A Distributed Collaborative Training Method and System for Railway Anomaly Detection Models

Abstract

This invention relates to the field of collaborative training technology, specifically to a distributed collaborative training method and system for railway anomaly detection models. The method includes: acquiring multi-dimensional time-series monitoring data of railways and calculating Granger causality relationships between features of each dimension; using a community detection algorithm to split the multi-dimensional time-series monitoring data into two independent views, obtaining initial model A and initial model B; during collaborative training iterations, alternately using one initial model to filter pseudo-label samples to update the other initial model, adding a new number of pseudo-label samples in each iteration; at the end of training, calculating the prediction variance of initial model A and initial model B, determining the fusion weights, and weightedly fusing the prediction results to obtain the anomaly detection result. In other words, the solution of this invention can split multi-dimensional time-series monitoring data into independent views, remove noisy samples, and obtain a more reliable model to improve detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of collaborative training technology. More specifically, this invention relates to a collaborative training method and system for a distributed railway anomaly detection model. Background Technology

[0002] The railway system is a vital pillar of the economy, and its safe and stable operation is of paramount importance. With the continuous development of the Internet of Things (IoT) and sensor technologies, numerous monitoring devices have been deployed on key railway equipment such as high-speed trains, tracks, and power supply systems to collect data in real time, accumulating massive amounts of multi-dimensional time-series data. How to effectively utilize this data for intelligent anomaly detection, and promptly identify equipment malfunctions and potential safety hazards, has become a core technology for ensuring railway transportation safety. Currently, deep learning-based anomaly detection methods have achieved certain application results in many fields; however, the performance of these methods highly depends on large-scale, high-quality labeled datasets, which are extremely costly to obtain in real-world industrial scenarios such as railways.

[0003] To address the shortcomings of the aforementioned deep learning methods and fully utilize massive amounts of unlabeled data, existing technologies have proposed applying semi-supervised learning methods such as co-training to anomaly detection tasks. The basic idea of ​​co-training is to construct two or more models trained on different views. By leveraging the complementarity between these models, they mutually provide high-confidence pseudo-labels for each other's unlabeled data. In this way, the models can learn together during the iterative process, thereby collectively improving overall anomaly detection performance even with only a small amount of labeled data.

[0004] However, the aforementioned semi-supervised learning methods based on collaborative training still have many shortcomings when applied to railway anomaly detection. First, there are usually complex nonlinear couplings and causal relationships between the features of multidimensional railway monitoring data, while existing technologies typically use simple random feature partitioning or partitioning methods based on single correlations. These methods are difficult to generate mutually independent views that meet the conditions for collaborative training. Second, existing technologies usually use fixed confidence thresholds when screening for pseudo-labels. This fixed screening method easily introduces noisy labels into the training process, causing errors to accumulate and amplify during the iterative training of the model. Finally, when making decisions and fusing multiple models, simple averaging or voting strategies ignore the reliability differences of different models on different data samples, which may lead to poor final detection accuracy. Summary of the Invention

[0005] The purpose of this invention is to propose a distributed collaborative training method and system for railway anomaly detection models, in order to solve the problems in the prior art that it is impossible to generate mutually independent views, introduces noisy samples during collaborative training, and ignores the reliability differences of different models during decision fusion, resulting in low detection accuracy; to this end, this invention provides solutions in the following two aspects.

[0006] In a first aspect, the present invention provides a distributed collaborative training method for railway anomaly detection models, comprising:

[0007] Acquire multidimensional time-series monitoring data of railways, calculate Granger causality relationships between features of each dimension in the multidimensional time-series monitoring data to construct a feature causality matrix; use a community detection algorithm to partition the feature causality matrix to obtain mutually independent views. Figure 1 Feature subsets and views Figure 2 Feature subsets are used to obtain modularity scores for community partitioning, and multidimensional time-series monitoring data is split into visual subsets. Figure 1 Heshi Figure 2 Using the original labeled dataset, respectively in the view Figure 1 Heshi Figure 2 The training process yields initial models A and B. During the co-training iterations, the following operations are performed alternately: One initial model is used to predict the unlabeled dataset. Samples with prediction confidence higher than a first preset threshold and local outlier factors calculated based on the k nearest neighbors in the unlabeled dataset lower than a second preset threshold are selected as pseudo-label samples. These pseudo-label samples are then added to the training set of the other initial model for updates. The number of newly added pseudo-label samples in each iteration is inversely proportional to the consistency rate of the prediction results of initial models A and B on the validation set at the end of the previous iteration. When training terminates, the prediction variances of initial models A and B on the nearest neighbors in the original labeled dataset are calculated based on the data to be detected. The fusion weights are then used to weight and fuse the prediction results of initial models A and B to obtain the anomaly detection results.

[0008] Preferably, the step of partitioning the feature causal relationship matrix using a community detection algorithm includes: employing the Louvain community detection algorithm, which continuously partitions nodes into communities that maximize the modularity gain of the feature causal relationship matrix, and iteratively aggregates these communities into new nodes for partitioning, until the modularity score no longer increases, resulting in mutually independent views. Figure 1 Feature subsets and views Figure 2 Feature subset.

[0009] Preferably, the prediction confidence is determined as follows: for any sample in the unlabeled dataset, the maximum class probability output by the initial model used for prediction is taken as the prediction confidence.

[0010] Preferably, the local outlier factor is calculated by setting a nearest neighbor parameter. The value is 20; for any sample in the unlabeled dataset, find the 20 nearest neighbor samples of the sample, calculate the reachability distance of each nearest neighbor and obtain the local reachability density, and calculate the local outlier factor of the sample.

[0011] Preferably, the number of newly added pseudo-label samples in each iteration is inversely proportional to the consistency rate of the prediction results of the initial model A and the initial model B on the validation set at the end of the previous iteration, including: calculating the initial model Compared with the initial model The consistency rate is calculated by predicting the proportion of samples with identical labels on the validation set. The number of newly added pseudo-labeled samples in this round is set to... ,in, This represents the number of newly added pseudo-labeled samples in this round. The value is the consistency rate, 100 is the baseline sample size, and 0.01 is a smoothing term to prevent the denominator from being zero.

[0012] Preferably, the step of calculating the prediction variances of initial model A and initial model B on the nearest neighbor samples in the original labeled dataset based on the data to be detected, and determining their respective fusion weights according to the reciprocal of the prediction variances, includes: using Euclidean distance as a metric, finding 15 nearest neighbor samples for the data to be detected in the original labeled dataset; calculating the prediction variances of the initial models respectively. and the initial model The variance of the predicted probability values ​​for 15 nearest neighbor samples; initial model The fusion weights satisfy: Initial model The fusion weights satisfy: In the formula, , The initial model , The fusion weight, , The initial model , The variance of the predicted probability values ​​for 15 nearest neighbor samples.

[0013] Preferably, the step of constructing a feature causal relationship matrix by calculating the Granger causal relationships between features of each dimension in the multidimensional time series monitoring data includes: for any two feature time series... and ,Establish The autoregressive model is then established and used simultaneously. and Historical values The regression model; the sum of squared residuals of the two models is analyzed. Test; when Inspection When determining features It is a feature Granger causes, in the characteristic causal matrix of the 1st... Line 1 Record the column as 1, otherwise record it as 0.

[0014] Preferably, the first preset threshold is The second preset threshold is ;in, The first preset threshold, The second preset threshold, Modularity score for dividing the community.

[0015] Preferably, the railway multi-dimensional time-series monitoring data includes axle box temperature, motor current, gearbox vibration, and vehicle body sway acceleration of the high-speed train.

[0016] In the second aspect, a distributed railway anomaly detection model collaborative training system includes:

[0017] The system includes a processor and a memory, the memory storing computer program instructions for collaborative training of a distributed railway anomaly detection model. When the computer program instructions are executed by the processor, the aforementioned collaborative training method for a distributed railway anomaly detection model is implemented.

[0018] The beneficial effects of this invention are as follows: In the view construction stage, by calculating the Granger causal relationship between features of each dimension and combining it with the community detection algorithm, this invention can decompose multi-dimensional time-series monitoring data into mutually independent views, ensuring the sufficiency and diversity of the views, providing a high-quality input foundation for collaborative training, and improving the learning potential of the model; In the collaborative training stage, this invention determines pseudo-label samples by combining two conditions: prediction confidence and local outlier factor, eliminating high-confidence noise samples and improving the quality of pseudo-labels; In the decision fusion stage, the fusion weight is determined by calculating the prediction variance of each model based on the nearest neighbor information of the data to be detected, which can take into account the reliability of different models on specific samples, obtain a more reliable model, and thus improve the accuracy of anomaly detection. Attached Figure Description

[0019] Figure 1 The flowchart illustrating the steps of the distributed railway anomaly detection model collaborative training method in this embodiment is shown in the illustration.

[0020] Figure 2 The schematic diagram illustrates the structural block diagram of the distributed railway anomaly detection model collaborative training system in this embodiment. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0022] like Figure 1 As shown, the distributed railway anomaly detection model collaborative training method in this embodiment includes steps S1 to S4:

[0023] Step S1: Acquire multi-dimensional time-series monitoring data of the railway, calculate Granger causality relationships between features of each dimension in the multi-dimensional time-series monitoring data to construct a feature causality matrix; use a community detection algorithm to partition the feature causality matrix to obtain mutually independent views. Figure 1 Feature subsets and views Figure 2 Feature subsets are used to obtain modularity scores for community partitioning, and multidimensional time-series monitoring data is split into visual subsets. Figure 1 Heshi Figure 2 .

[0024] In one embodiment, the railway multidimensional time-series monitoring data includes axle box temperature, motor current, gearbox vibration, and vehicle body sway acceleration of the high-speed train.

[0025] The acquired railway multidimensional time-series monitoring data contains multiple dimensions of features; specifically, the railway multidimensional time-series monitoring data includes dimensions such as axle box temperature, motor current, gearbox vibration, and vehicle body sway acceleration of high-speed trains.

[0026] In one embodiment, constructing a feature causal relationship matrix by calculating Granger causal relationships between features of various dimensions in multidimensional time-series monitoring data includes:

[0027] For any two characteristic time series and ,Establish The autoregressive model is then established and used simultaneously. and Historical values The regression model; the sum of squared residuals of the two models is analyzed. Test; when Inspection When determining features It is a feature Granger causes, in the characteristic causal matrix of the 1st... Line 1 Record the column as 1, otherwise record it as 0.

[0028] Specifically, for any two features and By establishing a regression model to test the causal relationship between two features, if the past value of one feature can improve the prediction accuracy of the future value of another feature, then a Granger causal relationship is considered to exist, thus enabling the determination of the feature's causal relationship. Is it a feature? Granger causation, by using Granger causation, can capture deeper, temporally sequential predictive relationships between features.

[0029] Understandably, firstly, regarding the features and To establish a system based solely on The regression model of its own historical values, that is The autoregressive model is used to predict... The current value; for example, the value from the previous time step. Value decremented by 1 and the previous two time points Subtract 2 from the value to predict the current time. Value. After model training, a set of prediction errors will be obtained. The sum of squares of these prediction errors is calculated and denoted as the residual sum of squares. Secondly, regarding features and Then, a second, more complex regression model is built, which simultaneously utilizes features. Historical values ​​and characteristics The historical value is used as the predictor variable, using the previous moment's value. Value decremented by 1 and the previous two time points Value minus 2 and using the value from the previous moment Value decremented by 1 and the previous two time points Subtract 2 from the value to predict the current time. Value; after model training, the sum of squared prediction errors is calculated and denoted as . Due to the establishment More information was input into the regression model. Must be less than or equal to ;right and implement The verification will generate a result after the verification. Statistic and a Value; when it is born When the value is less than a preset level, in this embodiment the preset level is set to 0.05 based on experience, which indicates that the feature... Historical information for prediction It played a role in the future, enabling the determination of features. It is a feature The Granger cause, at this point the th of the characteristic causal relationship matrix Line 1 The column is recorded as 1. Conversely, when it is generated... When the value is ≥0.05, then the first [value] of the characteristic causal relationship matrix is ​​[value]. Line 1 The column is recorded as 0.

[0030] The above process will be repeated on all feature pairs in the multidimensional time series monitoring data, and any two feature time series in the data will be analyzed in pairs. The results will be recorded in the corresponding positions of the feature causal relationship matrix to construct a complete feature causal relationship matrix.

[0031] In one embodiment, the partitioning of the feature causal relationship matrix using a community detection algorithm includes:

[0032] The Louvain community detection algorithm is employed. This involves iteratively assigning nodes to communities that maximize the modularity gain of the feature causal relationship matrix, and then aggregating these communities into new nodes for further assignment, until the modularity score no longer increases, thus obtaining mutually independent views. Figure 1 Feature subsets and views Figure 2 Feature subset.

[0033] Specifically, the feature causal relationship matrix is ​​viewed as a network graph, where features are nodes and Granger causal relationships are weighted edges. To features of At that time, draw a line from the network diagram. Pointing to The arrow, the weight is The statistical parameters are processed using community detection algorithms such as Louvain. The Louvain community detection algorithm iteratively optimizes the partitioning of closely related feature nodes into the same community until the module degree score of the entire network is maximized, resulting in two or more communities. In this embodiment, two communities are obtained, i.e., [the network is considered to have two communities]. Figure 1 Feature subsets and views Figure 2 Feature subsets. At this point, the modularity score is recorded, and the original multidimensional time-series monitoring data matrix is ​​split into two data subsets based on the feature subset division, corresponding to the view respectively. Figure 1 Heshi Figure 2 .

[0034] Understandably, taking 10 features as an example, there are 10 independent communities initially. The first stage, modularity optimization, involves iterating through each feature node (e.g., feature 1) and attempting to move it to the community of a neighboring node (e.g., feature 2's community). For each move attempt, a modularity gain is calculated; for example, moving feature 1 into feature 2's community yields a gain of 0.05. Then, the move that yields the maximum positive gain is selected and executed. This process is repeated for all nodes until no further movement of any node in the network can improve the overall modularity. Once the first stage stabilizes, the second stage, community aggregation, begins: the multiple communities formed in the previous stage are treated as new single nodes; for example, features 1, 2, and 3 constitute community one, and features 4, 5, and 6 constitute community two. The connection weights between the new nodes are determined by the sum of the connection weights between the original nodes, thus constructing a smaller network. Then, the process of modularity optimization is repeated on the new aggregated network, returning to the first stage. The two phases described above are executed alternately and iteratively until the module score of the entire network no longer increases. At this point, the community detection algorithm will output a stable community partitioning result. For example, the 10 features are partitioned into two largest, tightly connected, and independent communities. These two communities are the view... Figure 1 Feature subsets and views Figure 2 Feature subset.

[0035] In obtaining vision Figure 1 Feature subsets and views Figure 2 When defining feature subsets, the modularity score is recorded. The modularity score is a quantitative evaluation of the community segmentation effect; a higher value indicates better independence and higher quality between the two segmented views. Then, based on the feature subset segmentation, the original multidimensional time-series monitoring data matrix is ​​split into two data subsets, corresponding to the views respectively. Figure 1 Heshi Figure 2 This allows for the generation of high-quality training views, providing a foundation for subsequent collaborative training.

[0036] Step S2, using the original labeled dataset, respectively in the view Figure 1 Heshi Figure 2 The training process yields initial model A and initial model B.

[0037] Specifically, using the original labeled dataset, which includes a small amount of labeled normal and abnormal sample data, based on the visual data obtained in step S1... Figure 1 Feature subsets and views Figure 2 Feature subsets, which also split the labeled data into two view datasets, correspond to the view respectively. Figure 1 Heshi Figure 2 In view Figure 1On the dataset, a deep learning model is trained to obtain an initial model A. In the view... Figure 2 On the dataset, another independent deep learning model is trained to obtain the initial model B. The two models are different due to their different input features.

[0038] The deep learning model is a Long Short-Term Memory (LSTM) network. Before training the model, the samples in each view are processed using a minimax normalization method to eliminate the influence of the dimensions and numerical scales of the features.

[0039] Step S3: During the collaborative training iteration, perform the following operations alternately: use one of the initial models to predict the unlabeled dataset, select samples with prediction confidence higher than the first preset threshold and local outlier factors calculated based on the k nearest neighbors in the unlabeled dataset lower than the second preset threshold as pseudo-label samples, and add the pseudo-label samples to the training set of the other initial model for updating; the number of new pseudo-label samples in each iteration is inversely proportional to the consistency rate of the prediction results of the initial model A and the initial model B on the validation set at the end of the previous iteration.

[0040] In one embodiment, the prediction confidence is determined in the following manner:

[0041] For any sample in the unlabeled dataset, the maximum class probability output by the initial model used for prediction is taken as the prediction confidence.

[0042] Specifically, the sample selection process is performed on every single sample in the unlabeled dataset. Taking data points as an example, calculate the sample The prediction confidence level. Specifically, during the collaborative training iteration process, the initial model is used... and the initial model One method involves predicting class labels from an unlabeled dataset. When calculating prediction confidence, the samples... The input is fed into the currently trained initial model, and the softmax layer of the initial model outputs a probability distribution vector; where, when using the initial model... When making predictions, the samples are... Input to the initial model In the middle, the initial model is also being used. When making predictions, the samples are... Input to the initial model The maximum value of the probability value in the probability distribution vector is defined as the sample. The prediction confidence level.

[0043] The local outlier factor is calculated as follows:

[0044] Set nearest neighbor parameters The value is 20; for any sample in the unlabeled dataset, find the 20 nearest neighbor samples of the sample, calculate the reachability distance of each nearest neighbor and obtain the local reachability density, and calculate the local outlier factor of the sample.

[0045] Specifically, find samples in the entire unlabeled dataset. The 20 nearest neighbors by Euclidean distance are used as samples. Find the 20 nearest neighbor samples; calculate the distance from each of the 20 nearest neighbor samples to its 20th nearest neighbor sample, i.e. Distance; Calculate samples The reachable distance to each neighbor sample, wherein the reachable distance is taken from the sample. The actual distance (i.e., Euclidean distance) to neighboring samples and the neighboring samples' The larger of the two distances; the sample Calculate the reciprocal of the average reachable distance to the 20 nearest neighbor samples to obtain the sample. Locally accessible density; calculate sample The local reachability density of 20 neighboring samples is calculated and averaged, then the average is divided by the sample size. The local reachability density of a given entity is used to obtain the Local Outlier Factor (LOF).

[0046] The prediction confidence level is compared with a first preset threshold, and the local outlier factor is compared with a second preset threshold. Only when the prediction confidence level is higher than the first preset threshold and the local outlier factor is lower than the second preset threshold is the sample... Only then can it be confirmed as a high-quality, non-outlier sample and selected as a pseudo-label sample, so that the pseudo-label sample can be added to the training set of another initial model, and then the other initial model can be updated.

[0047] Before calculating the prediction confidence and local outlier factor, the samples in the unlabeled dataset must be processed using the minimax normalization method to eliminate the influence of the dimensions and numerical scales of the features.

[0048] In one embodiment, the first preset threshold is The second preset threshold is ;in, The first preset threshold, The second preset threshold, Modularity score for dividing the community.

[0049] Specifically, after executing the community detection algorithm, a modularity score is obtained. Modularity scores typically range from 0 to 1. A higher modularity score indicates better view partitioning and stronger independence between views. Based on the obtained modularity scores, a first preset threshold (prediction confidence threshold) and a second preset threshold (local outlier factor threshold) are calculated. For example, the modularity score obtained after one feature partitioning... A score of 0.8 indicates a high-quality view partition. The first preset threshold is calculated to be 0.76, and the second preset threshold is 1.875. Therefore, when the view quality is high, only samples with a prediction confidence higher than 0.76 and a local outlier factor lower than 1.875 are trusted. When the partitioning effect is poor, the obtained modularity score... If the value is relatively low, the first preset threshold will be lowered accordingly, while the second preset threshold will be raised accordingly. This will relax the screening criteria when the view quality is poor, ensuring that enough pseudo-label samples are added to the training set of another initial model. The other initial model will be retrained using this new training set, thereby updating the other initial model.

[0050] In one embodiment, the number of newly added pseudo-label samples in each iteration is inversely proportional to the consistency rate of the prediction results of the initial model A and the initial model B on the validation set at the end of the previous iteration, including:

[0051] Calculate the initial model Compared with the initial model The consistency rate is calculated by predicting the proportion of samples with identical labels on the validation set. The number of newly added pseudo-labeled samples in this round is set to... ,in, This represents the number of newly added pseudo-labeled samples in this round. The value is the consistency rate, 100 is the baseline sample size, and 0.01 is a smoothing term to prevent the denominator from being zero.

[0052] Specifically, after each round of collaborative training iteration, an independent validation set is used to evaluate the initial model. and the initial model That is, the initial model Compared with the initial model The consistency of predictions; for example, the validation set contains 1000 samples, and the initial model... and the initial model Predictions were made for 1000 samples, resulting in 1000 predicted class labels, categorized as either normal or abnormal. The initial model was statistically analyzed by comparing the prediction results one by one. Compared with the initial model The number of samples with identical predicted class labels; for example, if 800 samples have identical predicted labels, then the consistency rate is calculated. It is 0.8.

[0053] Using consistency rate Through formula Determine how many new pseudo-labeled samples need to be selected from the unlabeled dataset in the next iteration; for consistency rate When it is 0.8, Approximately equal to 123. (This refers to the initial model in the previous round.) Compared with the initial model When the consistency rate is low, then It will increase; for example, the consistency rate. When it is 0.5, Approximately 196. The inverse relationship indicates that when the initial model... Compared with the initial model They exhibit a high degree of consistency, i.e., a high consistency rate. When the difference is large, it indicates that the two models have converged. In this case, the number of new pseudo-label samples should be reduced to prevent overfitting. Conversely, when the difference between the two models is large, the number of pseudo-label samples should be increased to promote mutual learning between the two models and thus accelerate convergence.

[0054] Step S4: When training terminates, based on the nearest neighbor samples of the data to be detected in the original labeled dataset, calculate the prediction variance of the initial model A and the initial model B on the nearest neighbor samples, and determine their respective fusion weights according to the reciprocal of the prediction variances; use the fusion weights to perform weighted fusion of the prediction results of the initial model A and the initial model B to obtain the anomaly detection result.

[0055] Specifically, training terminates when the preset number of iterations is reached or when the performance of the initial model on the validation set no longer improves.

[0056] For a new dataset to be detected, find similar datasets in the original labeled dataset using Euclidean distance or Mahalanobis distance. The nearest neighbor samples. Each nearest neighbor sample is input into the initial model. and the initial model In the middle, we get Each prediction result. Calculate the initial model separately. of The variance of each prediction result and the initial model of The variance of each predicted outcome. Initial model. The fusion weights are determined as the initial model. Divide the reciprocal of the prediction variance by the sum of the reciprocals of the prediction variances of the two initial models; similarly, calculate the initial model... The fusion weight.

[0057] Initial model The direct prediction result of the data to be detected is multiplied by the initial model. The fusion weights, plus the initial model The direct prediction result of the data to be detected is multiplied by the initial model. The fusion weights and sum of the values ​​represent the anomaly detection results.

[0058] In an optional embodiment, the step of calculating the prediction variances of initial model A and initial model B on the nearest neighbor samples in the original labeled dataset based on the data to be detected, and determining their respective fusion weights according to the reciprocal of the prediction variances, includes:

[0059] Using Euclidean distance as the metric, the 15 nearest neighbor samples for the data to be detected are found in the original labeled dataset; the initial model is calculated respectively. and the initial model The variance of the predicted probability values ​​for 15 nearest neighbor samples;

[0060] Initial model The fusion weights satisfy: Initial model The fusion weights satisfy: In the formula, , The initial model , The fusion weight, , The initial model , The variance of the predicted probability values ​​for 15 nearest neighbor samples.

[0061] Specifically, when classifying a piece of data to be detected, the local neighborhood is located in the original labeled dataset. The value is set to 15. Specifically, the minimum-maximum normalization method is applied to all samples in the data to be detected and the original labeled dataset to eliminate the influence of the dimensions and numerical scales of the features. Then, by calculating the Euclidean distance between the data to be detected and all samples in the original labeled dataset, the 15 labeled samples that are most similar to the data to be detected (i.e., the closest in distance) are found to form the nearest neighbor samples.

[0062] The 15 nearest neighbor samples constitute a reference set for evaluating the local stability of the model. Initial model and the initial model Each initial model makes predictions for its 15 nearest neighbor samples, outputting 15 predicted probability distribution vectors. The stability (prediction variance) of the two initial models is calculated independently. For example, for the initial model... The variance of the predicted probability values ​​for each of the 15 output probability distribution vectors is calculated, and the variances are summed or averaged to obtain a total variance value. ; The smaller the value, the more representative the initial model. The prediction results for samples within the local neighborhood are highly consistent and stable. Similarly, the initial model is calculated. Total variance .

[0063] The fusion weights are assigned based on the reciprocal of the prediction variance. The smaller the prediction variance, the more stable the model performance, and the higher the fusion weight can be obtained. This ensures that the opinions of the more stable models dominate in the fusion decision.

[0064] This invention also provides a distributed collaborative training system for railway anomaly detection models. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions for collaborative training of a distributed railway anomaly detection model. When the computer program instructions are executed by the processor, the collaborative training method for a distributed railway anomaly detection model according to the present invention is implemented.

[0065] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0066] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented by computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0067] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0068] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A collaborative training method for a distributed railway anomaly detection model, characterized in that, include: Acquire multidimensional time-series monitoring data of railways, calculate Granger causal relationships between features of each dimension in the multidimensional time-series monitoring data to construct a feature causal relationship matrix; use community detection algorithm to divide the feature causal relationship matrix to obtain mutually independent feature subsets of view one and view two, and obtain the modularity score of community division, thus splitting the multidimensional time-series monitoring data into view one and view two. Using the original labeled dataset, we trained on View 1 and View 2 respectively to obtain initial model A and initial model B; During the collaborative training iteration, the following operations are performed alternately: one of the initial models is used to predict the unlabeled dataset, and samples with prediction confidence higher than a first preset threshold and local outlier factors calculated based on the k nearest neighbors in the unlabeled dataset lower than a second preset threshold are selected as pseudo-label samples. The pseudo-label samples are then added to the training set of the other initial model for updates. The number of new pseudo-label samples in each iteration is inversely proportional to the consistency rate of the prediction results of the initial model A and the initial model B on the validation set at the end of the previous iteration. When training terminates, the prediction variances of the initial model A and the initial model B on the nearest neighbor samples in the original labeled dataset are calculated based on the data to be detected, and their respective fusion weights are determined according to the reciprocal of the prediction variances. The prediction results of the initial model A and the initial model B are weighted and fused using fusion weights to obtain the anomaly detection results.

2. The collaborative training method for a distributed railway anomaly detection model according to claim 1, characterized in that, The process of partitioning the feature causal relationship matrix using a community detection algorithm includes: The Louvain community detection algorithm is adopted. By continuously dividing nodes into communities that maximize the modularity gain of the feature causal relationship matrix, and iteratively aggregating communities into new nodes for further division, until the modularity score no longer increases, we obtain mutually independent feature subsets of view one and view two.

3. The collaborative training method for a distributed railway anomaly detection model according to claim 1, characterized in that, The prediction confidence level is determined in the following way: For any sample in the unlabeled dataset, the maximum class probability output by the initial model used for prediction is taken as the prediction confidence.

4. The collaborative training method for a distributed railway anomaly detection model according to claim 1, characterized in that, The local outlier factor is calculated as follows: Set nearest neighbor parameters The value is 20; for any sample in the unlabeled dataset, find the 20 nearest neighbor samples of the sample, calculate the reachability distance of each nearest neighbor and obtain the local reachability density, and calculate the local outlier factor of the sample.

5. The collaborative training method for a distributed railway anomaly detection model according to claim 1, characterized in that, The number of newly added pseudo-label samples in each iteration is inversely proportional to the consistency rate of the prediction results of the initial model A and the initial model B on the validation set at the end of the previous iteration, including: Calculate the initial model Compared with the initial model The consistency rate is calculated by predicting the proportion of samples with identical labels on the validation set. The number of newly added pseudo-labeled samples in this round is set to... ,in, This represents the number of newly added pseudo-labeled samples in this round. The value is the consistency rate, 100 is the baseline sample size, and 0.01 is a smoothing term to prevent the denominator from being zero.

6. The collaborative training method for a distributed railway anomaly detection model according to claim 1, characterized in that, The process involves calculating the prediction variances of initial model A and initial model B on the nearest neighbor samples in the original labeled dataset based on the data to be detected, and determining their respective fusion weights based on the reciprocal of the prediction variances, including: Using Euclidean distance as the metric, the 15 nearest neighbor samples for the data to be detected are found in the original labeled dataset; the initial model is calculated respectively. and the initial model The variance of the predicted probability values ​​for 15 nearest neighbor samples; Initial model The fusion weights satisfy: Initial model The fusion weights satisfy: In the formula, , The initial model , The fusion weight, , The initial model , The variance of the predicted probability values ​​for 15 nearest neighbor samples.

7. The collaborative training method for a distributed railway anomaly detection model according to claim 1, characterized in that, The calculation of Granger causality relationships between features of various dimensions in multidimensional time-series monitoring data to construct a feature causality matrix includes: For any two characteristic time series and ,Establish The autoregressive model is then established and used simultaneously. and Historical values The regression model; the sum of squared residuals of the two models is analyzed. Test; when Inspection When determining features It is a feature Granger causes, in the characteristic causal matrix of the 1st... Line 1 Record the result as 1 if it is positive, otherwise record it as 0.

8. The collaborative training method for a distributed railway anomaly detection model according to claim 1, characterized in that, The first preset threshold is The second preset threshold is ;in, The first preset threshold, The second preset threshold, Modularity score for dividing the community.

9. The collaborative training method for a distributed railway anomaly detection model according to claim 1, characterized in that, The railway multidimensional time-series monitoring data includes axle box temperature, motor current, gearbox vibration, and vehicle body sway acceleration of high-speed trains.

10. A distributed collaborative training system for railway anomaly detection models, characterized in that, include: A processor and a memory, the memory storing computer program instructions for collaborative training of a distributed railway anomaly detection model, wherein when the computer program instructions are executed by the processor, the method for collaborative training of a distributed railway anomaly detection model according to any one of claims 1-9 is implemented.

Images