A method for identifying risk factors of operation safety of a high-speed train system

Abstract

The application relates to the technical field of risk identification, and discloses a high-speed train system operation safety risk factor identification method, which comprises the following steps: obtaining historical event reports and fault handling reports of high-speed trains as original data sets; performing text preprocessing on the obtained original data sets to obtain preprocessed texts; performing risk factor extraction on the preprocessed texts based on a DeBERTa-BiLSTM model, and outputting the probabilities of various risk factors. The DeBERTa model and the BiLSTM model are combined, the advantages of the DeBERTa in deep semantic feature extraction and the ability of the BiLSTM in bidirectional time sequence dependence capture are fully exerted, risk factors can be accurately extracted from unstructured texts such as high-speed train historical event reports and fault handling reports, not only the explicit risks can be identified, but also the hidden risk words and characters that are covered can be effectively restored through an enhanced mask decoder, and the accuracy and recall rate of risk factor extraction are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of risk identification technology, and in particular to a method for identifying risk factors in the operation safety of high-speed train systems. Background Technology

[0002] The safe operation of high-speed trains is crucial to people's lives and property and the stability of the national transportation network. Timely and accurate identification of risk factors during operation is key to preventing accidents and improving system reliability. Currently, the high-speed train system has accumulated a large number of historical event reports and fault handling reports. These texts record a wealth of information on operational anomalies and serve as an important data source for risk identification.

[0003] However, traditional risk analysis relies heavily on expert experience and structured statistical data, making it difficult to efficiently process massive amounts of unstructured text reports. Event reports often contain complex technical terms, implicit risk descriptions, and intertwined causal chains. Conventional keyword matching or shallow natural language processing methods cannot fully extract the deep semantic information in the text, easily missing hidden risk terms, resulting in low accuracy and recall rates for risk factor extraction.

[0004] Furthermore, existing risk identification methods based on single deep learning models (such as standard BERT, LSTM, etc.) have limitations in feature extraction. While the standard BERT model can capture contextual semantics, it is insufficient in modeling the logical dependencies between sentences in long texts; while the LSTM model excels at temporal processing, it is weak in understanding complex professional semantics. Neither model, when used alone, can adequately address both "deep semantic connections" and "bidirectional temporal dependencies," thus limiting the improvement of risk identification accuracy. Summary of the Invention

[0005] This invention provides a method for identifying safety risk factors in high-speed train systems to address existing technical problems. It solves the technical problems of existing methods for identifying safety risks in high-speed train systems, such as difficulty in efficiently processing unstructured text reports, easy omission of hidden risk terms, inability of a single deep learning model to take into account both deep semantics and bidirectional temporal dependencies, lack of quantitative analysis of the coupling strength of multiple risk factors, and inability to adapt to multi-label classification scenarios.

[0006] To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a method for identifying safety risk factors in high-speed train system operation, comprising the following steps:

[0007] Step S1: Obtain historical event reports and fault handling reports of high-speed trains as the raw dataset;

[0008] Step S2: Perform text preprocessing on the original dataset obtained in step S1 to obtain preprocessed text;

[0009] Step S3: Extract risk factors from the preprocessed text based on the DeBERTa-BiLSTM model and output the probability of each risk factor;

[0010] Step S4: Based on the NK model, perform risk coupling analysis on the extracted risk factors, calculate the coupling value between different risk factors, and measure the degree of impact of risk coupling on the accident.

[0011] Furthermore, the text preprocessing in step S2 includes: filtering and deleting stop words, terminology standardization, establishing risk labels, and data annotation of the text.

[0012] Furthermore, the risk labels are established as primary labels according to the "human-machine-environment-management" framework, and secondary labels are established for each of them; the data labeling adopts a combination of machine labeling and manual labeling to label each event into one or more corresponding risk categories.

[0013] Furthermore, in step S3, the risk factor extraction based on the DeBERTa-BiLSTM model specifically includes:

[0014] Step S301: The preprocessed text is embedded and encoded using the DeBERTa model. Deep semantic features are extracted through the deentanglement attention mechanism, and the restored risk semantic feature sequence is output through the enhanced mask decoder.

[0015] Step S302: Input the risk semantic feature sequence into the BiLSTM model, extract bidirectional temporal dependency features through forward LSTM and backward LSTM respectively, and concatenate the forward output features and backward output features to obtain a bidirectional fused feature sequence.

[0016] Step S303: Perform global pooling on the bidirectional fused feature sequence to obtain a global feature vector, and then output the probability of each risk factor through a multi-label classifier.

[0017] Furthermore, the embedding encoding of the DeBERTa model includes: converting text words into content vectors and relative position vectors;

[0018] The content vector represents the semantics of the word itself, and the relative position vector represents the positional relationship of the text word relative to other words in the risk semantic feature sequence.

[0019] Furthermore, when calculating the attention weights between word pairs, the de-entanglement attention mechanism includes content-content attention, content-position attention, and position-content attention, and then adds and fuses these three types of attention.

[0020] Furthermore, in step S4, the formula for calculating the risk coupling value using the NK model is as follows:

[0021] ;

[0022] In the above formula, This represents the coupling value between two or more risk factors. This indicates the status of each risk factor, with a value of 1 indicating the presence of the risk factor and a value of 0 indicating the absence of the risk factor. This represents the joint probability of all risk factors simultaneously being in a specific combination of states. These represent the marginal probabilities of each risk factor being in a certain state independently.

[0023] Furthermore, the risk coupling analysis includes the analysis of single-factor coupling risk, two-factor coupling risk, and multi-factor coupling risk.

[0024] Furthermore, the historical event reports and fault handling reports are unstructured text data.

[0025] The present invention provides a method for identifying safety risk factors in high-speed train systems. Compared with existing technologies, the advantages of this method are as follows:

[0026] 1. This invention combines the DeBERTa model with the BiLSTM model, giving full play to the advantages of DeBERTa in deep semantic feature extraction and the ability of BiLSTM in capturing bidirectional temporal dependencies. It can accurately extract risk factors from unstructured texts such as high-speed train historical event reports and fault handling reports. It not only identifies explicit risks, but also effectively restores the hidden risk words that are masked through the enhanced mask decoder, which significantly improves the accuracy and recall of risk factor extraction.

[0027] 2. This invention constructs a risk coupling analysis method based on the NK model. By calculating the coupling values ​​between single factors, two factors and multiple factors, it quantifies the degree of joint impact of different risk factor combinations on the occurrence of accidents. It can reveal the interaction and co-evolution law between risk factors, thereby helping to identify high-risk coupling patterns that are easily overlooked in single risk analysis. This provides a more comprehensive theoretical basis for the formulation of risk warning and prevention and control strategies for high-speed train operation safety.

[0028] 3. This invention establishes multi-level risk labels according to the "human-machine-environment-management" system and adopts a multi-label classification method that combines machine labeling and manual labeling. Each event is simultaneously labeled into multiple risk categories, adapting to the complex situation of multiple risk factors coexisting in actual operational events. Combined with the probabilities of each risk factor output by the DeBERTa-BiLSTM model, it can simultaneously identify and quantify multi-dimensional risk factors, avoiding the limitations of traditional methods that can only attribute to a single cause, and improving the comprehensiveness and engineering applicability of risk identification.

[0029] 4. This invention utilizes the deentanglement attention mechanism in the DeBERTa model, fusing three attention weights: content-content attention, content-position attention, and position-content attention. This preserves semantic relationships between lexical units while also considering the logical order of the text, achieving risk feature focusing under the dual constraints of semantics and logic. Compared to traditional static word embedding methods, this dynamic contextual embedding approach is better suited to high-speed train terminology and complex sentence structures, significantly improving the model's understanding of domain texts and the depth of risk discovery. Attached Figure Description

[0030] Figure 1 This is a risk coupling analysis diagram for the present invention;

[0031] Figure 2 This is a structural diagram of the risk identification model based on DeBERTa-BilSTM of the present invention;

[0032] Figure 3 This is a logical structure diagram of the DeBERTa model in this invention;

[0033] Figure 4 This is a logical structure diagram of the BiLSTM model in this invention;

[0034] Figure 5 This is a risk coupling analysis diagram of high-speed train operation events in this invention. Detailed Implementation

[0035] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1

[0037] like Figure 1 As shown, according to one aspect of the present invention, a method for identifying safety risk factors in the operation of a high-speed train system is provided, comprising the following steps:

[0038] Step 1: Obtain historical event reports and fault handling reports of high-speed trains as the raw dataset;

[0039] Step 2: Perform text preprocessing on the original dataset obtained in Step 1 to obtain preprocessed text; text preprocessing includes filtering and deleting stop words, terminology standardization, establishing a risk database, and data annotation.

[0040] 1. Filter and delete stop words

[0041] Stop words are function words that are common in report text but have no actual meaning. Removing stop words during report text processing reduces the amount of indexing data, thereby improving retrieval efficiency and enhancing search results. Stop words fall into three categories:

[0042] 1) Words that are not meaningful for accident risk assessment, including "is", "of", "this", "that", as well as conjunctions, interjections, punctuation marks, etc.

[0043] 2) Words that appear frequently in the report but have no significant impact on revealing accident risks include "high-speed train", "accident", and "railway".

[0044] 3) Train numbers, routes, and station names appearing in the accident text are also listed as stop words. Appropriately handling stop words and cleaning up symbols can provide the model with cleaner and more useful input data, thereby improving the model's performance and accuracy.

[0045] 2. Terminology standardization

[0046] Terminology standardization unifies different variations of the terminology used to describe the same concept in different documents into a standard terminology to help models accurately understand and extract relevant information. During model building, different reports and data sources may use different terms to describe the same concept. For example, in high-speed rail operation safety incidents, terms such as "excessive axle temperature," "axle temperature exceeding limits," and "axle temperature exceeding threshold" may be synonyms.

[0047] Without standardization, the model might treat these terms as different entities, affecting its accuracy and consistency. By standardizing the above terms to "axis temperature anomaly," the model can consistently understand these terms, avoiding information misunderstandings caused by inconsistent terminology. The terminology standardization process not only improves the accuracy of word embeddings in the DeBERTa model but also enhances the overall performance of the model.

[0048] 3. Establish risk labels

[0049] Risk labels are fundamental to guiding model learning during training and are also key to evaluating model performance. High-speed rail operation risk data is mostly unstructured data such as event reports and fault handling reports, which cannot be recognized by the model if directly inputted.

[0050] Establishing risk labels and annotating text accordingly is equivalent to attaching "machine-understandable labels" to the text. Only by constructing a "text-label" training dataset can the model learn "which text features correspond to which risk types," achieving a leap from "text reading" to "risk identification." Risk labels are established from four aspects: "human-machine-environment-management," and secondary labels are established from each of these four aspects, resulting in a total of 20 categories. The risk labels are shown in Table 1 below.

[0051] Table 1 Risk Labels

[0052] 4. Data labeling

[0053] Data annotation is a prerequisite and crucial step for achieving automatic risk factor identification based on the DeBERTa model. By labeling risk entities and risk types in text according to risk tags, high-quality supervision signals can be provided to the model, clarifying the boundaries, categories, and distribution characteristics of risk-related semantic units. High-quality data annotation helps the DeBERTa model effectively learn the linguistic patterns, contextual dependencies, and semantic structures of risk expressions, thereby improving the model's accuracy and generalization ability in risk entity extraction tasks.

[0054] Without labeled information, models can only capture general semantic features and struggle to accurately identify risk factors in specific domains. This study employs a combination of machine and human labeling, assigning each event to one of 20 risk labels matching that event. In other words, events are categorized into the same class only if they involve the same risk factors.

[0055] Because the same problem may be related to multiple risk factors, any event may belong to multiple categories, making them multi-label problems. For example, a high-speed train experienced a 25-minute delay due to a pantograph failure to raise, requiring a backup train to take over the route. The investigation revealed the cause was that the driver failed to operate the pantograph lowering button. This event, caused by both equipment malfunction and improper operation, can therefore be labeled into two categories. Standardized, consistent, and high-quality labeling is a crucial foundation for ensuring the effectiveness of subsequent model training, optimization, and practical applications, directly determining the reliability and practicality of the risk identification system.

[0056] Step 3: Extract risk factors from the preprocessed text based on the DeBERTa-BiLSTM model and output the probability of each risk factor;

[0057] 1. Risk Factor Extraction

[0058] To address the limitations of single models in feature extraction and temporal processing, this paper develops a risk identification model based on DeBERTa-BilSTM, which combines the advantages of deep semantic feature extraction of DeBERTa with the advantages of bidirectional temporal dependency capture of BilSTM, thereby improving the accuracy of risk identification.

[0059] First, the DeBERTa model pre-trains the input report text, transforming each word into two vectors that encode its content and position respectively, stacking them into matrices, calculating attention weights, and extracting deep feature vectors.

[0060] Secondly, BilSTM captures the bidirectional temporal dependencies of data through its bidirectional structure, obtaining a fused feature sequence containing "deep correlation features + bidirectional temporal features." Finally, it combines this with a risk labeling system and uses feature importance analysis to obtain key risk factors. This structure is as follows: Figure 2 As shown.

[0061] 2. DeBERTa Embedding

[0062] To improve the semantic representation capability of text vectors in professional domains, the Deberta (Decoding-Enhanced BERT with Disentangled Attention) model, which has strong semantic representation capabilities, is used for text vector representation.

[0063] DeBERTa is a pre-trained language model, an improved version of the BERT model. It mainly optimizes the shortcomings of the BERT model in terms of attention mechanism and masking strategy, and has achieved excellent performance in the field of natural language processing.

[0064] The core advantage of the DeBERTA model lies in its introduction of an unentangled attention mechanism and an enhanced mask decoder, enabling it to extract deep semantic and relational features from text and structured data more efficiently. Compared to models such as BERT and RoBERTa, it shows significant improvements in feature extraction accuracy, training efficiency, and generalization ability. The logical structure of the DeBERTA model is as follows: Figure 3 As shown

[0065] DeBERTa's multi-layer feature extraction is a progressive semantic encoding process "from local to global, from surface to deep". It captures complex semantic dependencies and temporal logic in text step by step by stacking multiple improved Transformer encoders and combining them with a decoupled attention mechanism. Specifically:

[0066] 1) Text embedding encoding

[0067] The semantic information of words can be represented by low-dimensional dense vectors, and the similarity between vectors can reflect the semantic relevance of words. Traditional embedding methods (such as Word2Vec) can only obtain the static semantics of words and cannot adapt to the context. DEBERTA's embedding layer, based on distributed representation theory, incorporates contextual dependence and relative position information to achieve "dynamic semantic embedding", which is more in line with the professional contextual needs of accident texts.

[0068] The input text is first broken down into word units. The DeBERTa model transforms the document's vocabulary into content vectors and relative position vectors. Content vectors Representing the semantics of the word itself, relative position vectors It indicates the positional relationship of the word relative to other words in the sequence.

[0069] The accident text embedding sequence integrates the fusion vector of all lexical units. This embedding sequence contains both the static professional semantics of the lexical units and the dynamic semantics of the context and relative positional information, laying the foundation for subsequent deep risk semantic mining and word order logic modeling.

[0070] 2) Deep risk semantic mining

[0071] The DeBERTa model extracts deep risk semantics through de-entangled attention. Firstly, it is based on content vector sequences. Embedded table with relative position The attention mechanism uses a "query-key-value" architecture, which is used to query the current lexical requirement, match related lexical terms, and output the features of related lexical terms, respectively:

[0072] , , , , ;

[0073] in, , , For content projection matrix; , This is the position projection matrix. , , These are content keys, content queries, and content value vectors, respectively, carrying only semantic information of word elements. , These are the position key and the position query vector, respectively, which only carry the relative position association information of the word elements.

[0074] The entanglement attention weights between two lexical units include content-content attention, content-position attention, and position-content attention. By adding and fusing these three attention types, the features of lexical units with high semantic relevance are preserved while also considering the rationality of word order logic, achieving risk feature focusing under the dual constraints of "semantics + logic". The attention weights of word pairs can be calculated using the following formula:

[0075] ;

[0076] ;

[0077] ;

[0078] in, This is the fused attention matrix; Content-content attention weight; the higher the weight, the higher the semantic relevance. Content-location attention weights associate semantic information with relative location information to capture the grammatical logic of the report text; The position-content attention weights invert relative position information and semantic information to further strengthen the grammatical logic of the report text. The attention matrix is ​​then normalized using softmax to ensure that the sum of the weights is 1, specifically:

[0079] ;

[0080] in, This is for the final unentanglement of attention weights. Based on the fused attention weights, the content value vector is... By performing a weighted summation, all lexical features highly relevant to the current word are integrated to obtain deep semantic features. Specifically:

[0081] ;

[0082] 3) Enhanced Mask Decoder (EMD) Semantic Restoration

[0083] The Enhanced Mask Decoder (EMD) is one of the core components that distinguishes DeBERTa from BERT

[46] , and it can meet the needs of identifying hidden risks in accident texts. EMD incorporates absolute position information into the decoding layer. The model decodes masked words based on the aggregated context embedding of word content and position. Through semantic enhancement and feature alignment, it strengthens the semantic restoration ability of hidden risk words and solves the problem that hidden risk words in event texts are easily masked. The Transformer Layer is used as EMD, and the decoding ability of the model is increased by increasing the number of network layers. Finally, the restored risk semantic feature sequence output by the Enhanced Mask Decoder is .

[0084] 3. BilSTM Risk Feature Fusion

[0085] The BiLSTM model solves the problem of risk logic breakage in long texts through a gating mechanism

[47] , focuses on the temporal correlation related to risk, and filters out irrelevant temporal information. It consists of a "forward LSTM" and a "reverse LSTM", and models through bidirectional contextual dependency. The restored feature sequence output by the DeBERTa mask decoder is used as input. The forward LSTM captures "risk inducement → failure phenomenon → accident consequence" in the forward word order of the text (from front to back), remembers the preceding risk information, forgets irrelevant information, and outputs the positive risk feature sequence HL;

[0086] The inverse LSTM captures the reverse logical connections of "accident consequences → failure phenomena → risk triggers" in reverse text order (from back to front), strengthening risk tracing and outputting the reverse risk feature sequence HR. The fusion of these two methods fully captures the bidirectional word order dependency in accident text, enhancing the coherence of risk logic and compensating for DeBERTa's shortcomings in capturing word order logic in long texts. The structure diagram of the BiLSTM model is shown below. Figure 4 As shown.

[0087] In BiLSTM, the forward LSTM and backward LSTM have the same structure and independent parameters. Both achieve the screening, storage, and output of risk information through the coordinated action of "forget gate, input gate, update cell state, and output gate". Taking the forward LSTM as an example, combined with the accident text scenario, the calculation of each part is as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] in, , , These are the outputs of the forget gate, input gate, and output gate, respectively. , , , These are the trainable weight matrices for the forget gate, input gate, candidate state, and output gate, respectively. , , , For bias terms; This involves concatenating the current word's restored features with the positive output features of the preceding word; It is the candidate lexical state; To update the final lexical state; To achieve element-wise multiplication, we can integrate previous risk information with current risk characteristics.

[0095] The core function of the forget gate is to control the forgetting ratio of preceding word states, accurately filtering redundant information in the incident text. For word information in the incident text that is irrelevant to risk, the forget gate outputs a value close to 0, thus forgetting it; for preceding information that is relevant to the current risk, the forget gate outputs a value close to 1, thus retaining it and reducing interference from irrelevant information.

[0096] The input gate can accurately capture the risk features of the current word. Working in conjunction with the forget gate, it filters out risk-related features in the current word, suppresses the input of irrelevant features, generates candidate states for the current word, provides high-quality input for updating the unit state, and ensures that the model only focuses on the core risk information of the current word.

[0097] The core function of a unit state is to store the risk context information of a lexical unit. Updating the unit state involves dynamically updating it using the outputs of the forget gate and the input gate, preserving prior risk information and incorporating current risk features to capture long-term risk dependencies in the incident text. Updating the unit state consists of two steps: first, generating candidate states, and then combining the outputs of the forget gate and the input gate to update the final unit state.

[0098] The core function of the output gate is to control the output ratio of the unit state, select the features most relevant to the current risk identification from the updated unit state, and output the positive (or negative) risk features of the current word, providing accurate input for subsequent bidirectional feature fusion and global feature aggregation.

[0099] The reverse LSTM has the same structure as the forward LSTM; the core difference lies in the direction of word order modeling (reverse LSTM is from back to front) and the trainable parameters (independent of the forward LSTM, avoiding parameter interference). The feature sequence output by the forward LSTM... Feature sequence of the inverse LSTM output By concatenating word-by-word meta-synthesis to achieve bidirectional feature fusion (as shown below), a more comprehensive risk feature sequence is formed, laying the foundation for subsequent output of key risk factors:

[0100] ;

[0101] in, To form a bidirectional fused feature sequence, it needs to be aggregated into a global feature vector through global pooling. Specifically:

[0102] ;

[0103] Finally, based on the global core risk characteristics By using a multi-label classifier, the corresponding probabilities of K preset risk factors are output, thereby achieving accurate identification and probability quantification of key risk factors in accident report texts.

[0104] 4. Risk Factor Output

[0105] Based on the global core risk feature sequence, multi-label classification is achieved through linear transformation and the Sigmoid activation function. The probability of K types of risk factors is output, and key risk factors are screened out, thus realizing the identification and probability quantification of key risk factors in accident report text.

[0106] ;

[0107] Where W is the trainable weight matrix of the classification layer. During the training process, the weights of the risk factor-related dimensions will be optimized to strengthen the mapping relationship between the global core risk features and risk categories, and ensure the accuracy of the basic probability output. For the classification layer bias term, K is the preset number of risk categories. It is a K-dimensional log-probability vector output by the classification layer, used for subsequent probability normalization.

[0108] By using the Sigmoid activation function, the log probabilities logits are normalized to the [0,1] interval to obtain the probability that each text belongs to each risk factor.

[0109] ;

[0110] in, This represents the basic probability of the existence of the Kth type of risk factor. The higher the probability, the more likely the risk factor is to exist. This is a sigmoid activation function, adapted for multi-label classification scenarios, ensuring that multiple risk factors can be identified simultaneously. When the threshold is set to 0.5, the sigmoid function can output multiple correct categories in multi-label classification.

[0111] Step 4: Based on the NK model, perform risk coupling analysis on the extracted risk factors, calculate the coupling value between different risk factors, and measure the degree of impact of risk coupling on the accident.

[0112] During high-speed rail operation, most malfunctions or incidents are not caused by a single risk; various risk factors interact and depend on each other.

[0113] like Figure 5 The diagram illustrates the risk coupling evolution mechanism of events during high-speed rail operation. A single risk breaching the defense system can lead to changes in local or overall safety attributes, resulting in the potential occurrence of an event. When different risk factors act synergistically, potential hazards within the system are triggered, thereby breaching the comprehensive defense system. Coupling modes include single-factor coupled risk, two-factor coupled risk, and multi-factor coupled risk. Single-factor coupled risk refers to the interaction between single risk factors affecting the safety of high-speed rail operation.

[0114] For example, if Figure 5 In this context, risk D represents the intrusion of external foreign objects. Therefore, when animals, humans, or other objects enter the track area, a coupled risk of risk D-risk D will occur. Two-factor coupled risk refers to the interaction between two types of risk factors affecting the safe operation of high-speed railways.

[0115] if Figure 5 In this context, risk A represents a failure risk factor in the air supply and braking system. When improper maintenance leads to this failure, the coupled risk of risk A leading to risk E (the improper maintenance risk factor) will occur. Similarly, multi-factor coupled risk refers to the interaction between three or more risk factors, such as risk A-risk B-risk C, risk A-risk B-risk C-risk D, or risk B-risk C-risk D-risk E. When coupled risks compromise the integrated defense system, under the strong coupling effect, even greater or new risks will emerge.

[0116] The NK model, proposed by Kauffman et al., is widely used to explore the interactions between various elements of a system and the impact of different combinations on the entire system. It can describe and understand the dynamic evolution and innovation processes in complex systems, and is particularly suitable for situations involving multiple interrelated factors.

[0117] This invention uses the NK model to explore the coupling relationship of various risk factors in unsafe events during high-speed train operation. N represents the number of risk factors; if each risk factor has... If there are several possible states, then the system will have The NK model identifies potential coupling scenarios. K represents the number of risk factors interacting and coupling in the system, with a maximum value of N-1 and a minimum value of 0. The NK model measures the impact of the coupling effect of two or more risk factors on an accident through the coupling value C between them. A larger C value indicates that these risk factors couple more frequently in some way, and are more likely to cause an accident. The risk coupling value is calculated as follows:

[0118] ;

[0119] in, Indicates risk factors The state is indicated by a value of 1, which means that there is a risk factor; and a value of 0, which means that there is no risk factor.

[0120] By using a risk factor coupling model, the degree of interaction between various risk factors can be calculated. The results can quantify the degree of danger brought about by the coupling of different risk factors, thus providing theoretical support for the analysis of high-speed train operation safety.

[0121] By deeply integrating DeBERTa with BiLSTM, an end-to-end risk factor identification process was constructed. Its core principle is as follows: First, by utilizing DeBERTa's deentanglement attention mechanism and enhanced mask decoder, words in the accident text are simultaneously encoded into content vectors and relative position vectors. Through the fusion of three types of attention—content-content, content-position, and position-content—the semantic relationships are preserved while accurately capturing the logical order of words, thereby extracting deep risk semantic features.

[0122] The feature sequence output by DeBERTa is then input into a BiLSTM. A forward LSTM captures forward dependencies in the order of "risk trigger → failure phenomenon → accident consequence," while a backward LSTM reinforces risk source tracing in reverse order. The two sequences are concatenated and then global pooled to obtain a global feature vector. Finally, a multi-label classifier outputs the probability of each risk factor. This principle effectively solves the problem that a single model cannot simultaneously handle deep semantics and bidirectional temporal dependencies, and can also restore hidden risk terms that have been masked.

[0123] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for identifying operational safety risk factors in a high-speed train system, characterized in that, Includes the following steps: Step S1: Obtain historical event reports and fault handling reports of high-speed trains as the raw dataset; Step S2: Perform text preprocessing on the original dataset obtained in step S1 to obtain preprocessed text; Step S3: Extract risk factors from the preprocessed text based on the DeBERTa-BiLSTM model and output the probability of each risk factor; Step S4: Based on the NK model, perform risk coupling analysis on the extracted risk factors, calculate the coupling value between different risk factors to measure the degree of impact of risk coupling on the accident; the formula for calculating the risk coupling value using the NK model is: ； In the above formula, This represents the coupling value between two or more risk factors. This indicates the status of each risk factor, with a value of 1 indicating the presence of the risk factor and a value of 0 indicating the absence of the risk factor. This represents the joint probability of all risk factors simultaneously being in a specific combination of states. These represent the marginal probabilities of each risk factor being independently in a certain state; Step S5: Based on the coupling value calculated in step S4, risk factor combinations with coupling values ​​greater than a preset coupling threshold are identified as high-risk coupling modes, and the high-risk coupling modes are output for use in the formulation of risk warning and prevention and control strategies for high-speed train operation safety.

2. The method for identifying safety risk factors in high-speed train system operation according to claim 1, characterized in that: The text preprocessing in step S2 includes: filtering and deleting stop words, terminology standardization, establishing risk labels, and data annotation of the text.

3. The method for identifying safety risk factors in high-speed train systems according to claim 2, characterized in that: The risk labels are established as primary labels according to the "human-machine-environment-management" framework, and secondary labels are established for each of them. The data labeling adopts a combination of machine labeling and manual labeling to label each event into one or more corresponding risk categories.

4. The method for identifying safety risk factors in high-speed train systems according to claim 1, characterized in that: Step S3, the extraction of risk factors based on the DeBERTa-BiLSTM model, specifically includes: Step S301: The preprocessed text is embedded and encoded using the DeBERTa model. Deep semantic features are extracted through the deentanglement attention mechanism, and the restored risk semantic feature sequence is output through the enhanced mask decoder. Step S302: Input the risk semantic feature sequence into the BiLSTM model, extract bidirectional temporal dependency features through forward LSTM and backward LSTM respectively, and concatenate the forward output features and backward output features to obtain a bidirectional fused feature sequence. Step S303: Perform global pooling on the bidirectional fused feature sequence to obtain a global feature vector, and then output the probability of each risk factor through a multi-label classifier.

5. The method for identifying safety risk factors in high-speed train systems according to claim 4, characterized in that: The embedding encoding of the DeBERTa model includes: converting text words into content vectors and relative position vectors; The content vector represents the semantics of the word itself, and the relative position vector represents the positional relationship of the text word relative to other words in the risk semantic feature sequence.

6. The method for identifying safety risk factors in high-speed train system operation according to claim 4, characterized in that: When calculating the attention weights between word pairs, the de-entanglement attention mechanism includes content-content attention, content-position attention, and position-content attention, and then adds and fuses these three types of attention.

7. The method for identifying operational safety risk factors of high-speed train systems according to claim 1, characterized in that: The risk coupling analysis includes the analysis of single-factor coupling risk, two-factor coupling risk, and multi-factor coupling risk.

8. The method for identifying operational safety risk factors of high-speed train systems according to claim 1, characterized in that: The historical event reports and fault handling reports are unstructured text data.

Images