A knowledge-driven high-speed rail emergency event evolution reasoning method

Abstract

The application relates to the technical field of deep learning and event reasoning, and discloses a high-speed rail emergency evolution reasoning method based on knowledge driving, which solves the technical problems that high-speed rail emergency evolution information lacks structured representation and evolution direction is uncertain, so that it is difficult to mine the internal logical correlation of event evolution and the method is suitable for high-speed rail field evolution reasoning. The method comprises the following steps: constructing a high-speed rail emergency evolution ontology and an evolution chain, defining a reasoning mode according to the length of the evolution chain; performing word vector coding on the evolution chain information, obtaining historical information embedding representation through a feature mapping network; inputting the historical information embedding representation into a deep feature extraction encoder module to capture deep semantic features; inputting the deep semantic features into a domain information enhanced attention mechanism decoder module to decode the predicted node feature vector representation and obtain a reasoning prediction result through linear transformation; constructing a binary cross-entropy loss function to calculate the information distribution difference between the output and candidate events, and obtaining the best candidate event through normalized probability scoring.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and event reasoning, specifically to a knowledge-driven method for reasoning about the evolution of high-speed rail emergencies. Background Technology

[0002] In the field of high-speed rail emergency evolution prediction, evolutionary information often exists within massive amounts of text, mostly described in short text formats, lacking structured information representation. Secondly, the evolution of emergencies is characterized by chain reactions and uncertain evolutionary directions, making it difficult to uncover the inherent logical connections within the event's evolution. In recent years, with the rise of the internet, the detection and evolution prediction of public safety emergencies have received widespread attention from academia and industry. Compared to traditional emergency evolution prediction technologies, ontology-based knowledge-driven methods have significant advantages in node feature representation, causal relationship extraction, and evolutionary pattern recognition for public safety emergencies. Therefore, how to effectively predict the evolutionary characteristics and direction of public safety emergencies based on ontology representation technology is of great significance for formulating effective emergency response measures and reducing personnel and property losses.

[0003] Generally, the evolution of an event is correlated with observed time-series data. Traditional sequence-based emergency event prediction mainly utilizes methods such as cluster analysis, probabilistic models, and artificial neural networks. However, traditional methods cannot characterize the hidden features of evolving events and lack substantial modeling and analysis of the event evolution process. Furthermore, considering the uncertainty and suddenness of emergencies, emergency event prediction methods based on the "scenario-response" model have gained attention to reduce reliance on historical data and experience. These methods can be modeled using probabilistic models, artificial neural networks, and fuzzy theory. However, the accuracy and effectiveness of scenario simulation models depend on the assumptions and parameter settings regarding the event development process. If the assumptions are inaccurate or the parameter settings are unreasonable, the simulation results are prone to distortion, and there is a lack of substantial modeling and analysis of the event elements involved in the event evolution process.

[0004] In recent years, driven by deep learning and knowledge graph technologies, knowledge-driven methods for predicting emergencies have developed rapidly. The main modeling principle of these methods is to use semantic information such as ontology knowledge, graph knowledge, and pre-trained knowledge to improve the scene representation ability of emergency prediction. On this basis, they use temporal modeling and attention mechanisms to mine the correlation of event evolution, which makes up for the shortcomings of traditional methods in semantic capture and alleviates the shortcomings of the "scene-response" model in not being able to deeply mine the potential information of evolution patterns.

[0005] In the context of knowledge-driven emergency prediction, given the highly specialized nature of emergency knowledge and the unique evolutionary patterns of emergencies in the high-speed rail sector, such methods are difficult to directly apply to emergency prediction tasks in the high-speed rail field. Specifically, existing methods lack a structured definition of the evolutionary ontology of high-speed rail emergencies, lack effective modeling of the logical evolutionary relationships in the evolutionary chain of high-speed rail emergencies, and do not fully consider the constraints of high-speed rail emergency domain information on evolutionary reasoning, resulting in insufficient prediction accuracy in high-speed rail emergency evolutionary reasoning tasks. Summary of the Invention

[0006] The purpose of this invention is to address the technical problem that existing methods struggle to uncover the inherent logical relationships in the evolution of high-speed rail emergencies due to the lack of structured representation and uncertain evolutionary directions, making it difficult to directly apply them to emergency evolution reasoning in the high-speed rail field.

[0007] To achieve the above objectives, the present invention employs the following technical means:

[0008] This invention provides a knowledge-driven method for reasoning about the evolution of high-speed rail emergencies, comprising the following steps:

[0009] Step 1: Construct the evolution ontology and event evolution chain of high-speed rail emergencies, and define the evolutionary reasoning mode according to the length of the event evolution chain;

[0010] Step 2: Under the given evolutionary reasoning model, word vector encoding is performed on the evolutionary chain information of high-speed rail emergencies, and the historical information embedding representation of the evolutionary chain of high-speed rail emergencies is obtained through a feature mapping network;

[0011] Step 3: Input the historical information embedding representation into the deep feature extraction encoder module, and capture the deep semantic features in the historical information embedding representation through the deep feature extraction encoder module;

[0012] Step 4: Input the deep semantic features into the domain information enhanced attention mechanism decoder module, and decode the feature vector representation of the node to be predicted in the high-speed rail emergency evolution chain through the domain information enhanced attention mechanism decoder module;

[0013] Step 5: Perform a linear transformation on the feature vector representation to obtain the comprehensive feature representation of the event evolution chain; construct a binary cross-entropy loss function, calculate the information distribution difference between the comprehensive feature representation and the candidate events, and use the sigmoid function to normalize the probability score to obtain the best candidate event.

[0014] Specifically, considering that the evolutionary information of high-speed rail emergencies often exists in the form of unstructured short texts, and that different events have logical evolutionary relationships such as sequence, causality, and parallelism, this invention addresses the problems of existing event representation methods failing to uniformly depict event elements, attribute knowledge, and logical evolutionary relationships between events, as well as the lack of targeted reasoning methods for historical evolution chains of different lengths. Based on the above ideas, step 1 specifically includes:

[0015] Step 1.1: Construct the evolution ontology and event evolution chain of high-speed rail emergencies. The evolution ontology is in triplet form, represented as follows: ,in As the main body of the event, The event trigger word, The event object is defined as the event evolution chain by adding event attribute knowledge to the evolutionary ontology.

[0016] Step 1.2: Define the evolutionary reasoning model based on the length of the evolutionary chain of high-speed rail emergencies. The evolutionary reasoning model is expressed as: given the preceding... The first historical evolutionary chain unit predicts the... Each evolutionary chain unit, namely Reasoning patterns; a total of 4 evolutionary reasoning patterns are defined, specifically as follows: , , , .

[0017] In the above scheme, step 2 includes:

[0018] Step 2.1: Under the given evolutionary reasoning model, use the word vector encoder word2vec to process the evolutionary chain information of high-speed rail emergencies. After vectorization, we get The feature embedding representation; where represents the event evolution chain, Represents a node. Includes modifiers of object attributes Attribute words that modify the subject Organizational entities Entity of the Person ,time Location of the incident Logical evolutionary relationship Attribute description;

[0019] Step 2.2: Construct a feature mapping network processing module, which includes linear layers, vector stacking, and vector averaging operations. This module is used to obtain the historical information embedding representation of the evolution chain of high-speed rail emergencies. .

[0020] In the above scheme, the feature embedding representation in step 2.1 is obtained through the following formula:

[0021]

[0022]

[0023] in, , , , , , , , , and These are the model weight parameters. and For bias parameters, This represents a logical evolutionary relationship;

[0024] The historical information embedding representation in step 2.2 It can be obtained through the following formula:

[0025]

[0026] in, It is an average function. For vector concat operation, Feature embedding representation for each event node. It embeds a representation of the logical evolution relationship between adjacent nodes.

[0027] In the above scheme, step 3 includes:

[0028] Step 3.1: Embed the historical information into a representation Add position encoding to obtain position representation After adding positional encoding, the input features become :

[0029]

[0030] Step 3.2: Calculate using a multi-head attention mechanism The correlation between nodes is used to obtain attention output;

[0031] Step 3.3: Obtain the output of the deep feature extraction encoder module through the ADD&Norm layer, FeedForward layer, and residual structure. :

[0032]

[0033] in, For layer normalization operation, For handling loss, This outputs the results for the FeedForward layer.

[0034] In the above scheme, the calculation process of the multi-head attention mechanism in step 3.2 is as follows:

[0035] For the Each attention head has a query matrix. Key matrix Sum matrix Represented as:

[0036]

[0037] in, , and For the first The model weight parameters corresponding to each attention head. The output of each attention head is:

[0038]

[0039] in, for transpose, To define the feature dimensions of the query matrix and key matrix, softmax is used for normalization. The outputs of each attention head are concatenated and linearly mapped to obtain the multi-head attention output:

[0040]

[0041] in, This represents a vector concatenation operation. For the number of attention heads, Output the corresponding model weight parameters for multi-head attention.

[0042] Furthermore, considering the highly specialized nature of knowledge in the field of high-speed rail emergencies, the different evolutionary patterns of different event categories, and the varying degrees of influence of different historical events on subsequent events within the event evolution chain, this invention addresses the challenges of existing methods in utilizing event domain information to constrain the evolutionary direction and in identifying crucial evolutionary knowledge from historical evolution chains for predicting subsequent events. This approach introduces event domain information and historical event importance information into the event evolution reasoning process to enhance the model's ability to represent the evolutionary characteristics and key historical evolutionary knowledge in the high-speed rail field. Step 4 specifically includes:

[0043] Step 4.1: Assign event domain category information The vectorized representation is obtained after encoding by the word2vec word vector encoder. ,Will The output of the embedded deep feature extraction encoder module In the process, domain-enhanced features are obtained. :

[0044]

[0045] in, For activation function, and These are model parameters;

[0046] Step 4.2: Decode the feature matrix using a feature-selective self-attention mechanism. Important feature matrices :

[0047]

[0048] in, , For model parameters, This is the bias parameter.

[0049] In the above scheme, step 5 includes:

[0050] Step 5.1: For the feature matrix A linear layer is applied to obtain a comprehensive feature representation of the event evolution chain. :

[0051]

[0052] Step 5.2: Given a set of candidate events for prediction The set of vector representations is obtained through the word vector encoder word2vec. The candidate events are The correlation score is calculated using the sigmoid function. :

[0053]

[0054] in, and For model parameters, For bias parameters; via The best candidate event was selected.

[0055] In the above scheme, the event evolution chain is represented as follows:

[0056]

[0057] in, This represents the knowledge of the instantiated properties of the ontology. Indicates the first The instance unit and the first The logical evolutionary relationships between instance units include sequential relationships, causal relationships, and parallel relationships.

[0058] In the above scheme, the position encoding in step 3.1 The calculation formula is:

[0059]

[0060] in, For node position, For the position vector dimension, The first position vector represents the position vector. dimension, Dimension index for location encoding.

[0061] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0062] 1. This invention addresses the challenges of high-speed rail emergency event evolution information often existing within massive amounts of text, mostly described in short text format, lacking structured information representation, and exhibiting chain effects and uncertain evolutionary directions, making it difficult to uncover the inherent logical connections within the event evolution. To resolve the technical problems of lacking structured representation and targeted reasoning model definitions for high-speed rail emergency event evolution information, this invention employs the technical approach described in step 1: constructing a high-speed rail emergency event evolution ontology and event evolution chain, and defining four evolutionary reasoning models based on the chain length. This is achieved by representing high-speed rail emergencies as triples. The evolutionary ontology is constructed, and event attribute knowledge is added to the ontology nodes to instantiate knowledge, forming an evolutionary chain containing sequential, causal, and parallel relationships. This allows the evolutionary information of high-speed rail emergencies, originally existing in the form of unstructured short text, to acquire a structured knowledge representation; simultaneously, by defining... , , , Four evolutionary reasoning modes are used to guide the reasoning learning of subsequent algorithm models, enabling the models to predict the next evolutionary chain unit for historical evolutionary chains of different lengths. This solves the technical problem of lacking structured modeling and reasoning mode definition in the evolutionary process of high-speed rail emergencies.

[0063] 2. This invention addresses the challenges posed by the highly specialized nature of emergency response knowledge and the unique evolutionary patterns of emergencies in the high-speed rail sector. Existing knowledge-driven methods are difficult to directly apply to emergency prediction tasks in the high-speed rail field. To resolve the technical issue of general event prediction methods lacking knowledge constraints specific to the high-speed rail domain, this invention employs the technique described in step 4: constructing a domain-increased attention mechanism decoder module. This involves using event domain category information... The vectorized representation is obtained after encoding by the word2vec word vector encoder. ,Will The output of the deep feature extraction encoder module in step 3 is embedded. In this study, event domain class label information is used as a feature-driven model for training to improve prediction performance. This makes the decoder constrained by the high-speed rail emergency domain information when decoding the feature vector representation of the node to be predicted in the evolution chain of high-speed rail emergencies. This solves the technical problem that existing methods are difficult to directly apply to the task of predicting emergencies in the high-speed rail field.

[0064] 3. This invention addresses the issue that different historical event nodes in an event evolution chain contribute differently to predicting the next event, necessitating the identification of crucial historical evolutionary knowledge for prediction. To resolve the technical problem of difficulty in highlighting key evolutionary knowledge when different historical event features are equally weighted in reasoning, thus affecting candidate event discrimination, this invention employs the feature-selective self-attention mechanism designed in step 4. This mechanism automatically assigns weights to the historical event feature representations and decodes the feature matrix. The historical event evolution knowledge feature matrix plays an important role in predicting candidate events. This enables the model to capture historical evolutionary knowledge features that play an important role in predicting candidate events from the evolutionary chain, solving the technical problem that it is difficult to distinguish the degree of contribution of different historical event nodes in the event evolution chain to the prediction of candidate events.

[0065] 4. This invention combines the event evolution chain containing logical evolutionary relationships constructed in step 1 with the technical means of using event domain class label information as a feature-driven model training in step 4, and the technical means of automatically assigning weights to the feature representations of different historical events by the feature-selective self-attention mechanism in step 4. In the domain information-enhanced attention mechanism decoder module, while the event domain class label information drives model training, the feature-selective self-attention mechanism automatically assigns attention weights to important historical events. The two work together in the same decoder module to decode the feature vector representation of the node to be predicted in the evolution chain of high-speed rail emergencies, and obtain the prediction result of event reasoning. This achieves accurate prediction and event reasoning of high-speed rail emergencies, and solves the technical problem that the event domain class label information-driven model training and the historical event attention weight assignment need to be implemented simultaneously in the evolution reasoning of high-speed rail emergencies. Attached Figure Description

[0066] Figure 1 This is a simplified flowchart of the present invention;

[0067] Figure 2 This is a block diagram of the device corresponding to the present invention. Detailed Implementation

[0068] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0069] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0070] This paper proposes a novel domain ontology-driven method for predicting sudden events. First, it constructs a knowledge-embedded representation of high-speed rail sudden events. Second, it builds an encoder module based on node position encoding and multi-head attention mechanisms to capture deep semantic features from the historical information embedding representation. Then, it constructs a domain-enhanced attention mechanism decoder module to decode the feature vector representations of the nodes to be predicted in the evolution chain of high-speed rail sudden events, obtaining the prediction results for event reasoning. Finally, it effectively validates the method in two typical sudden event scenarios.

[0071] Example 1

[0072] Please see Figure 1 This invention provides a knowledge-driven method for reasoning about the evolution of high-speed rail emergencies, comprising the following steps:

[0073] Step 1: Construct the evolution ontology and event evolution chain of high-speed rail emergencies, and define the evolutionary reasoning mode according to the length of the event evolution chain.

[0074] Step 1.1: Construct the evolution ontology and event evolution chain for high-speed rail emergencies. Using a case study of a high-speed rail emergency, such as the example: "High-speed train departs normally, but the carriage experiences circuit aging, resulting in a short circuit in the carriage's wiring," the ontology node can be represented as a triple: "Departure [High-speed train, NAN], Experiencing [Carriage, Circuit Aging], Short Circuit [Carriage Wiring, NAN]," where "NAN" indicates that the missing part of the event description is missing. The evolution ontology is in triple form, represented as follows: ,in As the main body of the event, The event trigger word, The object of the event.

[0075] After obtaining the ontology node, the ontology node is instantiated into knowledge and represented as an evolutionary chain. The evolutionary chain information is represented as follows: :

[0076]

[0077] in, This represents the knowledge of the instantiated properties of the ontology. To modify the subject's attribute words, To modify the object attribute word, For organizational entities, For the physical form of a person, For time, The location where the incident occurred; Indicates the first The instance unit and the first The logical evolutionary relationships between individual instance units include sequential, causal, and parallel relationships. For a specific example... =Departure [High-speed rail, NAN, ...] Problems include [carriage, aging electrical circuits, ...] Short circuit [carriage wiring, NAN, ...].

[0078] Step 1.2: Define the evolutionary inference model based on the length of the evolutionary chain of high-speed rail emergencies. This method trains four evolutionary inference models, where RM=0 indicates that given a historical event evolutionary chain sequence of length 3, it is necessary to predict the fourth evolutionary event triplet instance unit, i.e. Mode; RM=1 is Mode; RM=2 is Mode; RM=3 is Pattern. Specifically, model:" =An incident occurred involving [high-speed rail, fire, ...] This resulted in [passengers being injured, ...] To hold [government meetings, conferences, etc.] "?" represents a candidate event to be predicted, where the set of candidate events is: ={Causes [NAN, high-speed rail suspension, ...], Announces [government, prohibition order, ...], Notifies [driver, emergency evacuation, ...], Stops [driver, NAN, ...]}, by calculating "?" and The relevance scores of candidate sub-events are used to deduce the best candidate event.

[0079] Step 2: Under the given evolutionary reasoning model, word vector encoding is performed on the evolutionary chain information of high-speed rail emergencies, and the historical information embedding representation of the evolutionary chain of high-speed rail emergencies is obtained through a feature mapping network.

[0080] Step 2.1: Under different evolutionary reasoning modes, analyze the generated event chain information. Obtained through the word vector encoder word2vec and The feature embedding representation, specifically, is as follows:

[0081]

[0082]

[0083] in, , , , , , , , , and These are the model weight parameters. and For bias parameters, This represents a logical evolutionary relationship.

[0084] Step 2.2: The feature embedding representation is processed through a linear layer, vector stacking, and vector averaging to obtain the encoder input. :

[0085]

[0086] in, It is an average function. For vector concat operation, Feature embedding representation for each event node. It embeds a representation of the logical evolution relationship between adjacent nodes.

[0087] Step 3: Input the historical information embedding representation into the deep feature extraction encoder module, and capture the deep semantic features in the historical information embedding representation through the deep feature extraction encoder module.

[0088] Step 3.1: Process the input First, add the corresponding position code to obtain... and Position representation :

[0089]

[0090] in, for and Location, For the position vector dimension, The first position vector represents the position vector. dimension, Dimension index for location encoding. Add Then, the input features become :

[0091]

[0092] Step 3.2: Calculate using a multi-head attention mechanism The correlation between nodes. For the first... Each attention head has a query matrix. Key matrix Sum matrix Represented as:

[0093]

[0094] in, , and For the first The model weight parameters corresponding to each attention head. The output of each attention head is:

[0095]

[0096] in, for transpose, To define the feature dimensions of the query matrix and key matrix, softmax is used for normalization. The outputs of each attention head are concatenated and linearly mapped to obtain the multi-head attention output:

[0097]

[0098] in, This represents a vector concatenation operation. For the number of attention heads, Output the corresponding model weight parameters for multi-head attention.

[0099] Step 3.3: To prevent overfitting, perform layer normalization on the data:

[0100]

[0101] in, For layer normalization operation, Loss handling helps prevent overfitting.

[0102] Subsequently, it is obtained through forward pass. :

[0103]

[0104] in, For activation function, and These are the model parameters.

[0105] Finally, using residual connections, and passing through the Add&Norm layer again, the final output is obtained. :

[0106]

[0107] in, For layer normalization operation, For handling loss, This outputs the results for the FeedForward layer.

[0108] Step 4: Input the deep semantic features into the domain information enhanced attention mechanism decoder module, and decode the feature vector representation of the node to be predicted in the high-speed rail emergency evolution chain through the domain information enhanced attention mechanism decoder module;

[0109] Step 4.1: The event domain class tag is represented as follows The vectorized representation is obtained after encoding by the word vector encoder word2vec. .Will Add encoder output In the process, the final feature vector is obtained. :

[0110]

[0111] in, For activation function, and These are the model parameters.

[0112] Step 4.2: To capture historical event evolution knowledge features that play an important role in predicting candidate events, a feature-selective self-attention mechanism is designed. This mechanism can automatically assign weights to the historical event feature representations and decode the feature matrix. Important feature matrices :

[0113]

[0114] in, , For model parameters, This is the bias parameter.

[0115] Step 5: Perform a linear transformation on the feature vector representation to obtain the comprehensive feature representation of the event evolution chain; construct a binary cross-entropy loss function, calculate the information distribution difference between the comprehensive feature representation and the candidate events, and use the sigmoid function to normalize the probability score to obtain the best candidate event.

[0116] Step 5.1: Use the binary cross-entropy loss function as the model's loss function, and obtain the result based on the linear transformation function. The linear transformation output. For Perform linear processing to obtain a comprehensive feature representation output of the event evolution chain. :

[0117]

[0118] Step 5.2: Given a set of candidate events for prediction The set of vector representations is obtained through the word vector encoder word2vec. The candidate events are The correlation score is calculated based on the linear transformation function and using the sigmoid function. , obtained Then it is determined to be the best candidate event:

[0119]

[0120] in, , and For model parameters, This is the bias parameter.

[0121] Example 2

[0122] This invention also provides a knowledge-driven high-speed rail emergency evolution reasoning device, comprising an emergency representation module, a deep feature extraction encoder module, and a domain information-enhanced attention mechanism decoder module, specifically:

[0123] First, the ontology, event evolution chain, and evolutionary reasoning mode of high-speed rail emergencies are defined. Second, under a specific evolutionary reasoning mode, the historical evolution chain is first vectorized using word2vec, followed by the construction of a feature mapping network processing module, which then becomes the input to the encoder. Next, a deep feature extraction encoder module is constructed, using positional encoding and multi-head attention operations to capture deep semantic features of historical evolution information. Then, a domain-enhanced attention mechanism decoder module is built, using event domain class label information as a feature-driven model for training to improve prediction performance. A self-attention mechanism for feature selection is designed to capture historical evolution knowledge features that are crucial for predicting candidate events, decoding the feature vector representation of the nodes to be predicted in the event evolution chain, and obtaining the prediction result of the event reasoning. Finally, in the output part, a linear transformation is performed, and a relevance score is calculated using the sigmoid function to determine the best candidate event.

[0124] The method of this invention starts from the perspective of reasoning about high-speed rail emergencies, takes the evolutionary chain as the research object, and uses word2vec to perform vector processing on high-speed rail emergency information; it also uses an encoder module to extract features from the event feature vector to enhance feature expression; and it constructs a domain information-enhanced attention mechanism decoder module to drive model training while automatically allocating attention weights to important historical events, thereby achieving accurate prediction and event reasoning of emergencies.

[0125] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge-driven method for reasoning about the evolution of high-speed rail emergencies, characterized in that, Includes the following steps: Step 1: Construct the evolution ontology and event evolution chain of high-speed rail emergencies, and define the evolutionary reasoning mode according to the length of the event evolution chain; Step 2: Under the given evolutionary reasoning model, word vector encoding is performed on the evolutionary chain information of high-speed rail emergencies, and the historical information embedding representation of the evolutionary chain of high-speed rail emergencies is obtained through a feature mapping network; Step 3: Input the historical information embedding representation into the deep feature extraction encoder module, and capture the deep semantic features in the historical information embedding representation through the deep feature extraction encoder module; Step 4: Input the deep semantic features into the domain information enhanced attention mechanism decoder module, and decode the feature vector representation of the node to be predicted in the high-speed rail emergency evolution chain through the domain information enhanced attention mechanism decoder module; Step 5: Perform a linear transformation on the feature vector representation to obtain the comprehensive feature representation of the event evolution chain; construct a binary cross-entropy loss function, calculate the information distribution difference between the comprehensive feature representation and the candidate events, and use the sigmoid function to normalize the probability score to obtain the best candidate event.

2. The method according to claim 1, characterized in that, Step 1 includes: Step 1.1: Construct the evolution ontology and event evolution chain of high-speed rail emergencies. The evolution ontology is in triplet form, represented as follows: ,in As the main body of the event, For event trigger words, The event object is defined as the event evolution chain by adding event attribute knowledge to the evolutionary ontology. Step 1.2: Define the evolutionary reasoning model based on the length of the evolutionary chain of high-speed rail emergencies. The evolutionary reasoning model is expressed as: given the preceding... The first historical evolutionary chain unit predicts the... Each evolutionary chain unit, namely Reasoning patterns; a total of 4 evolutionary reasoning patterns are defined, specifically as follows: , , , .

3. The method according to claim 1, characterized in that, Step 2 includes: Step 2.1: Under the given evolutionary reasoning model, use the word vector encoder word2vec to process the evolutionary chain information of high-speed rail emergencies. After vectorization, we get The feature embedding representation; where represents the event evolution chain, Represents a node. Includes modifiers of object attributes Attribute words that modify the subject Organizational entities Entity of the Person ,time Location of the incident Logical evolutionary relationship Attribute description; Step 2.2: Construct a feature mapping network processing module, which includes linear layers, vector stacking, and vector averaging operations. This module is used to obtain the historical information embedding representation of the evolution chain of high-speed rail emergencies. .

4. The method according to claim 3, characterized in that, The feature embedding representation in step 2.1 is obtained through the following formula: in, , , , , , , , , and These are the model weight parameters. and For bias parameters, This represents a logical evolutionary relationship; The historical information embedding representation in step 2.2 It can be obtained through the following formula: in, It is an average function. For vector concat operation, Feature embedding representation for each event node. It embeds a representation of the logical evolution relationship between adjacent nodes.

5. The method according to claim 1, characterized in that, Step 3 includes: Step 3.1: Embed the historical information into a representation Add position encoding to obtain position representation After adding position encoding, the input features become : Step 3.2: Calculate using a multi-head attention mechanism The correlation between nodes is used to obtain attention output; Step 3.3: Obtain the output of the deep feature extraction encoder module through the ADD&Norm layer, FeedForward layer, and residual structure. : in, For layer normalization operation, For handling loss, This outputs the results for the FeedForward layer.

6. The method according to claim 5, characterized in that, The calculation process for the multi-head attention mechanism in step 3.2 is as follows: For the Each attention head has a query matrix. Key matrix Sum matrix Represented as: in, , and For the first The model weight parameters corresponding to each attention head. The output of each attention head is: in, for transpose, To define the feature dimensions of the query matrix and key matrix, softmax is used for normalization. The outputs of each attention head are concatenated and linearly mapped to obtain the multi-head attention output: in, This represents a vector concatenation operation. For the number of attention heads, Output the corresponding model weight parameters for multi-head attention.

7. The method according to claim 1, characterized in that, Step 4 includes: Step 4.1: Assign event domain category information The vectorized representation is obtained after encoding by the word2vec word vector encoder. ,Will The output of the embedded deep feature extraction encoder module In the process, domain-enhanced features are obtained. : in, For activation function, and These are model parameters; Step 4.2: Decode the feature matrix using a feature-selective self-attention mechanism. Important feature matrices : in, , For model parameters, This is the bias parameter.

8. The knowledge-driven high-speed rail emergency evolution reasoning method according to claim 1, characterized in that, Step 5 includes: Step 5.1: For the feature matrix A linear layer is applied to obtain a comprehensive feature representation of the event evolution chain. : Step 5.2: Given a set of candidate events for prediction The set of vector representations is obtained through the word vector encoder word2vec. The candidate events are The correlation score is calculated using the sigmoid function. : in, and For model parameters, For bias parameters; via The best candidate event was selected.

9. The knowledge-driven high-speed rail emergency evolution reasoning method according to claim 8, characterized in that, The event evolution chain is represented as follows: in, This represents the knowledge of the instantiated properties of the ontology. Indicates the first The instance unit and the first The logical evolutionary relationships between instance units include sequential relationships, causal relationships, and parallel relationships.

10. The knowledge-driven high-speed rail emergency evolution reasoning method according to claim 5, characterized in that, Position encoding in step 3.1 The calculation formula is: in, For node position, For the position vector dimension, The first position vector represents the position vector. dimension, Dimension index for location encoding.

Images