Railway emergency auxiliary decision-making method and system

Abstract

The application discloses a railway emergency auxiliary decision-making method and system, wherein the method comprises the following steps: collecting railway accident reports and constructing a railway emergency case library; pre-training a Rail-BERT model by using a mask language model based on the case library; fine-tuning the Rail-BERT model based on event extraction, event relation classification and semantic role labeling tasks; aligning the event extraction, event relation classification and semantic role labeling results, thereby constructing a railway event graph; injecting railway event graph information into a Transformer encoder layer of the Rail-BERT model through a knowledge weight matrix; performing similarity calculation on semantic vectors corresponding to a current emergency event and historical events generated by the Transformer encoder layer, thereby obtaining K historical events with the highest similarity; and performing emergency auxiliary decision-making based on the disposal schemes of the K historical events with the highest similarity. The application can improve the emergency decision-making effect of railway emergencies.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of railway transportation safety and intelligent decision-making, in particular to a railway emergency auxiliary decision-making method and system. BACKGROUND

[0002] The railway transportation system is large in scale and technology-intensive, involving the coordination of multiple departments such as train service, locomotive service, rolling stock, signal service, and passenger transportation. With the development of high-speed, heavy-load, and high-altitude railways, the complexity of emergency rescue in the event of accidents has further increased. Railway accident report historical data is an important auxiliary tool for railway emergency decision-making and has strong reference value. However, how to improve the causal logic mining and dynamic matching capability through effective methods remains a core challenge. Traditional case-based reasoning methods rely on keyword surface matching and are difficult to capture the deep causal logic in accident reports. The precision of similar case matching is insufficient, resulting in a decrease in the reference value of historical events.

[0003] Previous research on railway accident report semantic model construction and case base construction often does not focus on the causal logic of events. Using cosine similarity method for matching can only capture surface semantic similarity and ignore the dynamic evolution law between events, resulting in a lack of explainability and logical consistency in the search results required for emergency decision-making. SUMMARY

[0004] The purpose of the present application is to provide a railway emergency auxiliary decision-making method and system that can improve the effectiveness of railway emergency decision-making in the event of emergencies.

[0005] In a first aspect, the present application provides a railway emergency auxiliary decision-making method, comprising:

[0006] Step 1: Collect railway accident reports and preprocess the collected railway accident reports to obtain a structured representation of the railway accident reports. Based on the text sequence of the railway accident reports, a railway emergency case base is constructed.

[0007] Step 2: Based on the railway emergency case base, a Rail-BERT model dedicated to the railway field is pre-trained using a masked language model.

[0008] Step 3: Fine-tune the Rail-BERT model based on downstream tasks. The downstream tasks include event extraction, event relation classification, and semantic role labeling.

[0009] Align the event extraction, event relation classification, and semantic role labeling results to construct a railway event graph. The railway event graph contains a node set and an edge set with relationship types, where the nodes represent events and the edges represent the relationships between events.

[0010] Step 4: Injecting railway affair graph information into the Rail-BERT model through the knowledge weight matrix to the Transformer encoder layer;

[0011] Step 5: Combining the semantic vectors corresponding to the current emergency event and historical events generated based on the Transformer encoder layer, performing similarity matching between the current emergency event and the historical events in the railway affair graph, and obtaining K historical events with the highest similarity; based on the disposal scheme of the K historical events with the highest similarity, making an emergency auxiliary decision.

[0012] In one possible implementation, in the event relationship classification in step 2, a multi-dimensional confidence model is constructed to quantify the confidence of the relationship triple {subject event, relationship type, object event} in the railway field knowledge; the multi-dimensional confidence model is represented as:

[0013] ;

[0014] ;

[0015] ;

[0016] ;

[0017] In the formula, denotes the normalized confidence, , and are weight coefficients, is the syntactic evidence strength, is the semantic similarity, is the co-occurrence frequency; denotes the semantic association tightness of the subject event and the object event; is the attenuation coefficient; denotes the causal predicate weight; denotes and the cosine similarity, denotes the semantic vector corresponding to the subject event, denotes the semantic vector corresponding to the object event; and are obtained based on the pre-trained Rail-BERT model; denotes the time attenuation factor, is the time corresponding to the subject event, is the time corresponding to the object event, is a preset time threshold; is an indicator function, and if the event pair In the document The CCP now appears, then ,otherwise ; Represents the spatial distance between two events; weighting coefficient .

[0018] In one possible implementation, the knowledge weight matrix is ​​calculated in step 4 as follows:

[0019] The input text sequence is embedded and positionally encoded to obtain a position-enhanced representation of the text sequence. ;

[0020] The node and edge embeddings of the railway logic graph are learned using a relational graph convolutional network, resulting in the node embedding matrix of the railway logic graph. Based on node embedding matrix Generate edge incidence matrix ;

[0021] By mapping events in a text sequence to nodes in a railway event graph through entity links, an alignment matrix is ​​generated. ;

[0022] Generate a knowledge weight matrix: .

[0023] In one possible implementation, step 4 involves injecting railway reasoning graph information into the Transformer encoder layer of the Rail-BERT model using a knowledge weight matrix, including:

[0024] Knowledge weight matrix The attention weights of each encoder layer in the Transformer model are calculated using the following formula:

[0025] ;

[0026] In the formula, Indicates the first Attention weights calculated from each attention head; This represents the knowledge fusion intensity coefficient. It is the dimension of attention head; , and The first The query matrix, key matrix, and value matrix of each attention head.

[0027] In one possible implementation, step 5 involves matching the current emergency with historical events in the railway event map, including: performing similarity matching based on three-level fusion similarity; the three-level fusion similarity includes event-level similarity, path-level similarity, and spatiotemporal similarity.

[0028] In a possible implementation, the event-level similarity calculation formula is:

[0029] ;

[0030] In the formula, denotes the event-level similarity of the current event node and the historical event node , denotes the semantic vector of the node and the node , which is generated based on a Transformer encoder layer; denotes the calculation of the cosine similarity; denotes the event type weight, if and belong to the same type of event, then , otherwise, 0; denotes the weight coefficient, which is set to 0.7 in order to strengthen the semantic matching.

[0031] In a possible implementation, the path-level similarity calculation formula is:

[0032] ;

[0033] In the formula, denotes the path-level similarity of the current event and the historical event, and denote the causal path of the current event and the historical event, respectively; denotes the graph structure edit distance, that is, the minimum operation cost based on node addition, deletion, and edge relationship adjustment; denotes the confidence weight of the relationship edge ; denotes the path length, that is, the number of nodes.

[0034] In a second aspect, the present application provides a railway emergency auxiliary decision-making system, comprising a memory and a processor.

[0035] The memory is configured to store a computer program.

[0036] The processor is configured to invoke the computer program to execute the method described above.

[0037] In a third aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program, when running on an electronic device, causes the electronic device to implement the method described above.

[0038] In a fourth aspect, the present application provides a computer program product comprising a computer program which, when running on an electronic device, causes the electronic device to implement the method described above.

[0039] The specific implementation manners of the above-mentioned second to fourth aspects of the present application can refer to the implementation manners of the above-mentioned first aspect, and will not be described here.

[0040] Beneficial effects:

[0041] The present application improves the semantic understanding precision in the railway field by the Rail-BERT model pre-trained by the term weight mask strategy. In the matter graph construction, the reliability of the logical relationship is quantified based on the multi-dimensional confidence model, and the explainability of the decision logic is enhanced; the knowledge-guided attention mechanism is used to dynamically fuse the causal logic of the railway matter graph in the Transformer encoder layer, the node embedding and the relationship edge weight are learned using the graph neural network, and the knowledge weight matrix is generated, so that the model focuses on the semantic association and causal path relationship related to the domain logic. The similarity matching of the current event and the historical case is carried out based on the three-level fusion of the event level, the path level and the space-time level, and the disposal scheme of the similar event is obtained to assist the decision, which improves the matching accuracy and further improves the emergency decision efficiency. The technical scheme of the present application realizes the technical closed loop of semantic understanding-logical modeling-dynamic matching to assist the decision of the railway emergency event, which can significantly improve the scientific nature and timeliness of the emergency response of the railway emergency event. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The method flowchart of an embodiment of the present application is shown in the figure;

[0043] Figure 2 The system structure diagram of an embodiment of the present application is shown in the figure;

[0044] Figure 3 The Transformer encoder layer schematic diagram of an embodiment of the present application injected with the knowledge-guided attention mechanism is shown in the figure;

[0045] Figure 4 The performance comparison schematic diagram of the event extraction of multiple models in an embodiment of the present application is shown in the figure;

[0046] Figure 5 The performance comparison schematic diagram of the event relation classification of multiple models in an embodiment of the present application is shown in the figure;

[0047] Figure 6 The performance comparison schematic diagram of the semantic role labeling of multiple models in an embodiment of the present application is shown in the figure;

[0048] Figure 7 The confusion matrix diagram of the Rail-BERT model in an embodiment of the present application is shown in the figure;

[0049] Figure 8 Confusion matrix result diagram of the BERT model in one of the examples of the present application;

[0050] Figure 9 Confusion matrix result diagram of the BiLSTM+Attention model in one of the examples of the present application;

[0051] Figure 10 Confusion matrix result diagram of the Bio-BERT model in one of the examples of the present application. DETAILED DESCRIPTION

[0052] In order to enable personnel in the technical field to better understand the scheme of the present application, the technical scheme of the present application will be further described in detail below in combination with the examples of the present application and the accompanying drawings.

[0053] The present application discloses a railway emergency auxiliary decision-making method and system. The method comprises the following steps: based on a railway accident report, a professional term library and a stop word library, a BERT model (referred to as Rail-BERT model) dedicated to the railway field is pre-trained through a term weight mask strategy to improve the semantic understanding ability of railway terms; the Rail-BERT model is used to extract event triples and their causal, chronological, conditional or concurrent relationships, and a multi-dimensional confidence model is constructed by combining syntactic evidence, semantic similarity and co-occurrence frequency, so as to construct a railway event graph; the knowledge weight matrix (causal logic weight) of the railway event graph is dynamically injected into the Transformer encoder layer, the causal logic of the railway event graph is dynamically fused in the Transformer encoder layer, the graph neural network is used to learn node embedding and relationship edge weight, and the knowledge weight matrix is generated, so that the model focuses on the semantic association and causal path relationship related to the field logic, and the semantic representation is realized through the knowledge guided attention mechanism; a historical event matching mechanism with three-level fusion of event-level semantic similarity, path-level topological similarity and space-time level evolution similarity is designed; a Top-K (K most similar) emergency disposal scheme is output based on the similarity score, and end-to-end response within 1 second is supported.

[0054] The present application solves the problem of ignoring logic in traditional case matching, provides a new idea for railway accident report similarity matching and causal logic modeling, and aims to solve the problem of ignoring event evolution causal logic in traditional case matching. The present application includes railway emergency case library construction, historical event matching driven by causal path and emergency disposal scheme generation, and can realize a closed-loop railway emergency event auxiliary decision-making of semantic analysis-logic modeling-causal path matching-decision output, and can provide an intelligent solution for railway emergency management. The present application realizes precise decision-making with high comprehensive similarity, and significantly improves the scientificity and timeliness of railway emergency response.

[0055] A specific embodiment according to the present application will be described below with reference to the accompanying drawings.

[0056] Embodiment one:

[0057] The embodiment of the present application provides a railway emergency auxiliary decision-making method, comprising:

[0058] Step 1: Collect railway accident reports, and pre-process the collected railway accident reports to obtain a text sequence of the railway accident reports, and construct a railway emergency case base based on the text sequence of the railway accident reports;

[0059] Further, the railway accident reports are collected, including: collecting a plurality of Chinese and English railway accident reports.

[0060] Further, the collected railway accident reports are pre-processed to obtain a text sequence of the railway accident reports, and a railway emergency case base (structured representation) is constructed based on the text sequence of the railway accident reports, including: based on the plurality of Chinese and English railway accident reports, a high-frequency professional vocabulary is screened by using a method combining TF-IDF (Term Frequency-Inverse Document Frequency) and mutual information, and a railway professional bilingual terminology library is constructed based on the screened high-frequency professional vocabulary; a railway special stop word library is constructed, which is used to remove part of useless words in the Chinese and English railway accident reports; the collected railway accident reports are segmented based on a segmentation tool to obtain a preliminary segmentation result, the preliminary segmentation result is matched with the railway professional bilingual terminology library, and the matched segmentation result is retained; based on the railway special stop word library, the stop words in the retained segmentation result are filtered to obtain a structured representation of the railway accident reports.

[0061] The scheme optimizes the segmentation processing result based on the construction of the railway professional bilingual terminology library and the railway special stop word library.

[0062] In some embodiments, the segmentation tool uses a Jieba segmentation tool.

[0063] The Jieba segmentation tool is one of the most commonly used open source segmentation tools in Chinese natural language processing (NLP), and is mainly used to divide continuous Chinese text into meaningful words (Token). In the pre-processing of the railway accident reports, the segmentation of the collected railway accident reports based on the Jieba segmentation tool means that the default segmentation mode of the Jieba segmentation tool is used to preliminarily divide the Chinese and English railway accident report texts.

[0064] In some embodiments, 220,000 Chinese and English railway accident reports are collected, and a railway professional bilingual terminology library is constructed based on the 220,000 Chinese and English railway accident reports; the railway professional bilingual terminology library includes 19,714 Chinese-English contrastive terms, covering 12 secondary subfields.

[0065] Step 2: Based on the railway emergency case library, a BERT model dedicated to the railway field (referred to as Rail-BERT model) is pre-trained using a masked language model (MLM).

[0066] Further, in the process of pre-training the Rail-BERT model using the masked language model, a term weight masking strategy is used to enhance the importance of professional terms in random masking.

[0067] The term weight masking strategy refers to strengthening the learning of railway field professional terms by differentiating the masking probability of the model. The core idea is to increase the masking probability of professional terms and reduce the masking probability of general terms. Masking probability refers to the probability of randomly covering (masking) words during pre-training, which aims to force the model to predict the covered words through context, thereby learning word semantics and enhancing term semantic disambiguation ability.

[0068] For example, the masking probability is set to 35%, and the other words are set to 15%, expressed as:

[0069] ;

[0070] wherein, represents the masking probability of the i-th word of the input model;

[0071] Further, for cross-language problems, alignment masking is performed on Chinese-English parallel sentences, that is, corresponding terms (such as “signal failure” and “signal failure”) in Chinese-English parallel sentences are masked simultaneously to ensure synchronous learning of cross-language terms.

[0072] In addition, the proportion of masking (MASK) in MLM is 15%, and after selecting 15% of the tokens, 80% are replaced with [MASK], 10% remain unchanged, and 10% are replaced with random words (another random token). This masking strategy can improve the robustness of the model. The task of MLM is to predict all selected 15% of the tokens, and gradually improve the accuracy of railway professional term prediction.

[0073] ​​​Further, in the pre-training process of the Rail-BERT model, an entity boundary detection task is added on the basis of the MLM patrol, and through Entity-Aware Training, the ability of the model to identify the boundary of the railway entity is improved, and the problem of identifying composite terms is solved.

[0074] The entity boundary detection task, i.e., labeling the boundary label of each word as: B (beginning), I (interior), E (end), O (outside), S (single), if it is a composite term, the whole word is labeled (such as "contact net icing" → B / I / I / I / E), and the letter-number combination is also regarded as a whole for labeling (such as "TEDS-3000" → S).

[0075] Therefore, the tasks of the Rail-BERT model in the pre-training stage include predicting the original word of the [MASK] position and entity boundary detection.

[0076] Step 3: fine-tuning the Rail-BERT model based on downstream tasks; wherein the downstream tasks include event extraction, event relation classification and semantic role labeling; aligning the event extraction, event relation classification and semantic role labeling results to construct a railway event graph;

[0077] Event extraction, event relation classification and semantic role labeling are also key steps for constructing a railway event graph.

[0078] The following will specifically explain this step.

[0079] 1): event extraction; the event extraction task elements include event trigger words, event types, argument extraction and semantic role labeling, identifying the core verbs or nouns directly representing the occurrence of events in the text as trigger words, assigning and defining event categories for each trigger word and outputting labels, and then extracting event-related arguments such as subjects, objects, times, places, understanding the overall process of the event, and semantic role labeling aims to determine the semantic role of arguments (elements) in the event, such as "cause of the accident" and "direct consequences", and the role system includes inducement, agent, recipient, disposal measures and consequences.

[0080] 2): event relation classification; event relation classification aims to identify the logical relationship between events, and defines four core event relations according to the characteristics of railway emergencies: cause-effect relationship (Cause), time sequence relationship (Follow), conditional relationship (Condition), and concurrent relationship (Co-occur). The extraction process adopts a double-track fusion strategy, including explicit relationship extraction and implicit relationship extraction. Explicit relationship extraction is based on language rules, including 20 words such as "cause", "trigger", and "result in", combined with dependency syntax analysis to locate event subjects and objects, and inject temporal and spatial constraints to parse event expressions and extract geographic locations to associate event attributes; implicit relationship extraction relies on the pre-trained Rail-BERT model, which vectorizes the event representation, integrates temporal and spatial features, and document-level co-occurrence features, and uses a relationship classifier to infer potential logical associations.

[0081] A multi-dimensional confidence model is constructed to quantify the confidence of the relationship triple {subject event, relationship type, object event} in the railway field knowledge, by weighted integration of syntactic evidence strength, semantic similarity and co-occurrence frequency, to form a standardized confidence (relationship weight) of 0~1.0. The multi-dimensional confidence model is represented as:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] In the formula, represents the standardized confidence, , and are weight coefficients, is the syntactic evidence strength, is the semantic similarity, is the co-occurrence frequency;

[0087] represents the semantic association tightness of the subject event and the object event; represents the cause-effect predicate weight, which is used to quantify the strength of the cause-effect relationship word, and is a constant based on language rules (such as "cause" weight = 1.0, "may trigger" weight = 0.6), is a decay coefficient (e.g., 0.5). The semantic correlation tightness can be quantified according to the dependency tree. The dependency tree is a tree-shaped graph in natural language processing to represent the syntax structure of a sentence, where the nodes are words (e.g., verbs, nouns), and the edges are the syntactic relationships between words (e.g., subject-verb, verb-object). The semantic correlation tightness can be quantified by the length of the shortest path between the two events in the dependency tree (e.g., the path length of "contact network-icing → cause → delay" is 3).

[0088] represents and the cosine similarity of represents the semantic vector corresponding to the subject event, represents the semantic vector corresponding to the object event; and is obtained based on the pre-trained Rail-BERT model, specifically: the position-enhanced representation of the text sequence corresponding to the event is input into the Transformer model in the Rail-BERT model, and the output matrix of the last encoder layer of the Transformer model is obtained. The vector corresponding to the event classification mark [CLS] or the event trigger in the output matrix can be used as the semantic vector corresponding to the event; where [CLS] (Classification) is the first special mark of the text sequence corresponding to the event, and the vector corresponding to [CLS] is used for overall sequence representation, i.e., event-level representation. The vector corresponding to [CLS] can be represented as , where is the output matrix of the last encoder layer, i.e., the layer of the Transformer model, is the dimension of the hidden state of the Transformer model. The vector corresponding to the event trigger is the vector corresponding to the position of the trigger in the input matrix, which is used for fine-grained event representation, and can be represented as , where is the position index of the event trigger in the text sequence corresponding to the time;

[0089] represents a time decay factor, is the time corresponding to the subject event, is the time corresponding to the object event, is a preset time threshold (e.g., = 72 hours);

[0090] is an indicator function, which is used to quantify the co-occurrence relationship of the event pair in the document . If the event pair co-occurs in the document , then , else . denotes the spatial distance (unit: km) between two events;

[0091] weight coefficient , which is set to 0.5 according to experience and can be verified according to historical data , , .

[0092] Finally, the structured event triple <event trigger word, event type, {argument 1: role 1, argument 2: role 2,...}> is output, which provides a basic unit for subsequent construction of nodes and edges in the railway event graph.

[0093] 3): Event alignment, align the events and their relationships identified and extracted above to a specific knowledge framework through knowledge fusion, and thus construct a complete railway event graph (Event Graph). Event alignment includes three aspects of time sequence chain alignment, spatial topology alignment and causal logic alignment. Time sequence chain alignment aims to solve the problem of time logic conflict of event occurrence, that is, the cause event must be earlier than the result event; spatial topology alignment aims to verify the feasibility of event propagation in space, filter the relationships that violate the spatial decay law, propose invalid spatial association, and retain topologically reachable relationships; causal logic alignment aims to complete the logic chain and ensure that the causal relationship complies with the physical law and industry standard.

[0094] Step 4: Inject railway event graph information into the Transformer encoder layer of the Rail-BERT model through the knowledge weight matrix.

[0095] This step injects knowledge-guided attention mechanism, dynamically fuses the causal logic of the railway event graph in the Transformer encoder layer, uses graph neural network to learn node embedding and relationship edge weight, and generates a knowledge weight matrix, so that the model focuses on semantic association and causal path relationship related to domain logic.

[0096] The Transformer model includes four parts: input part (input-output embedding and position encoding), multi-layer encoder, multi-layer decoder and output part (output linear layer and Softmax).

[0097] This step specifically includes:

[0098] Step 4.1: Embedding and position encoding of the input text sequence to obtain the position-enhanced representation of the text sequence .

[0099] The original input vector needs to be injected with position information and domain knowledge to make up for the defects of the permutation invariance of the Transformer model.

[0100] First, the input text sequence is mapped into a dense vector for each word in the text sequence . The dimension of the dense vector is , i.e. ; is the length of the text sequence (the number of words in the text sequence); the embedding matrix of the entire text sequence is composed of . In some embodiments, may take a value of 768.

[0101] For the position of the word and the dimension , a sine-cosine function is used to generate a position vector:

[0102] ;

[0103] ;

[0104] wherein , represents the position vector of the corresponding dimension, which is the sine component; is the position vector of the corresponding dimension, which is the cosine component; the parameter is used to realize frequency attenuation.

[0105] and are taken as the elements of the position matrix in the row, the column and the column, respectively, to obtain the position matrix .

[0106] The position-enhanced representation of the text sequence is calculated as follows:

[0107] ;

[0108] After that, the knowledge representation is added to vectorize the railway affair graph.

[0109] ​​​The railway event graph is a typical directed heterogeneous graph. A Relational Graph Convolutional Network (R-GCN) is used to learn the node (event) and edge (relation) embeddings of the railway event graph. Each layer of R-GCN is a Relational Graph Convolutional Network layer, represented as... This is used to iteratively update the node embedding matrix, and the calculation formula is:

[0110] ;

[0111] in, For R-GCN The node embedding matrix of the layer, the total number of layers in R-GCN is , , For R-GCN Layer node embedding matrix, ; The railway logic graph structure consists of a set of nodes and a set of edges with relational types. It is the total number of nodes in the railway logic diagram;

[0112] in, ,Right now It is the embedding vector of the node. ~ A matrix formed by piecing together nodes in order.

[0113] For the nodes of the railway logic diagram The update formula for its embedding vector (hidden state) is:

[0114] ;

[0115] in, Represents a node In R-GCN Layer embedding vectors, Represents a node In R-GCN Layer embedding vectors, A set of relation types; ={causal relationship, temporal relationship, conditional relationship, concurrency relationship} For nodes In relation The set of neighboring nodes below, ; The normalization constant is usually taken as . , Represents a set the number of nodes contained in the middle; for the relationship the transformation matrix specific to the node, for the self-connection transformation matrix of the R-GCN first layer, for the activation function, for the neighbor nodes of the node the embedding vector of the node in the R-GCN first layer. For each node its initial embedding vector is obtained based on the pre-trained Rail-BERT model. For example, the node the position enhanced representation of the corresponding text sequence can be input into the Transformer model in the Rail-BERT model to obtain the output matrix of the last encoder layer of the Transformer model, and the vector corresponding to the event classification mark [CLS] or the event trigger word in the output matrix can be taken as the semantic corresponding to the event as If there is a new node in the railway affair graph that does not appear in the training data, the random initialization is adopted:

[0116] ;

[0117] wherein, represents a normal distribution (Gaussian distribution) with a mean of 0 and a variance of

[0118] Finally, the node embedding matrix of the railway affair graph is obtained by iteration , and the edge association matrix is generated based on the node embedding matrix :

[0119] ;

[0120] ;

[0121] The embedding vector of the edge is represented as:

[0122] ;

[0123] ;

[0124] wherein, represents the embedding vector of the edge between the node and the node ; represents the embedding vector of the node , and the embedding vector of the node is​​​​​ the first row of ; the embedding vector of node , the first row of ; and respectively represent the weight matrix of the first layer of fully connected layers and the second layer of fully connected layers, and represent the bias term; is an intermediate variable. is the embedding of the relationship type between node and node , the value of is realized by a trainable relationship embedding matrix. Four kinds of relationships are respectively assigned a trainable low-dimensional dense vector: ; is the relationship embedding dimension, which can be set to . Xavier uniform distribution is used for initialization:

[0125] ;

[0126] indicates that obeys the Xavier uniform distribution, the layer input dimension is , and the layer output dimension is .

[0127] Xavier uniform distribution is used for initialization to ensure stable gradient propagation at the beginning of training.

[0128] By entity linking, the events in the text sequence are mapped to the nodes of the railway affair graph, and the alignment matrix is generated.

[0129] First, the set of event mentions in the text sequence is identified:

[0130] ;

[0131] Among them, each event mention corresponds to a continuous word sequence in the input text, which is composed of trigger words and its arguments.

[0132] For each event mention and node in the railway affair graph, the semantic similarity is calculated:

[0133] ;

[0134] wherein, represents and semantic similarity; represents and cosine similarity, is the corresponding semantic vector, which can be obtained based on the pre-trained Rail-BERT model, such as and the position enhanced representation of the corresponding text sequence is input into the Transformer model of the Rail-BERT model, and the output matrix of the last encoder layer of the Transformer model is obtained, and the vector corresponding to the event classification mark [CLS] or the event trigger word in the output matrix can be taken as the semantic of the event; represents the event mention, represents the node , is the semantic vector of , is the embedding vector of the node .

[0135] For each word in the input text, the association probability thereof with the node is calculated:

[0136] ;

[0137] wherein, represents the association probability of , that is, the association strength between the word in the input text and the node in the railway affair graph, ; represents the probability that the word belongs to the node in the railway affair graph; represents the word in the input text, represents that the word belongs to the event mention , is the total number of nodes in the railway affair graph.

[0138] Finally, the alignment matrix is generated:

[0139] ; wherein N is the number of nodes in the railway affair graph.

[0140] The association probability of each word is summed over all events to be 1.

[0141] Generate the knowledge weight matrix:

[0142] ;

[0143] This matrix quantifies the association strength between any two words in the text based on the domain logic, for calculating the association strength of each word with all events, represents the mapping of the logical association of the event space back to the word space. The knowledge weight is added to the multi-head self-attention mechanism in the subsequent steps.

[0144] Step 4.2: In the Transformer encoder layer, inject the knowledge weight matrix to guide the attention weight correction;

[0145] For the 1st encoder layer: , ;

[0146] The encoder layer: , ;

[0147] wherein, is the total number of encoder layers of the Transformer model, represents the output of the encoder layer of the Transformer model, represents the Transformer model encoder layer operation, the input of the 1st encoder layer , the encoder layer is the output of the previous encoder layer; the final output is ;

[0148] In some embodiments, = 12.

[0149] In the subsequent steps, the semantic vector of the node directly comes from the hidden state of the corresponding position in , and the relationship between events is identified through the classification head of the Rail-BERT model . At the same time, it lays the foundation for the subsequent three-level similarity calculation.

[0150] Each encoder layer operation is as follows:

[0151] Step 4.2.1: First, linear projection and split head are performed on the input are projected to one independent attention head:

[0152] ;

[0153] wherein, is the attention head number; , and are the query, key and value projection matrices of the th attention head, and is the learnable parameter; is the dimension of the attention head; , and are the query matrix, key matrix and value matrix of the th attention head, respectively.

[0154] The knowledge weight matrix is then integrated into the scaled dot-product attention score calculation, and the attention weight is calculated independently for each attention head:

[0155] ;

[0156] wherein, represents the attention weight calculated by the th attention head; is the knowledge fusion strength coefficient, and the default value is 0.5; is used to inject the causal logic weight of the railway affairs graph The role of is to prevent gradient explosion; obtains a similarity matrix, wherein each element in the matrix represents the dot product similarity between the query vector of position and the key vector of position , which is used to measure the “attention degree” of position to position .

[0157] is the output of the th attention head, i.e. ; ;

[0158] Finally, the outputs of each attention head are spliced and fused, and linear transformation is performed:

[0159] ;

[0160] wherein, is the output of the multi-head self-attention layer; is the linear projection weight matrix, is a learnable parameter, used to map the multi-head output after concatenation back to the original dimension space.

[0161] Step 4.2.2: Residual connection and layer normalization.

[0162] The input is added to the output, alleviating the vanishing gradient problem:

[0163] ;

[0164] where, is the output after residual connection;

[0165] Normalization along the feature dimension:

[0166] ;

[0167] where, the layer normalization output, is the mean and standard deviation of the features, and the matrix is the mean of the first row features, and the standard deviation is , and the calculation formula is:

[0168] ;

[0169] ;

[0170] Only here represents the elements within the matrix . Layer normalization independently standardizes all feature values of each word vector in the sequence, ensuring model training stability and providing normalized input for subsequent FFN layers; is a learnable scaling and bias parameter, , .

[0171] Step 4.2.3: Two linear transformations and activations on the normalized vector:

[0172] ;

[0173] where, and is the linear projection weight matrix, , is a learnable parameter; is the intermediate dimension, usually taking the value 2048, is the activation function, and is the bias term, ​​Position-wise FFN.

[0174] Step 4.2.4: Residual connection and layer normalization are performed again:

[0175] ;

[0176] ;

[0177] wherein, and are vectors obtained after residual connection and layer normalization are performed again; is the semantic vector corresponding to the input generated by the Transformer encoder layer;

[0178] Step 5: Combine the semantic vectors corresponding to the current emergency event and historical events (cases) generated based on the Transformer encoder layer, match the current emergency event with historical events in the railway affair graph, and obtain K historical events with the highest similarity; make an emergency auxiliary decision based on the disposal scheme of the K historical events with the highest similarity;

[0179] wherein, the value of K can be set as needed.

[0180] To break through the limitations of traditional case matching methods in dynamic evolving events, the embodiment of the present application proposes a causal path driven matching mechanism based on the evolution path of the railway affair graph. The mechanism first parses the description of the current emergency event into a structured causal path (causal chain), deeply understands the event semantics and their logical associations through the Rail-BERT model; then, searches for a historical evolution path with homologous head and tail events in the railway affair graph, and calculates the similarity using a three-level fusion strategy, including event level, path level and spatiotemporal level similarity.

[0181] In response to the user inputting the current emergency event in the case library, the current emergency event input by the user is parsed into an initial causal path:

[0182] ;

[0183] wherein, is a node, is a relationship edge between nodes, including weight and spatiotemporal attributes.

[0184] In some embodiments, this step includes extracting event triples from the current emergency event input by the user based on the Rail-BERT model, and parsing the current emergency event input by the user into an initial causal path based on the time triples;

[0185] wherein the similarity adopts a three-level fusion similarity, and the calculation formula is:

[0186] ;

[0187] wherein, denotes the three-level fusion similarity, denotes the event-level similarity, denotes the path-level similarity, denotes the spatio-temporal similarity; and parameters are set as .

[0188] The three-level fusion similarity calculates the similarity from the event level, the path level and the spatio-temporal level, and obtains the comprehensive similarity through fusion calculation.

[0189] wherein the calculation formula of the event-level similarity is defined as:

[0190] ;

[0191] wherein, denotes the event-level similarity of the current event node and the historical event node , denotes the semantic vector of the node and the node , which is taken from the hidden state of the corresponding position and quantifies the event semantic matching degree through the calculation of the cosine similarity; denotes the event type weight, if and belong to the same type of event, then , otherwise 0; denotes the weight coefficient, which is set as 0.7 in order to strengthen the semantic matching. wherein the calculation formula of the path-level similarity is defined as:

[0192]

[0193] ; wherein,

[0194] denotes the path-level similarity of the current event and the historical event, and denote the causal path of the current event and the historical event, respectively; denotes the graph structure edit distance, which is the minimum operation cost based on the addition, deletion and modification of nodes and the adjustment of edge relationships; denotes the confidence weight of the relationship edge ; denotes the path length, i.e. the number of nodes.

[0195] ​The calculation formula of the spatiotemporal level similarity is defined as:

[0196]

[0197]

[0198]

[0199]

[0200] In the formula, represents the spatiotemporal level similarity of the current event and the historical event, represents the time similarity, represents the space similarity, is a consistency factor and are the causal time interval of the current event and the historical event, respectively; is a preset time threshold, which can be set to 72 hours; is the event space position distance; is the maximum influence radius, which can be set to 5 km; and are decay coefficients, which can be both set to 0.5.

[0201] The parameters in the present application can be valued according to experience if not specially explained, and the model parameters can be optimized based on learning.

[0202] The training process of the model in the present application is described below.

[0203] After 220,000 pieces of data after word segmentation and stop word preprocessing are divided into a training set, a validation set and a test set according to a ratio of 8:1:1. The overall training process is as follows: for the event extraction task, the format of [CLS] + accident text + [SEP] is adopted, and the BIOES label is output, wherein [CLS] is a classification mark, BIOES label is a boundary label (including B, I, O, E, S), and [SEP] (Separator) is a separation mark; for the relationship classification task, the format of [CLS] + event 1 description + [SEP] + event 2 description + [SEP] is adopted; for the semantic role labeling task, the format of [CLS] + complete sentence + [SEP] + target argument + [SEP] is adopted. All inputs are truncated to 512 tokens, and are jointly optimized through multi-task learning. For the English-Chinese parallel sentence pair, the “term alignment mask” strategy is adopted to ensure that the bilingual terms are learned synchronously in fine-tuning, and the random dropout rate (Dropout) is adopted to avoid overfitting. The AdamW optimizer is used, and the parameters = 0.9, ​​​​= 0.999, weight decay 0.01. Linear learning rate warm-up + cosine decay strategy was adopted, with initial learning rate , warm-up steps accounting for the first 10% steps, and maximum learning rate . The batch size was set to 16, gradient accumulation was 4 times, and the equivalent batch size was 64. A total of 10 rounds of training were performed, and the validation set was evaluated (F1 score) every round. If there was no improvement for 3 consecutive rounds, the training was stopped. The core parameter settings are shown in Table 1.

[0204] ;

[0205] In the pre-training stage, the Rail-BERT model was trained using MLM, and the standard cross-entropy loss function was used:

[0206] ;

[0207] wherein, is the MLM task loss function, is the set of masked word positions, is the original word at position , is the version of the text sequence after being masked, is the probability of the Rail-BERT model predicting the original word at position .

[0208] In the fine-tuning stage, the Rail-BERT model adapts to three downstream tasks simultaneously, and a multi-task joint loss function is used, which is composed of the following three parts:

[0209] Event relation classification task loss function The expression is:

[0210] ;

[0211] wherein, is the number of relation categories; is the one-hot encoding vector of the true label; is the probability of the Rail-BERT model predicting the relation category .

[0212] Event extraction task loss function is obtained by adding the trigger word recognition loss, event type classification loss, and argument extraction and role labeling loss.

[0213] Trigger word recognition loss in event extraction task The expression is:

[0214] ;​

[0215] where, is the number of trigger word labels BIEOS classes, is the true trigger word label of the t-th word, is the probability predicted by the Rail-BERT model that the t-th word belongs to label q.

[0216] Event type classification loss in event extraction task The expression is:

[0217] ;

[0218] where, is the number of identified trigger words in the sentence, is the number of event classes, is the true event class label of the t-th trigger word, is the probability predicted by the Rail-BERT model that the t-th trigger word belongs to class q. .

[0219] Argument extraction and semantic role labeling loss in event extraction task The expression is:

[0220] ;

[0221] where, is the number of events, is the number of words, is the number of semantic role classes, is the true role label of the t-th word in the f-th event, is the probability predicted by the Rail-BERT model that the word belongs to role r, if .

[0222] The total loss function is:

[0223] ;

[0224] Ablation study is a core method in machine learning to analyze the contribution of model components. Its essence is to systematically remove or modify specific parts of the model and observe performance changes to quantify the importance of each component. Four sets of weight combinations are set for the parameters The ablation experiment results show that when the weight combination is ​​​​​At this time, the relationship classification task F1 value, the event extraction task F1 value and the decision response time are all the best, so the weight combination is adopted,

[0225] Compared with multiple models, the performance of the designed Rail-BERT model is evaluated. The performance evaluation results of the model can be seen from Figures 4 to 10 . Figures 4 to 6 The performance comparison diagrams of the multiple model event extraction, event relationship classification and semantic role labeling are respectively Figures 4 to 6 It can be seen that for the event extraction task, the relationship classification task and the semantic role labeling task, the precision (Precision), recall (Recall) and F1 value of the Rail-BERT model of the present application are all significantly higher than those of other models such as BERT. In the original data set, 10,000 "causal relationship" documents, 6,500 "time sequence relationship" documents, 4,000 "condition relationship" documents and 1,500 "concurrent relationship" documents are selected. In order to evaluate the effectiveness of event relationship classification, the confusion matrix of different models in the railway field is calculated, Figures 7 to 10 The confusion matrix diagrams of the Rail-BERT model, the BERT model, the BiLSTM+Attention model and the Bio-BERT model are respectively shown. Figures 7 to 10 The confusion matrix in the figure shows the prediction results of different models in four categories (causal relationship, time sequence relationship, condition relationship and concurrent relationship). The number in each cell represents the number of samples predicted by the model when the actual category is the row label and the predicted category is the column label. For example, in the confusion matrix of the "Rail-BERT" model: in the "causal relationship" category, the model correctly predicts 9,019 samples, in the "time sequence relationship" category, the model correctly predicts 5,679 samples, in the "condition relationship", the model correctly predicts 3,562 samples, and in the "concurrent relationship" category, the model correctly predicts 1,235 samples. The prediction accuracy rates of the four categories are 90.4%, 87.3%, 89.0% and 82.2% respectively, and the recall rates are 91.5%, 85.0%, 81.8% and 79.8% respectively. It can be seen from Figures 7 to 10 that the prediction accuracy of the Rail-BERT model in the embodiment of the present application is obviously improved compared with the other three models.

[0226] In summary, the Rail-BERT model constructed in the present application has achieved good results in the field of railway emergency decision-making.

[0227] Embodiment Two

[0228] The railway emergency auxiliary decision-making system provided in the embodiment comprises a memory and a processor.

[0229] The memory is used to store a computer program.

[0230] The processor is configured to invoke the computer program to execute the method of embodiment one.

[0231] The embodiments of the present application also construct a railway emergency decision system based on the Rail-BERT model. The system adopts a hierarchical architecture design method, visualizes the model retrieval results, and is divided into four levels, namely, a front-end display layer, a back-end control layer, a core retrieval layer, and a data protection layer, to construct a Web-based intelligent management platform for railway accident cases. The platform is used to store historical event reports and causal paths, and realizes end-to-end response within 1s.

[0232] A RESTful service is built through a Flask framework, a non-structured data feature extraction is realized by integrating a PyPDF2 text analysis engine, a multi-modal data warehouse is constructed by combining MongoDB and GridFS, and distributed storage and metadata management of case documents are realized. Based on the Rail-BERT model and the causal path modeling capability of the railway affairs graph, a three-level fusion similarity retrieval mechanism of “event level-path level-time and space level” is established, which effectively improves the case matching accuracy in complex accident scenarios. The system adopts a modular development strategy, and through technical means such as service decoupling, asynchronous loading, and cache optimization, it ensures that the retrieval response speed is guaranteed while having good function expansion and large-scale data carrying capacity, providing a complete solution for case knowledge precipitation and intelligent analysis in the field of railway safety.

[0233] The front-end display layer uses HTML5 and Bootstrap 5 to build a responsive interactive interface, and integrates a PDF upload portal, multi-dimensional retrieval, accident statistics, and decision suggestion modules through a card layout. The front-end display is based on a current event analysis causal path graph drawn by Neo4j. The case detail page shows the sorting results of Top-K similar cases, the evolution path of the railway affairs graph, and the emergency disposal measures, and the right side is embedded with a rule library and real-time decision suggestions.

[0234] The back-end control layer relies on the Python Flask framework to build a RESTful service, and analyzes the uploaded PDF accident report text through PyPDF2 to realize non-structured data feature extraction. MongoDB is used to store case metadata and establish a composite index, and files larger than 16 MB are stored through GridFS block storage; a multi-layer security mechanism is designed: file hash verification is used to ensure data integrity; IP access frequency limitation is implemented to defend against malicious requests; and replica set redundancy backup is used to ensure data availability and disaster recovery capability. The Flask backend calls the Cypher query interface of Neo4j through the py2neo library to map the causal path after analyzing the current event into a graph query.

[0235] The core retrieval layer fuses the Rail-BERT model and the railway affair graph technology: the Rail-BERT model is generated based on the pre-training of the term weight mask strategy The semantic vector of the event element is represented (as 768); the "event level-path level-time and space level" three-level fusion similarity algorithm based on the railway affair graph is used to comprehensively calculate the causal path evolution path similarity with a weight of 0.4:0.4:0.2, GPU acceleration calculation and Faiss vector index structure are combined to improve the efficiency of large-scale vector similarity retrieval, and end-to-end response within 1 second is achieved.

[0236] The data protection layer uses the TLS protocol to establish an encrypted channel between the client and the server at the transmission layer, realizes transmission link protection through certificate bidirectional authentication, key negotiation and dynamic session key generation, ensures that the PDF file uploading, case retrieval and result transmission process resist the risks of man-in-the-middle attack and eavesdropping, integrates the MongoDB field-level encryption (CSFLE) technology at the storage layer, generates a master key based on the key management service (KMS) to perform client-side encryption processing on sensitive fields (such as the number of casualties, geographic location coordinates, equipment number, etc.) in the accident report, so that the encrypted data can only be stored in the form of ciphertext, and the database server and unauthorized users cannot directly access the plaintext, preventing data leakage and unauthorized access.

[0237] Embodiment three

[0238] The embodiment provides a computer readable storage medium, and the computer readable storage medium stores a computer program.

[0239] Embodiment four

[0240] The embodiment provides a computer program product, which comprises a computer program.

[0241] The specific implementation manners of the system, the electronic device, the computer readable storage medium and the computer program product provided in the embodiment can refer to the specific embodiments of the above method, and details are not described here.

[0242] The technical contents of the above embodiments can be referred to each other, and for the same or similar technical features, part of the embodiments is omitted.

[0243] Obviously, those skilled in the art should understand that the units or steps of the present application described above can be realized by universal computing devices, and they can be centralized on a single computing device or distributed on a network composed of multiple computing devices. Alternatively, they can be realized by program codes executable by the computing devices, so that they can be stored in storage devices and executed by the computing devices, or they can be respectively manufactured into individual integrated circuit modules, or multiple modules or steps among them can be manufactured into a single integrated circuit module to realize. Thus, the present application is not limited to any particular combination of hardware and software.

[0244] The preferred embodiments of the present application are described above, but the present application is not limited to the above. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for railway emergency assistance decision making, characterized in that, The application relates to a method for constructing a railway incident case base and an emergency decision-making method based on the railway incident case base. Step 1: collecting railway accident reports and preprocessing the collected railway accident reports to obtain a structured representation of the railway accident reports, and constructing a railway incident case base based on a text sequence of the railway accident reports; Step 2: based on the railway incident case base, a Rail-BERT model is pre-trained by using a mask language model; Step 3: the Rail-BERT model is fine-tuned based on a downstream task; wherein the downstream task includes event extraction, event relation classification and semantic role labeling; wherein in the event relation classification, a multi-dimensional confidence model is constructed to quantify the confidence of a relation triple {subject event, relation type, object event} in the railway field knowledge; event extraction, event relation classification and semantic role labeling results are aligned to construct a railway event graph, wherein the railway event graph includes a node set and a relation type edge set, wherein the node represents an event, and the edge represents the relation between events; Step 4: injecting railway event graph information into the Transformer encoder layer of the Rail-BERT model through a knowledge weight matrix; Step 5: combining the semantic vectors corresponding to the current incident and the historical events generated based on the Transformer encoder layer to perform similarity matching on the current incident and the historical events in the railway event graph based on three-level fusion similarity to obtain K historical events with the highest similarity; wherein the three-level fusion similarity includes event-level similarity, path-level similarity and space-time-level similarity; Based on the K historical events with the highest similarity, an emergency auxiliary decision is made based on the disposal scheme.

2. The method of claim 1, wherein, The multi-dimensional confidence model is represented as: ； ； ； ； In the formula, represents the normalized confidence, , and is a weight coefficient, is the syntactic evidence strength, is the semantic similarity, is the co-occurrence frequency; represents the semantic correlation tightness of the subject event and the object event; is the decay coefficient; represents the causal predicate weight; represents and the cosine similarity, represents the semantic vector corresponding to the subject event, represents the semantic vector corresponding to the object event; and are obtained based on the pre-trained Rail-BERT model; represents the time decay factor, is the time corresponding to the subject event, is the time corresponding to the object event, is a preset time threshold; is an indicator function, if the event pair co-occurs in the document , then , otherwise ; represents the spatial distance of the two events; the weight coefficient .

3. The method of claim 1, wherein, In step 4, the knowledge weight matrix calculation method is: The input text sequence is embedded and positionally encoded to obtain a position-enhanced representation of the text sequence ; The node and edge embeddings of the railway logic graph are learned using a relational graph convolutional network, resulting in the node embedding matrix of the railway logic graph. Based on node embedding matrix Generate edge incidence matrix ; Mapping events in a text sequence to railroad event graph nodes by entity linking, generating an alignment matrix ; Generating a knowledge weight matrix: .

4. The method of claim 3, wherein, In step 4, the railway event graph information is injected into the Transformer encoder layer of the Rail-BERT model through the knowledge weight matrix, including: The knowledge weight matrix is incorporated into the attention weight calculation of each encoder layer of the Transformer model, and the formula is: ； In the formula, Indicates the first Attention weights calculated from each attention head; This represents the knowledge fusion intensity coefficient. It is the dimension of attention head; , and The first The query matrix, key matrix, and value matrix of each attention head.

5. The method of claim 1, wherein, The event-level similarity calculation formula is: ； wherein, represents the event level similarity of the current event node and the historical event node , are semantic vectors of the node and the node , which are generated based on a Transformer encoder layer; represents calculating the cosine similarity; is an event type weight, if and belong to the same type of event, then , otherwise 0; is a weight coefficient.

6. The method of claim 1, wherein, The path-level similarity calculation formula is: ； wherein, represents the path-level similarity of the current event and the historical event, and respectively represent the causal path of the current event and the historical event; is the graph structure edit distance, i.e., the minimum operation cost based on node addition, deletion, and modification and edge relationship adjustment; is the confidence weight of the relationship edge ; and is the path length, i.e., the number of nodes.

7. A railway emergency assistance decision system, characterized in that, The application relates to a method for constructing a railway incident case base and an emergency decision-making method based on the railway incident case base. The memory is used for storing a computer program; The processor is used for calling the computer program to execute the method in any one of claims 1 to 6. The computer program stored in the computer readable storage medium makes the electronic device realize the method in any one of claims 1 to 6 when the computer program runs on the electronic device.

8. A computer-readable storage medium, characterized in that, The computer program makes the electronic device realize the method in any one of claims 1 to 6 when the computer program runs on the electronic device.

9. A computer program product comprising a computer program, characterized in that, ​

Images