model) contains two residual blocks, each of which is configured with a spectral normalization layer. SN controls the Lipschitz continuity condition constant of the network by limiting the spectral norm of the weight matrix. The calculation formula is as follows: (16); in, is the weight matrix, is its maximum singular value.

[0034] The network parameters are shown in Table 1. The output vector is connected to the random Fourier feature module. This module approximates the Gaussian process kernel function through random feature mapping to determine the similarity between unlabeled samples and labeled samples. Specifically, the Gaussian kernel function is first used to measure the similarity between two points in the input space, which decays exponentially as the distance between the points increases: (17); in l is the length scale parameter, and its initial value is set to 0.1. RFF is mapped by the following function To approximate the kernel function: (18); in, From the normal distribution The sampled random projection matrix, is a random bias vector sampled from a uniform distribution , is the feature dimension (512); through this mapping, any two input and The inner product of satisfies: (19); The predicted value is obtained by linear combination of the mapped features, and the uncertainty can be calculated by the kernel matrix in the feature space. For the evaluation of the SNGP output value in each iteration, this embodiment designs a multi-dimensional uncertainty evaluation strategy, such as Figure 2 As shown in Figure 2, a multidimensional uncertainty evaluation value is designed based on the predicted value and uncertainty value output by SNGP. , considering four metrics: is entropy, which represents the uncertainty of distribution; is the marginal difference; is the confidence level, taking the highest prediction probability; is the prediction variance, which characterizes the stability of the model, as follows: (20); (twenty one); (twenty two); (twenty three); (twenty four); in, denotes the diagonal elements of the kernel matrix, Indicates that the sample belongs to The predicted probability of each category, represents the highest predicted probability of the sample, Represents the second highest predicted probability of the sample; , , and are the weights of entropy, marginal difference, confidence and prediction variance respectively.

[0035] Then, a threshold adaptive mechanism is designed to determine the multi-dimensional uncertainty evaluation value. Whether it exceeds the dynamic threshold to select the training set for the next iteration, the specific method is: set the dynamic low threshold , Lower than It means that the sample is predicted to be a known type of sample with a high certainty, that is, the certainty of the real disease is high, and the sample prediction value is set as a pseudo label. Set dynamic high threshold , Higher than Indicates that the sample may have characteristics different from known labeled samples and is more likely to belong to a new category, that is, if there is no disease, the sample prediction value is set to the new category '0'. Add the above sample to the training set for the next training. First, the initial threshold is set as follows based on statistical principles: (25); (26); in, and are the mean and standard deviation of the model uncertainty estimate; 3 The interval setting provides a statistically reliable boundary that can effectively identify high-confidence samples and potential new category samples. Considering that the model's prediction ability gradually improves with training iterations, the threshold needs to be adjusted accordingly to adapt to changes in model performance: (27); (28); in, , is the threshold of the previous iteration, , ; is the average confidence of the newly labeled samples, and 0.95 is the benchmark confidence; after reaching the predetermined number of iterations, the mature model is used to predict the categories of the remaining samples, and the samples of the new category in the remaining samples are screened according to the last threshold.

[0036] Table 1. Model architecture parameter configuration of adaptive threshold self-training method based on spectral normalized Gaussian process

[0037] The disease recognition module adopts an attention-guided multi-task cascade CNN architecture, that is, the disease location is detected through a binary classification layer, the detection results are filtered with the help of a mask layer, and the disease level is evaluated by a multi-classification layer, and finally a multi-level recognition result including the disease location and severity is output.

[0038] The binary classification layer is used to identify the location of the defect. The specific structure is as follows: a channel attention mechanism is introduced after the original input signal, and the importance weights of different vehicle posture features are captured through parallel global average pooling and maximum pooling branches. The features of the two branches are reduced in dimension and After the nonlinear transformation of the activation function, the attention weights are generated by fusion and multiplied with the original input to obtain weighted features. The weighted features are extracted through a two-layer convolutional network, and each convolution layer is followed by a maximum pooling operation to compress the feature dimension. After flattening, the extracted features are input into a three-layer fully connected network for feature conversion, all using Activation function, and adopt Prevent overfitting. Finally, The output node of the activation function produces a binary classification result. Since misdiagnosing a diseased sample as disease-free (false negative) is more dangerous than misdiagnosing a disease-free sample as diseased (false positive). The binary classification layer uses a weighted binary cross entropy loss function, which introduces a penalty factor Increase the penalty for false negative samples: (29); in, is the true label of each sample, that is =1 means there is disease, =0 means there is no disease. is the predicted value, between 0 and 1; is the penalty factor value, when =1, the penalty term will be activated. The closer the predicted value is to 0, the larger the loss function value will be.

[0039] The mask layer is used to filter the results of the binary classification layer. The mask layer includes a screening layer and a multi-classification input layer. The screening layer applies a non-zero mask to the output of the binary classification layer, retains the positive predictions in the output, that is, the prediction value>0.5, and sets all other prediction values ​​to zero; the filtered binary classification layer output is element-wise multiplied with the original input to filter out samples that the binary classification layer believes to have diseases; the screened samples form the multi-classification input layer for training the multi-classification layer to assess disease levels.

[0040] The multi-classification layer is used to evaluate the disease grade. The specific method is as follows: the channel attention mechanism is used to weight the output features of the convolution block, and the first-stage channel attention weights are nonlinearly transformed through feature remapping, and then the dimension adjustment and Activation function; through The function generates normalized guidance weights, which are element-wise multiplied with the channel attention outputs of the multi-classification stage; the weighted features are transformed through a multi-layer fully connected network, and each layer uses Prevent overfitting; finally, according to the number of levels of different diseases, The function outputs the multi-classification prediction results. This stage uses the standard cross entropy loss function, which is weighted and added to the loss function of the binary classification layer to form the overall optimization goal of the model: (30); in, is the binary classification loss, is the multi-classification loss, and is the corresponding weight coefficient.

[0041] Case Study: Metro Line 10 in a certain city was selected as the specific test and verification object. Based on the above-mentioned defect identification system, measured data collection and defect identification experiments were carried out on the line. A total of 82,400 complete test samples were obtained through the portable detector. Each set of samples contains vehicle body vibration data, heterogeneous data and corresponding defect marking information. The data collection covers a range of 20.6095 kilometers of the entire line, including various typical working conditions such as different slopes and different roadbed structures. The collected defect samples include two major categories: track geometric diseases and structural diseases. Among them, geometric diseases cover five main types: height, level, track direction, track gauge and triangular pits, and structural diseases select two typical disease forms: rail side wear and vertical wear.

[0042] Verification of the effectiveness of the dual-view spectrum clustering algorithm: In the data preprocessing stage, the data set obtained by associating the vehicle body vibration data with the heterogeneous data is first standardized. In order to eliminate the impact of different dimensions of the heterogeneous data and retain relatively complete information, all data are normalized in the interval [-1, 1]. For non-numerical roadbed type parameters, special processing is required: the integral roadbed, rubber vibration isolation pad roadbed, double-layer nonlinear fastener roadbed and steel spring floating plate roadbed are assigned values ​​of 1, 2, 3, and 4 respectively. After normalization, their values ​​are -1, -0.33, 0.33, and 1 respectively. This mapping method maintains the relative relationship between different roadbed types.

[0043] In terms of grid division and feature extraction, this embodiment divides the line into grid units of 200 meters in length, and obtains a total of 103 grid samples, each of which contains 800 sampling points (200 / 0.25=800). The mean of the vehicle body vibration and heterogeneity data in each grid is calculated according to formulas (1)-(3) as the feature representation of the grid. Then, the correlation between each vibration feature and heterogeneity data is calculated using formulas (4)-(8), as shown in Figure 4 The heat map shown in the figure shows that the ordinate represents four heterogeneous factors, the abscissa represents six vehicle body vibration parameters, blue represents negative correlation, red represents positive correlation, and the color depth reflects the strength of the correlation. Based on the calculation results, the weight values ​​of each factor are obtained: speed 0.45, slope 0.25, track curvature 0.25, and roadbed 0.05.

[0044] In terms of clustering parameter settings, set larger connection parameters for high-weight perspectives. = 0.9 and structural protection parameters =0.8, allowing more samples to establish connections and protecting local structural characteristics; setting smaller parameters for low-weight perspectives =0.2 and =0.2, limiting the connection between samples and reducing the reliance on local structure; the fusion coefficient γ is set to 0.7 to balance the contribution of high- and low-weight perspectives. By comparing the silhouette coefficients of 2-10 clusters (such as Figure 5 As shown in the figure, when the number of clusters is 2, the highest silhouette coefficient of 0.46 is obtained, and the optimal number of clusters is determined to be 2.

[0045] In order to intuitively demonstrate the clustering effect, we use The technology maps high-dimensional data into two-dimensional space. Figure 6 As shown, the blue cluster 0 contains 74 grids (59,200 sampling points), and the green cluster 1 contains 29 grids (23,200 sampling points). The two types of data show a clear separation in the feature space, indicating that the clustering model successfully captures the data structure differences in the high-dimensional heterogeneous factor space.

[0046] In order to comprehensively evaluate the superiority of this scheme, a comparative experiment with traditional clustering algorithms was conducted. Figure 7 As shown in Table 2 below, the weighted spectral clustering method proposed in this embodiment achieved a silhouette coefficient of 0.4567, which is 35.7% higher than the unweighted method (0.3365). Weighted spectral clustering, unweighted spectral clustering and density-based spatial clustering all determined that the optimal number of clusters is 2, indicating that the spectral clustering method has good stability. Although K-means gives a more detailed four-category division, its lower silhouette coefficient (0.1774) suggests the risk of overfitting. In terms of sample distribution balance, this method reasonably divides 103 samples into two groups of 29 and 74, while the distribution of density-based spatial clustering (86:17) shows obvious imbalance. Through the visualization of clustering results, it is further verified that the clustering boundaries obtained by this method are the clearest and the inter-cluster separation is the highest, while the results of other methods, especially K-means clustering, show relatively fuzzy cluster boundaries. The experimental results fully demonstrate the effectiveness and superiority of the spectral clustering algorithm in dealing with the heterogeneity of complex orbital environments.

[0047] Table 2 Comparison of clustering method performance and sample distribution

[0048] Verification of the effectiveness of the adaptive threshold self-training method based on spectral normalized Gaussian process for structural disease samples: The random Fourier feature module sets the feature dimension to 512, and the initial value of the length scale parameter is 0.1. In the multidimensional uncertainty assessment mechanism, the entropy weight and marginal difference weight are both set to 0.3, and the confidence weight and prediction variance weight are set to 0.2. As shown in Table 4 below, the label distribution of the two cluster samples is as follows (as shown in Table 3 below): labels 0, 1, 2, and 3 represent samples without disease, disease level 1, 2, and 3, respectively. There are a large number of unlabeled samples in each sample set, and pseudo labels need to be labeled through self-training. Since the number of label 0 samples is 0, label 0 will be identified as a new category during model training.

[0049] Table 3 Initial distribution of track structure disease labels in cluster samples

[0050] Experimental data show that: As shown in Table 4, the mean variance of the model in all scenarios is less than 0.19, and the standard deviation is within 0.04. The average confidence value is 0.95. In terms of the length scale interval, the vertical wear interval of cluster 0 is [0.20024903, 0.21437594], and the side wear interval of cluster 1 is [0.1, 0.1457911].

[0051] Table 4 Performance indicators of adaptive threshold self-training method based on spectral normalized Gaussian process

[0052] The following data are obtained through tracking and analyzing the training process: Figure 8 As shown in the figure, in the side wear training of cluster 0, the average confidence of the model increased from 0.8 to 0.95. In the first iteration, 13,486 samples were annotated, of which 1,759 (13.04%) were identified as new categories. In the fourth iteration, the recognition rate of new categories reached 99.32%. For vertical wear training, the confidence increased from 0.5 to 0.9, and the recognition rate of new category samples in each iteration was in the range of 11-15%. In cluster 1, 5,490 samples were annotated in the first iteration of side wear, and the recognition rate of new categories was 16.67%. In the third iteration, the recognition rate of new categories increased to 34.48%.

[0053] The final label distribution results are shown in Table 5 below, where cluster 0 side wear and vertical wear contain 39,948 and 1,165 label 0 samples, respectively, and cluster 1 side wear and vertical wear contain 5,198 and 472 label 0 samples, respectively. In order to verify the effectiveness of the adaptive threshold self-training method strategy based on spectral normalized Gaussian process in pseudo-labeling, as shown in Table 6, the baseline model (convolutional neural network + self-training model), convolutional neural network + Monte Carlo sampling + self-training model, convolutional neural network + deterministic uncertainty estimation + self-training model and convolutional neural network + spectral normalized Gaussian process + self-training model were selected for comparison. The adaptive threshold self-training method based on spectral normalized Gaussian process has better performance indicators than other methods in terms of average confidence, computational efficiency, convergence performance and confidence variance.

[0054] Table 5 The final distribution of track structure disease labels after processing by the adaptive threshold self-training method based on spectral normalized Gaussian process

[0055] Table 6 Performance comparison of different models on pseudo-label tasks

[0056] Experimental data show that the adaptive threshold self-training method based on spectral normalized Gaussian process can handle the problem of sparse rail structure disease labels. Through the dynamic threshold adjustment mechanism and multi-dimensional uncertainty assessment strategy, this method can label data for scenes of different complexity. The data set processed by the adaptive threshold self-training method based on spectral normalized Gaussian process can provide training samples for subsequent multi-level rail disease identification tasks.

[0057] Verification of the effectiveness of multi-task cascaded CNN guided by multi-level disease attention: The annotated multi-level track defect data is input into the attention-guided multi-task cascaded CNN for training. Figure 4 As shown in the figure, the model contains seven branches, corresponding to the identification of height, level, gauge, track direction, triangular pit, side wear and vertical wear. The penalty factor value of the binary loss function is set to 2. The weights of the binary loss and multi-classification loss in the total loss are set to 2 and 1 respectively to ensure that the model gives priority to accurately detecting the location of the disease.

[0058] Parameter search is performed on the two types of samples to determine the optimal hyperparameter combination. In cluster 0, the optimal hyperparameter combination is a learning rate of 0.001 and a regularization term of 0.01; in cluster 1, it is a learning rate of 0.01 and a regularization term of 0.1. The hyperparameter selection for cluster 0 (59,200 samples) and cluster 1 (23,200 samples) is reasonable. Cluster 0 has significantly more samples, and a smaller learning rate and a smaller regularization term enable the model to learn more carefully. In contrast, cluster 1 has fewer samples, and a larger learning rate can speed up the convergence of the model on less data, while a larger regularization term enables the model to better learn the true characteristics of the data.

[0059] The number of training rounds is set to 100. Fig. 9 As shown in the figure, the output loss function and accuracy change with the training rounds, showing the recognition effect of different types of diseases. Fig. 9 (ad) shows the training results of cluster 0, including the unevenness in geometric defects ( Fig. 9 -a), triangular pit ( Fig. 9 -b), and vertical wear in structural damage ( Fig. 9 -c) and side wear ( Fig. 9 -d). Fig. 9 (eh) shows the training results of cluster 1, including the track gauge ( Fig. 9 -e), track direction ( Fig. 9 -f), uneven level ( Fig. 9 -g) and vertical wear in structural diseases ( Fig. 9 -h). Each subgraph is divided into two parts, the upper part shows the accuracy evolution curve, and the lower part shows the loss function evolution curve.

[0060] From the training results, as shown in Table 7 below, the model has good performance in both geometric and structural track defect classification. It performs well in the binary classification task: the precision, recall and F1 score of all defect types are close to or reach 1.00. From the Kappa coefficient, the Kappa values ​​of geometric defects such as height, level, gauge and track direction are all above 0.9, while the Kappa values ​​of structural defects such as side wear and vertical wear are relatively low, especially the Kappa value of vertical wear in cluster 0 is 0.8010.

[0061] Table 7 Performance indicators of the attention-guided multi-task cascade CNN model in the recognition of various track defects

[0062] The recognition performance in multi-classification tasks remains at a high level. The F1 scores of geometric damage types are generally above 0.95, while the F1 score of structural damage such as side wear in cluster 1 is 0.93. Overall, the model shows stable performance in both datasets, especially in geometric damage recognition. Although the recognition of structural damage is relatively challenging, its performance indicators remain at an acceptably high level.

[0063] To further understand the learning mechanism of the model, this example analyzes the evolution characteristics of attention weights in the recognition of high and low roughness and side wear. Fig.10 As shown, the model demonstrates an accurate grasp of the laws of physics. Fig.10 -a shows that in the height roughness identification of cluster 1, the pitch angular velocity weight increases from 0.52 to 0.67 and remains stable. Fig.10 -b shows that in the side wear identification of cluster 1, the yaw angular velocity weight increases rapidly from 0.5 to 0.7, and the roll angular velocity weight increases steadily to 0.55. These changes in weight distribution reflect the model's ability to effectively identify the characteristics of different types of diseases.

[0064] The above are only embodiments of the present invention, and the common knowledge such as the known specific structures and / or characteristics in the scheme are not described in detail here. It should be pointed out that for those skilled in the art, several deformations and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A multi-level track defect identification system based on vehicle body vibration data, characterized in that: include: Data acquisition module, used to collect heterogeneous data of the track environment; The environment adaptation module is used to analyze the correlation between vehicle vibration data and heterogeneous data, construct a dual-view similarity matrix with differentiated weights, and realize adaptive classification of track environment through spectral clustering method; The label enhancement module calculates the multi-dimensional uncertainty index of the sample through the self-training method of the spectral normalized Gaussian process model, and gradually generates high-quality pseudo-labels through a dynamic threshold mechanism, ultimately forming a complete annotated sample set; Disease recognition module, using attention-guided multi-task cascade The architecture is to detect the disease location through the binary classification layer, filter the detection results of the binary classification layer with the help of the mask layer, evaluate the disease level through the multi-classification layer, and finally output a multi-level recognition result including the disease location and severity.

2. The multi-level track defect identification system based on vehicle body vibration data according to claim 1 is characterized in that: The basic rules for dataset construction are as follows: ； in, This is a vehicle body vibration data set. Each sampling point contains 6 vibration parameters. , specifically: represents the longitudinal acceleration, represents the lateral acceleration, is the vertical acceleration, represents the longitudinal angular velocity, represents the lateral angular velocity, Represents the vertical angular velocity; B is the heterogeneous factor data set, each sampling point contains Environmental parameters ; In order to eliminate the dimension differences between different parameters, the [-1, 1] interval normalization process is adopted; For non-numerical ballast type parameters, numerical mapping is performed according to the following rules: the integral ballast is assigned a value of -1, the rubber vibration isolation pad ballast is assigned a value of -0.33, the double-layer nonlinear fastener ballast is assigned a value of 0.33, and the steel spring floating plate ballast is assigned a value of 1; is the disease label set, and the label information of each sampling point is It indicates that, Corresponding to Different types of track disease label values; In order to obtain statistically representative training samples, the method of constructing grid samples is as follows: ； ； in, Indicates floor operation. Representative grid samples, is the total number of grids, which is determined by the total number of sampling points The number of sampling points of a single grid Sure: , and Respectively represent In the grid The vibration characteristics and heterogeneity factor information of each sampling point.

3. The multi-level track defect identification system based on vehicle body vibration data according to claim 1 is characterized in that: The process of the spectral clustering algorithm is as follows: the correlation strength between the vehicle body vibration and the data heterogeneity data is analyzed to determine the feature weights, and based on this, an improved dual-view similarity matrix is ​​designed to make full use of the differential information of the weights, and then spectral clustering is performed to convert the line grid sample data set into Divided into different categories.

4. The multi-level track defect identification system based on vehicle body vibration data according to claim 3 is characterized in that: The method of constructing the dual-view similarity matrix is ​​as follows: the feature space is divided into high-weight viewpoints based on the median of the feature weights and low-weight perspective , high-weight perspectives use larger connection parameters and structural protection parameters , to allow more samples to establish connections and protect local structural characteristics; low-weight perspectives use smaller connection parameters and structural protection parameters , limiting the connection between samples and reducing the dependence on local structure, characterizing the local structure association through local density and reconstruction relationship within each perspective: Local density through the sample The overlap of neighbor sets is measured: ； in, Representation sample of Number of neighbors , each perspective quilt adjust: ,in Take 10; The reconstruction relationship is calculated by normalizing the distance between samples: ； Gaussian kernel function in the design perspective based on local indicators: ； in, is the feature weight vector, is the local adaptive bandwidth, taking samples The average distance of the nearest neighbors, It is a local structure protection term, which consists of local density and reconstruction relationship.

5. The multi-level track defect identification system based on vehicle body vibration data according to claim 4 is characterized in that: The similarity matrices are fused to balance the contribution of the two perspectives to the final similarity: ； in, and Represent the similarity matrix of two perspectives respectively, is the fusion coefficient, which is used to control the dominance of the high-weight perspective; Based on the fused similarity matrix Creation Matrix : ； Among them, all non-diagonal elements are 0, and the Laplace matrix is ​​calculated from the similarity matrix and the degree matrix , using the symmetric normalization method: ； in, The symmetric normalization method can maintain the geometric structure of high-dimensional data while having the advantages of high computational stability and efficient processing of data of different scales. Before The eigenvector corresponding to the smallest eigenvalue , and use it as a column vector to construct the eigenvector matrix ,Will Each row of is normalized to unit length to obtain the matrix : ； Will Each row in the dimensional points to cluster, samples The category is the cluster to which it finally belongs.

6. The multi-level track defect identification system based on vehicle body vibration data according to claim 1 is characterized in that: The pseudo-label generation method includes a spectral normalized Gaussian process model and an adaptive threshold training mechanism: Spectrally normalized Gaussian process model: The Gaussian kernel function is used to measure the similarity between two points in the input space, which decays exponentially as the distance between the points increases: ； in is the length scale parameter, and its initial value is set to 0.1; RFF is mapped by the following function To approximate the kernel function: ； in, From the normal distribution The sampled random projection matrix, is a random bias vector sampled from a uniform distribution , is the feature dimension (512); through this mapping, any two input and The inner product of satisfies: ； The predicted value is obtained by linear combination of the mapped features, and the uncertainty can be calculated by the kernel matrix in the feature space; Multi-dimensional uncertainty assessment and adaptive threshold training mechanism: Design of multidimensional uncertainty assessment value based on prediction value and uncertainty value output by spectral normalized Gaussian process model , considering four metrics: is entropy, which represents the uncertainty of distribution; is the marginal difference; is the confidence level, taking the highest prediction probability; To predict the variance, characterize the model stability; ； ； ； ； ； in, denotes the diagonal elements of the kernel matrix, Indicates that the sample belongs to The predicted probability of each category, represents the highest predicted probability of the sample, Represents the second highest predicted probability of the sample; , , and are the weights of entropy, marginal difference, confidence and prediction variance respectively; By determining the multidimensional uncertainty assessment value Whether it exceeds the dynamic threshold to select the training set for the next iteration, the specific method is: set the dynamic low threshold , Lower than Indicates that the sample is predicted to be a known type of sample with a high degree of certainty, that is, the certainty of the real disease is high, then the sample prediction value is set as a pseudo label; set a dynamic high threshold , which is higher than that, indicates that the sample may have characteristics different from those of known labeled samples and is more likely to belong to a new category, that is, if there is no disease in reality, the sample prediction value is set to the new category '0'.

7. The multi-level track defect identification system based on vehicle body vibration data according to claim 1 is characterized in that: The initial threshold is set as follows: ； ； Among them, and are the mean and standard deviation of the model uncertainty estimate; 3 The interval setting provides a statistically reliable boundary. Considering that the model's prediction ability gradually improves with training iterations, the threshold needs to be adjusted accordingly to adapt to changes in model performance: ； ； in, , is the threshold of the previous iteration, , ; is the average confidence of the newly labeled samples, and 0.95 is the benchmark confidence; after reaching the predetermined number of iterations, the mature model is used to predict the categories of the remaining samples, and the samples of the new category in the remaining samples are screened according to the last threshold.

8. The multi-level track defect identification system based on vehicle body vibration data according to claim 1 is characterized in that: The binary classification layer uses a weighted binary cross entropy loss function, which introduces a penalty factor Increase the penalty for false negative samples: ； in, is the true label of each sample, that is =1 means there is disease, =0 means there is no disease. is the predicted value, between 0 and 1; is the penalty factor value, when =1, the penalty term will be activated. The closer the predicted value is to 0, the larger the loss function value will be.

9. The multi-level track defect identification system based on vehicle body vibration data according to claim 8 is characterized in that: The mask layer includes a filtering layer and a multi-classification input layer. The filtering layer applies a non-zero mask to the output of the binary classification layer, retains the positive predictions in the output, that is, the prediction value>0.5, and sets all other prediction values ​​to zero; the filtered binary classification layer output is element-wise multiplied with the original input to filter out samples that the binary classification layer believes to have diseases; the filtered samples form the multi-classification input layer for training the multi-classification layer to assess disease levels.

10. The multi-level track defect identification system based on vehicle body vibration data according to claim 9 is characterized in that: The specific method of multi-classification layer disease grade assessment is as follows: the channel attention mechanism is used to weight the output features of the convolution block, the first-stage channel attention weights are nonlinearly transformed through feature remapping, and then the dimension adjustment and Activation function; through The function generates normalized guidance weights, which are element-wise multiplied with the channel attention outputs of the multi-classification stage; the weighted features are transformed through a multi-layer fully connected network, and each layer uses Prevent overfitting; finally, according to the number of levels of different diseases, The function outputs the multi-classification prediction results. This stage uses the standard cross entropy loss function, which is weighted and added to the loss function of the binary classification layer to form the overall optimization goal of the model: ； in, is the binary classification loss, is the multi-classification loss, and is the corresponding weight coefficient.

Images