CN-121980305-A - Railway accident cause identification method and system based on double-cost sensitive graph attention network

CN121980305ACN 121980305 ACN121980305 ACN 121980305ACN-121980305-A

Abstract

The invention provides a railway accident cause identification method and a system based on a double-cost sensitive graph attention network, which belong to the technical field of railway accident identification, and utilize a pre-trained railway accident cause identification model to process acquired real railway accident data to be processed to obtain a railway accident cause analysis result, wherein in the training of the railway accident cause identification model, graph theory is introduced into a railway accident scene, an accident scene information graph structure which represents internal association of a complex railway accident scene is constructed as the input of the railway accident cause identification model, and based on a cost sensitive classification layer and a graph characteristic distance separation layer, misjudgment cost distribution is optimized, and class generalization capability is enhanced by utilizing the geometric characteristics of a characteristic space. The invention optimizes the misjudgment cost distribution and enhances the class generalization capability by utilizing the geometric characteristics of the feature space, realizes high robustness performance in the unbalanced classification task of the accident cause, and shows stronger recognition sensitivity to the low-frequency key cause.

Inventors

MA XIAOPING
WANG RUOJIN

Assignees

北京交通大学

Dates

Publication Date: 20260505
Application Date: 20251231

Claims (10)

1. A railway accident cause identification method based on a double-cost sensitive graph attention network is characterized by comprising the following steps: acquiring real railway accident data to be processed; The method comprises the steps of obtaining real railway accident data to be processed by utilizing a pre-trained railway accident cause identification model, obtaining a railway accident cause analysis result, introducing graph theory into a railway accident scene in training of the railway accident cause identification model, constructing an accident scene information graph structure which characterizes the internal association of a complex scene of the railway accident, taking the graph structure as the input of the railway accident cause identification model, optimizing misjudgment cost distribution based on a cost sensitive classification layer and a graph characteristic distance separation layer by the railway accident cause identification model, and enhancing class generalization capability by utilizing the geometric characteristics of a characteristic space, wherein the cost sensitive classification layer is used for distributing different cost values for each class in cross entropy loss based on the number of class samples, so as to obtain unbiased classification hyperplane, forcing the classification layer to strengthen the discrimination capability of few classes by amplifying punishment force of few classes of misclassification, and the cost sensitive graph characteristic distance separation layer is used for calculating the characteristic vector-to-feature Euclidean distance between samples in the same training batch, distributing a sensitivity coefficient according to class labels, and extracting characteristic representation similarity of similar samples by restraining the similar samples.
2. The method for identifying the cause of a railway accident based on a dual-cost sensitive graph attention network according to claim 1 is characterized in that a railway accident scene information graph structure is constructed, graph theory is introduced into a railway accident scene, edge weights are calculated by using a maximum mutual information coefficient algorithm, directed edges are determined by using the mean deviation of the relation strength, and the railway accident scene information graph is expressed as a node set Weighted adjacency matrix Drawing of composition The method specifically comprises the steps of respectively calculating the maximum Mutual Information Coefficient (MIC) between railway scene elements Thereby constructing a scene element relation matrix , Determining weight matrix, defining edge weight of graph structure as MIC score between scene elements, determining correlation threshold between elements so as to construct weighted adjacent matrix of railway accident scene information graph And determining the direction of the graph structure side, wherein the average deviation of the correlation strength among the accident scene elements is defined as the direction of the graph structure side.
3. The railway accident cause identification method based on the dual-cost sensitive graph attention network according to claim 1 is characterized in that the cost sensitive classification layer is used for distributing different cost values for each class in cross entropy loss based on the number of class samples so as to obtain unbiased classification hyperplane, the classification layer is forced to strengthen the discrimination capability of the minority class by amplifying the punishment force of minority class misclashing, and concretely comprises the steps of determining a class weight function, firstly, defining a weight function related to the number of class samples in order to ease the neglect of the tendency of a classifier to a majority class accident cause sample and minority class samples, wherein the core idea of the function is to distribute misclassification value for each class according to the number of samples of different accident cause classes, distributing different cost values for each class in the cross entropy loss based on the number of class samples in the classification layer so as to obtain unbiased classification hyperplane, and integrating the class weight function into a standard cross entropy loss function after obtaining the class weight function so as to construct a cost sensitive cross entropy loss function to evaluate inferred probability representation.
4. The railway accident cause identification method based on the dual cost sensitive graph attention network according to claim 1 is characterized by constructing a cost sensitive graph feature distance separation layer, calculating Euclidean distance between samples in a batch and feature vectors in the same training batch, distributing cost sensitive coefficients according to class labels, and extracting more discriminative accident cause feature representations by restraining feature representations similarity of similar samples and feature representations separation of heterogeneous samples, and specifically comprises determining a distance weighting coefficient, applying constraint on feature distance between samples in an embedding space for forcing models to learn features with high discriminant, and applying constraint on feature distance between samples for any two samples And To extract the accident cause characteristic representation with more discrimination, the following constraint is satisfied when the sample is And When the accident cause category is the same category, the method is characterized in that the distance in an embedded space is smaller, the characteristic distance of a heterogeneous sample is larger, a cost sensitive graph characteristic distance separating layer is constructed, in order to further enable a minority sample to form a tighter cluster structure in the characteristic space, the separation degree of the minority sample from a majority decision boundary is enhanced, and the category weight function of the category is fused And weighted distance coefficient And constructing a characteristic distance cost sensitivity loss function.
5. The method for identifying railway accident causes based on the dual-cost sensitive graph attention network according to claim 4 is characterized in that a weighting function is designed for Euclidean distance, wherein similar features are promoted to be tightly gathered through positive weighting, and different features are forced to be far away through negative weighting, so that feature separation loss is constructed.
6. The railway accident cause identification method based on the double-cost sensitive graph attention network is characterized by comprising the steps of constructing a weighted loss function of accident cause classification, combining a cost sensitive cross entropy loss function and a characteristic distance cost sensitive loss function, introducing a characteristic alignment coefficient, balancing the cross entropy loss function and the characteristic distance cost sensitive loss function weight to obtain a total loss function of a model, and fusing a cost sensitive classification layer and a characteristic distance separation strategy by the model in the training process, so that the classification performance of the model on unbalanced data is improved.
7. A railway accident cause identification system based on a dual cost sensitive graph attention network, comprising: The acquisition module is used for acquiring real railway accident data to be processed; The processing module is used for processing the acquired real railway accident data to be processed by utilizing a pre-trained railway accident cause identification model to obtain a railway accident cause analysis result, wherein in the training of the railway accident cause identification model, a graph theory is introduced into a railway accident scene to construct an accident scene information graph structure which represents the internal association of a complex scene of the railway accident, the graph structure is used as the input of the railway accident cause identification model, the railway accident cause identification model is based on a cost sensitive classification layer and a graph feature distance separating layer to optimize misjudgment cost distribution and enhance class generalization capability by utilizing the geometric characteristics of a feature space, the cost sensitive classification layer is used for distributing different cost values for each class in cross entropy loss based on the number of class samples, so that an unbiased classification hyperplane is obtained, the classification layer is forced to strengthen the discrimination capability of a few classes by amplifying punishment force of minority misclasis achieved, the cost sensitive graph feature distance is used for calculating the inter-feature vector Euclidean distance of samples in the same training batch, the cost sensitive graph separation layer is used for distributing a sensitivity coefficient according to class labels, and the feature of the similar class samples is more similar in terms, and the feature of the similar class samples are more similar in terms is represented by constraint on the class labels.
8. A non-transitory computer readable storage medium storing computer instructions which, when executed by a processor, implement the method of double cost sensitive graph attention network based railway accident cause identification of any one of claims 1 to 6.
9. A computer device comprising a memory and a processor, the processor and the memory being in communication with each other, the memory storing program instructions executable by the processor, the processor invoking the program instructions to perform the dual cost sensitive graph attention network based railway accident cause identification method according to any one of claims 1-6.
10. An electronic device comprising a processor, a memory, and a computer program, wherein the processor is coupled to the memory, the computer program being stored in the memory, the processor executing the computer program stored in the memory when the electronic device is operating, to cause the electronic device to execute instructions for implementing a method of identifying railway accident causes based on a dual cost sensitive graph attention network as claimed in any one of claims 1 to 6.

Description

Railway accident cause identification method and system based on double-cost sensitive graph attention network Technical Field The invention relates to the technical field of railway accident identification, in particular to a railway accident cause identification method and system based on a double-cost sensitive graph attention network. Background The safety and stability of railway transportation are guaranteed, and the method is an important foundation for realizing sustainable development. The occurrence of railway accidents often causes the serious consequences of infrastructure damage, equipment failure, casualties and the like. The accurate cause identification of railway accidents can effectively help to formulate a targeted prevention strategy, is a key point for improving the running safety of trains, and is a primary step of accident prevention. Railway accidents are typically caused by random combinations of various accident scenario elements, such as track conditions, train conditions, and signal control systems. Under different accident scenarios, there may be differences in the leading cause categories of the induced accidents. For example, in the case of "level crossings", "low visibility", the main cause of a collision accident is often "human operation factor (such as lookout negligence)", whereas in the case of "high temperature" and "truck loading", the main cause of the same collision accident may become "mechanical and electrical faults". The characteristic that the cause category dynamically evolves along with the scene makes an accident prevention strategy based on a macroscopic rule limited. Therefore, there is a need for an accident cause identification model considering multiple scene elements, which reveals essential association between different element combinations and accident causes by accurately classifying railway accident causes, and can also predict the accident causes of key risk scenes during train operation, thereby implementing targeted intervention to reduce accident occurrence probability. The traditional accident analysis of multi-focus isolated accident elements and post statistics of macroscopic rules often neglect the decisive influence of specific accident scenes generated by accidents on the identification of accident causes. The accident modeling method based on the graph structure provides scientific basis for formulating preventive measures and strengthening a safety system, wherein the graph neural network has obvious application potential in the railway accident analysis field by virtue of the advantages of fitting accident randomness and multi-factor characteristics, being good at revealing interaction relation among nodes and strong learning and computing capability. However, when the graphic neural network is directly applied to railway accident cause identification under different accident scenes, two fundamental challenges are still faced. First of all, in the high complexity inherent in the railway accident scenario itself. Accident scene elements are complex and nonlinear data, and railway accidents often occur due to the coupling effect of multiple scene elements. If potential dependencies between accident scene elements are ignored, the information is underutilized, which weakens the accuracy and interpretation of the accident cause predictions. In addition, the railway accident data set covers various accident types such as derailment, collision, and the like. Under certain accident types, the accident cause distribution in different scene element combinations presents a significant imbalance. If the accident cause prediction model fails to effectively process the high unbalance of the cause distribution among the scenes in the training stage, the model is excessively biased to identify the high-frequency causes, and the low-frequency causes under the same scene are ignored. The model evaluation index can not reflect the balanced recognition capability of the model to each category, even if the overall accuracy value of the model is higher, the serious missed detection risk to the low-frequency key causes can be covered, and finally, a potential safety hazard blind area is formed. Disclosure of Invention The invention aims to provide a railway accident cause identification method and system based on a double-cost sensitive graph attention network, so as to solve at least one technical problem in the background technology. In order to achieve the above purpose, the present invention adopts the following technical scheme: In a first aspect, the present invention provides a railway accident cause identification method based on a dual cost sensitive graph attention network, including: acquiring real railway accident data to be processed; The method comprises the steps of obtaining real railway accident data to be processed by utilizing a pre-trained railway accident cause identification model, obtaining a railway accident cause analysis result, int