An accident scene reasoning method and device

Abstract

The application discloses an accident scene reasoning method and device, the method comprises the following steps: extracting a causal relationship mention from historical data and constructing a causal relationship mention dataset; the historical data records the cause event and the result event of the historical accident; the causal relationship mention represents the text containing the causal relationship pair; the causal relationship pair includes the cause event and the result event related to each other; the causal relationship mention in the causal relationship mention dataset is parsed into a causal event; based on the causal event, at least two causal events with commonality are integrated to form a generalized causal event network; and the accident to be analyzed is reasoned based on the generalized causal event network. The accident scene reasoning method constructed by the application can mine valuable causal knowledge from a large number of English accident reports, and form a useful tool capable of predicting future accidents.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information processing, in particular to a reasoning method and device based on accident scene. BACKGROUND

[0002] The accident scene reasoning method is a complex and evolving field that involves the analysis, simulation, and prevention of accidents. The conventional reasoning method is case-based reasoning, which mainly uses historical cases to guide emergency decision-making and has important application prospects for emergency response. However, China currently lacks a complete emergency case library, and cases are mostly manually sorted and scattered in different departments, resulting in low reasoning efficiency and accuracy, which limits the practical application of accident scene reasoning. These technologies and problems indicate that the accident scene reasoning method is constantly evolving and faces many challenges, requiring further research and innovation to overcome these problems. SUMMARY

[0003] The present application provides a reasoning method and device based on accident scene, which is used to solve the problem of low reasoning efficiency and accuracy in the accident scene reasoning process in the prior art.

[0004] The present application provides a reasoning method for accident scene, the method comprises:

[0005] Extracting causal relationship mentions from historical data and constructing a causal relationship mention dataset; the historical data records the cause events and result events of historical accidents; the causal relationship mention represents text containing a causal relationship pair; the causal relationship pair includes: related cause events and result events;

[0006] Analyzing the causal relationship mentions in the causal relationship mention dataset into causal events;

[0007] Based on the causal events, at least two causal events with commonality are integrated to form a generalized causal event network;

[0008] Reasoning on the accident to be analyzed based on the generalized causal event network.

[0009] Optionally, extracting causal relationship mentions from historical data includes:

[0010] Determining the causal relationship connecting words in the historical data;

[0011] Extracting the cause events and result events associated with the causal relationship connecting words based on the causal relationship connecting words;

[0012] Obtaining the causal relationship mention based on the cause events, causal relationship connecting words, and result events.

[0013] Optionally, the parsing the cause-effect relation mention into cause-effect events comprises:

[0014] determining data features of the cause-effect relation based on the cause-effect relation mention to select a cause-effect relation template; the cause-effect relation template represents a grammatical format describing the cause-effect relation;

[0015] parsing the cause-effect relation mention into the cause-effect events based on the cause-effect relation template.

[0016] Optionally, the integrating at least two cause-effect events with commonality to form a generalized cause-effect event network based on the cause-effect events comprises:

[0017] calculating similarity of at least two cause-effect events;

[0018] when the similarity is greater than a preset threshold, integrating two groups of the cause-effect events into one generalized cause-effect event;

[0019] mapping the generalized cause-effect event to the generalized cause-effect event network.

[0020] Optionally, before the calculating similarity of at least two groups of the cause-effect events, the method further comprises:

[0021] vectorizing each of the cause-effect events to generate a feature vector of the cause-effect event.

[0022] Optionally, the reasoning on an accident to be analyzed based on the generalized cause-effect event network comprises:

[0023] constructing a graph neural network based on the generalized cause-effect event network;

[0024] taking the feature vector of each cause-effect event as a node attribute of each node in the graph neural network;

[0025] taking the number of simultaneous occurrence of two nodes as an edge attribute in the graph neural network;

[0026] constructing an accident prediction model based on the node attribute and the edge attribute to reason on the accident to be analyzed.

[0027] Optionally, the method further comprises:

[0028] constructing a training set and a test set based on first graph structure data;

[0029] taking an edge in the first graph structure data as a positive sample;

[0030] for a cause node of each edge, sampling multiple non-existing edges as negative samples in the training set and the test set respectively;

[0031] establish a model composed of two layers of GraphSAGE based on first graph structure data corresponding to positive samples in the training set, and calculate first possibility scores of edges existing between positive sample nodes;

[0032] calculate second possibility scores of edges not existing between negative sample nodes;

[0033] evaluate the accident prediction model based on the first possibility scores and the second possibility scores.

[0034] Optionally, the method further comprises:

[0035] optimize the accident prediction model through a cross-entropy loss function and a marginal loss function.

[0036] The application provides an accident scenario reasoning device, which comprises:

[0037] a construction unit configured to extract causal relationship mentions from historical data and construct a causal relationship mention dataset; the historical data records cause events and result events of historical accidents; the causal relationship mentions represent texts containing causal relationship pairs; the causal relationship pairs include cause events and result events related to each other;

[0038] a parsing unit configured to parse the causal relationship mentions in the causal relationship mention dataset into causal events;

[0039] a generalization unit configured to integrate at least two causal events with commonalities to form a generalized causal event network based on the causal events;

[0040] a reasoning unit configured to reason about an accident to be analyzed based on the generalized causal event network.

[0041] The application further provides a computer readable storage medium storing one or more programs, which can be executed by one or more processors to implement the method according to any one of the above.

[0042] The accident scenario reasoning method constructed by the application can mine valuable causal knowledge from a large number of English accident reports and form a useful tool capable of predicting future accidents. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a flowchart of an accident scenario reasoning method in an embodiment of the application;

[0044] Figure 2 is a construction diagram of a generalized causal event network in an embodiment of the application;

[0045] Figure 3is a generalization causal event network construction graph in an embodiment of the present application;

[0046] Figure 4 is a model establishment schematic diagram of a two-layer GraphSAGE in an embodiment of the present application;

[0047] Figure 5 is a reasoning device structure schematic diagram of an accident scene in an embodiment of the present application;

[0048] Figure 6a and 6b are model performance evaluation index schematic diagrams based on cross-entropy loss function and marginal loss function respectively in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The present application will be further described below in conjunction with the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, only the parts related to the present application are shown in the drawings, but not all the structures.

[0050] The present application provides an accident scene reasoning method, as shown in Figure 1 The method comprises the following steps:

[0051] Step 100, extracting causal relation mentions from historical data and constructing a causal relation mention dataset; the historical data records cause events and result events of historical accidents; the causal relation mention represents a text containing a causal relation pair; the causal relation pair includes: a cause event and a result event related to each other. Specifically, the accident scene reasoning of the embodiment of the present application is based on an accident scene inference method (ASIM) constructed by the present application, and the historical data comes from English accident reports, in which the cause events and result events of historical accidents are recorded in English. A cause event and a corresponding result event can form a causal relation pair, and the text pair of the causal relation is a causal relation mention. For example, “liquefied ammonia [causes] frostbite on contact” is a typical causal relation mention, which contains a causal relation pair “liquefied ammonia → frostbite on contact”, wherein [causes] is a causal relation connecting word. This step can extract all the causal relation mentions from the historical data and construct a causal relation mention dataset. In this step, the causal relation mention in a single complete sentence is extracted, but not the causal relation implied in multiple sentences.

[0052] Step 200, parsing the causal reference mentions in the data set into causal events. The causal reference mentions extracted in step 100 are not convenient for storage, retrieval and application, so the second task of the ASIM method is to further parse the extracted causal reference mentions into specific causal events and simplify the representation, so as to more concisely and effectively store the accident information. Specifically, the causal event is a causal reference mention expressed in a concise and effective manner.

[0053] Step 300, integrating at least two common causal events to form a generalized causal event network based on the causal events. In the embodiment of the present application, the causal relationship between specific events is mined, but it is still not convenient to explore the deeper and more common causal laws behind a large number of historical accidents, and it is also not conducive to practical application. Therefore, a generalized causal event network is also constructed to more conveniently guide accident prevention and safety decision-making.

[0054] Step 400, reasoning on the accident to be analyzed based on the generalized causal event network. After the construction of the generalized causal event network, the prediction of future accident scenarios is the ultimate goal of the ASIM model, and the ASIM model constructed by machine learning in the embodiment of the present application is used to reason on the accident to be analyzed.

[0055] The reasoning method of the accident scenario constructed in the embodiment of the present application can mine valuable causal knowledge from a large number of English accident reports, and form a useful tool capable of predicting future accidents.

[0056] The reasoning method of the accident scenario in the embodiment of the present application, preferably, the causal reference mentions extracted from the historical data include:

[0057] Identifying causal connection words in historical data. It is the primary task of accident scenario reasoning model to extract high-quality causal mentions from a large number of English accident reports comprehensively and systematically. However, it is almost impossible to extract all the causal mentions from a large number of English reports by manual analysis. Therefore, the embodiment of the present application can efficiently identify and extract causal mentions in the text based on rules by machine learning method. Specifically, the causal connection word refers to the word or phrase used to connect the cause event and the result event in the accident record. For example, "because", "because of", "lead to" and "after" are the four common causal connection words, accounting for more than 80% of all connection words. On this basis, the connection words are extended to make them more suitable for application in the accident analysis process. For example, in the accident report, when describing the changes of temperature, pressure and liquid level in chemical accidents, words such as "increases" and "decreasing" are often used to describe the causal relationship between chemical substances, reactions and states.

[0058] Extracting cause events and result events associated with the causal connection words based on the causal connection words. The following are two examples of causal mentions, and the causal connection words are shown in bold:

[0059] ·

[0060] ·

[0061] In addition, "the / a result of", "result in", "due to", "lead to" and "creating" can also be determined as causal connection words, and the embodiment of the present application lists 12 causal connection words as shown in Table 1.

[0062] Based on the cause event, the cause-effect conjunction word and the result event, a cause-effect mention is obtained. The embodiment of the present application formulates the extraction rule of the cause-effect mention of each conjunction word according to the template of <Pattern, Constraint>. Specifically, the Pattern in the template refers to the regular expression containing the cause-effect conjunction word, and the Constraint refers to the grammar constraint. For example, for the cause-effect conjunction word "cause", the Pattern (regular expression) of the cause-effect mention thereof is designed as "[sentence1][caus\S+][sentence2]", wherein [sentence1] is the cause event, [sentence2] is the result event, and the regular expression [caus\S+] is used to match different forms of "cause", such as "causes" and "cause". In addition, since "cause" has the same last five letters as "because", in order to avoid matching the text segment containing "because", the Constraint stipulates that the position before the matched "cause" must be a space instead of a letter. Using the customized template, all the cause-effect pairs connected by "cause" can be quickly identified, and two examples are as follows.

[0063] ·

[0064] ·

[0065] The cause-effect conjunction word and the cause-effect mention rule are shown in Table 1, wherein the sentence (sentence) marked in bold is the cause event, and the sentence not marked in bold is the result event.

[0066] Table 1 Cause-effect conjunction word and cause-effect mention rule

[0067]

[0068]

[0069] The method described in the above embodiment can efficiently and accurately extract the cause-effect mention from the English accident report.

[0070] The reasoning method of the accident scene described in the embodiment of the present application preferably analyzes the cause-effect mention into the cause-effect event, which includes:

[0071] Based on the cause-effect mention, the data features of the cause-effect relationship are determined to select a cause-effect template; the cause-effect template represents the grammar format describing the cause-effect relationship.

[0072] Based on the cause-effect template, the cause-effect mention is analyzed into the cause-effect event.

[0073] Specifically, the causal relationship templates include at least the following three types, respectively: (1) <subject-predicate-object> tuples; (2) noun phrase sets; and (3) verb and noun sets. These three types of causal relationship templates are suitable for different data characteristics. First, for verb and noun sets, the method aims to abstractly represent a specific event as a set of verbs and nouns. For example, the causal relationship mention "a BLEVE cause the tank's failure" can be represented as a verb and noun set {BLEVE, cause, tank, failure}. For concise textual descriptions, such as news headlines, a set of verbs and nouns can be used. For example, in the causal relationship mention "a BLEVE cause the tank's failure", the cause event "The vapor ignited" and the result event "explosion" can be clearly represented as {vapor, ignited} and {explosion} in the form of verb and noun sets, and the causal relationship can be represented as {vapor ignited}→{explosion}.

[0074] However, for complex sentences like "This process greatly compounded the total amount of explosive energy within each box", abstractly representing it as a set of verbs and nouns will increase the difficulty of understanding it, and even introduce irrelevant information interference. In this case, it is a better choice to extract the subject, predicate, and object of the sentence according to the syntactic structure of the sentence. The nominal subject of the sentence is "process", the direct object is "amount", the indirect object is "energy" and "box", and the predicate verb is "compounded". Therefore, this sentence can be simply represented as {process, compounded, amount, energy} using <subject-predicate-object> tuples. It can also be represented as {this process, the total amount, explosive energy, each box} using noun phrase sets.

[0075] Accident reports often have both simple sentences and complex sentences, and representing causal events in a single way can lose key accident information. Therefore, the embodiments of the present application use all three types of causal relationship templates to represent causal events, which can help researchers better utilize and understand accident information.

[0076] ​The accident scenario reasoning method of the embodiment of the present application, preferably, based on the causal events, at least two causal events with commonality are integrated to form a generalized causal event network, which comprises:

[0077] calculating the similarity of at least two causal events;

[0078] when the similarity is greater than a preset threshold, integrating two groups of the causal events into one generalized causal event;

[0079] mapping the generalized causal event to the generalized causal event network.

[0080] Specifically, the specific causal relationship pair of "miscommunication between the two shift supervisors→the unsafe use of steam in the reactor vessel" is converted into a more generalized representation "miscommunication→unsafe use". Obviously, the latter is a more concise causal knowledge, and such generalized causal information is beneficial to discovering more profound and common rules hidden behind specific causal events. For example, as shown in Figure 2 {lack, communication}, {supervisor, miscommunication}, {communication, cease}, and the like can be generalized to {lack, communication}. The red nodes represent specific events, and the directed arrows point from the cause event to the result event. The dashed line is the conversion process from the specific event to the generalized event (blue node).

[0081] The accident scenario reasoning method of the embodiment of the present application, preferably, before calculating the similarity of at least two groups of the causal events, the method further comprises:

[0082] Each of the causal events is vectorized to generate a feature vector of the causal event. Specifically, in order to identify and integrate highly similar causal events, it is necessary to quantize each causal event and represent it in the form of a feature vector. At the same time, the quantization process of the causal event should preserve the original event information as much as possible. Traditional text mining models usually use One-Hot encoding or TF-IDF to calculate the text similarity, ignoring the context information, which may lead to the loss of key information. Embodiments of the present application use a word vector representation model, which converts words into fixed-dimensional vectors and then calculates the similarity between words. This model maps semantically similar words in an embedding space, so that the word vector contains context information, and the model has the advantages of fast calculation speed and high accuracy. In addition, it is specifically divided into two models: Skip-gram and CBOW. Compared with Skip-gram, CBOW requires less training time and is more suitable for processing large amounts of text data. Preferably, the embodiments of the present application calculate the word vector based on the CBOW model.

[0083] The corpus of the embodiments of the present application includes 1175 accident reports from the eMARS database and 100 accident reports from the CSB website. First, the accident reports are processed sentence by sentence using regular expressions based on Python language, and useless characters and stop words are removed to obtain a corpus containing 73075 English sentences. Then the corpus is input into Word2vec for modeling, and finally each word in the corpus is represented by a 300-dimensional vector as shown in equations 1 and 2. Each specific causal event is composed of a series of words. Therefore, the feature vector of the causal event is defined as formula 3, and the similarity of two causal events is defined as formula 4.

[0084] Corpus=[W1,W2,W3,…,W n ] (1)

[0085] W=[x1,x2,x3,…,x 300 ] (2)

[0086]

[0087] In the formula, Corpus represents a corpus containing n words, W is a word vector, S j is the feature vector of the jth causal event, X j,i is the ith value in the feature vector of the jth causal event. Similarity j,k is the similarity of S j and S k , that is, the feature vector S j and the feature vector S kThe cosine value of the included angle, the closer the cosine value is to 1, the smaller the included angle of the two vectors, indicating that the two events are more similar. The similarity of the two causal events is calculated in this way.

[0088] Specifically, the preset threshold is 0.95, and for the pair of causal events with a similarity greater than 0.95, the causal event with a shorter tuple length is replaced by the causal event with a longer tuple length to generalize it. Finally, a plurality of generalized causal events construct a generalized causal event network, which is composed of nodes (i.e., causal events) and directed edges (i.e., causal relationships between causal events).

[0089] The inference method of the accident scene according to the embodiments of the present application is preferably based on the generalized causal event network to infer the accident to be analyzed, which includes:

[0090] Building a graph neural network based on the generalized causal event network;

[0091] Taking the feature vector of each causal event as the node attribute of each node in the graph neural network;

[0092] Taking the number of simultaneous occurrence of two nodes as the edge attribute in the graph neural network;

[0093] Building an accident prediction model based on the node attribute and the edge attribute to infer the accident to be analyzed.

[0094] After the generalized causal event network is constructed, the final goal of the ASIM model is to predict future accident scenarios. Similar to social networks and knowledge graphs, the generalized causal event network belongs to typical graph structure data, which is composed of nodes and edges, and is used to describe the internal structure of complex systems, such as Figure 2 However, compared with structured data, the processing and analysis of graph structure data are more complex, and the scale of the graph (i.e., the number of nodes and edges) can be very large, so it is difficult for traditional machine learning to efficiently identify the potential patterns and features in the graph structure data.

[0095] The embodiment of the application adopts a graph neural network (GNN) to process graph structure data, that is, the possibility of the existence of a causal relationship between nodes is predicted by modeling the information transmission mechanism between nodes in a knowledge network. Among them, the graph sample and aggregate (GraphSAGE) is one of the classic GNN algorithms, which is suitable for learning information propagation mechanisms from complex networks. The algorithm generates new node embeddings by aggregating the information (i.e., feature vectors) of adjacent nodes. The prediction tasks based on GraphSAGE include node classification, edge classification and link prediction (i.e., predicting whether there is a relationship between two nodes). Therefore, the specific embodiment of the application uses a causal event network and a GraphSAGE algorithm to construct an accident prediction model. The model is used to predict whether there is an edge between two specific nodes (i.e., whether there is a causal relationship between two events), and to reveal the possible development trend of the accident.

[0096] Specifically, the generalized causal event network (graph structure data) is denoted as G=(V,E), where V is a set of nodes and E is a set of edges. In the edge set E, c and e represent two nodes, (c,e) represents a directed edge connecting node c and node e, which means that there is a causal relationship between the two nodes. Secondly, as shown in Figure 3 The feature vector of each event is taken as a node attribute, and the co-occurrence frequency of two nodes (i.e., the number of simultaneous occurrences) is taken as an edge attribute, which can represent the relationship strength between two nodes.

[0097] Preferably, when constructing the generalized causal event network, the causal relationship template will be further integrated. Specifically, first, use the <subject-predicate-object> tuple method to detect whether the subject, predicate and object are contained in the sentence of a causal event, if yes, prefer to use the <subject-predicate-object> tuple to represent the event; otherwise, use the verb and noun set.

[0098] The reasoning method of the accident scene described in the embodiment of the application, preferably, the accident prediction model constructed by the application is essentially a link prediction task based on GNN, which aims to infer whether there is a causal relationship between two nodes in the graph. Specifically, it is constructed with the help of a deep graph library (DGL) based on pytorch. The specific construction method includes:

[0099] Based on the first graph structure data, a training set and a test set are constructed. The first graph structure is divided into a training set and a test set according to a certain proportion.

[0100] The edges in the first graph structure data are regarded as positive samples;

[0101] For each reason node of each edge, multiple non-existing edges (i.e. no edge between nodes) are sampled in the training set and the test set as negative samples, and each existing edge (c, e) has k negative samples (c, e <negative,i> ).

[0102] Based on the first graph structure data corresponding to the positive samples in the training set, a model composed of two layers of GraphSAGE is established, and the first possibility score of the existing edge between the positive sample nodes is calculated according to formula 5. Wherein the establishment of the model includes the following four steps, as shown in formula 6: a. input the first graph structure data with node attributes and edge attributes; b. for each node, sample neighbor samples with a depth of 2; c. aggregate the information of the green nodes to the blue nodes; d. aggregate the information of the red nodes to the green nodes, update the embedding of the green nodes, and then update the embedding of the blue nodes again. Figure 4

[0103] The second possibility score of the non-existing edge between the negative sample (c, e <negative,i> ) nodes is calculated.

[0104] The accident prediction model is evaluated based on the first possibility score and the second possibility score. Specifically, based on the test set, the model is evaluated using the Area Under Curve (AUC) index. For a specific node pair, if the score of the existing edge is greater than all the scores of the non-existing edges, add 1; if the score of the existing edge is equal to the maximum score value of all the non-existing edges, add 0.5; otherwise, add 0. After n iterations (i.e. n existing edges), n1 represents the number of the first case, and n2 represents the number of the second case, and the performance AUC of the obtained model is as shown in formula 8.

[0105]

[0106] The inference method of the accident scene according to the embodiments of the present application, preferably, the method further comprises:

[0107] The accident prediction model is optimized by a cross-entropy loss function and a marginal loss function.

[0108] In order to make the score of the existing edge higher than all the scores of the non-existing edges, two loss functions are used to update and optimize the model, such as the cross-entropy loss function of formula 6 and the marginal loss function of formula 7.

[0109] y c,e = φ (S c , S e ) = S c × S e (5)

[0110] ​

[0111] In the formula, S' c and S' e These are the embeddings of nodes c and e, respectively, calculated based on the GraphSAGE model; y c,e Let S' be the score for the existence of edge (c,e), i.e. c and S' e dot product; e i P is the negative sample node of the cause node c. n (e) represents a negative sample distribution. For negative sample edges (c,e) i The second probability score is k, where k is the number of negative sample nodes and M is a constant hyperparameter.

[0112] This invention also provides a reasoning device for accident scenarios, such as... Figure 5 As shown, the device includes:

[0113] Construction unit 501 is used to extract causal relationship mentions from historical data and construct a causal relationship mention dataset; the historical data records the causal events and result events of historical accidents; the causal relationship mentions represent text containing causal relationship pairs; the causal relationship pairs include: related causal events and result events;

[0114] Parsing unit 502 is used to parse the causal relationship mentions in the causal relationship mention dataset into causal events;

[0115] The generalization unit 503 is used to integrate at least two causal events with common characteristics to form a generalized causal event network based on the causal events.

[0116] The reasoning unit 504 is used to perform reasoning on the incident to be analyzed based on the generalized causal event network.

[0117] To verify the practicality of the ASIM model, this embodiment of the invention also modeled 100 accident reports from the US CSB website. Furthermore, using the established model, the accident process of a certain explosion was simulated, verifying the model's predictive ability for unknown accidents.

[0118] Specifically, the ASIM method was constructed to analyze 100 major chemical accident reports from the US CSB. These accidents occurred between 1998 and 2022, as shown in Table 2. The accident reports ranged in length from tens to hundreds of pages, containing rich details of the accidents. Furthermore, the recording format and style of these reports were relatively uniform, which was beneficial to the conduct of this study.

[0119] The content of the accident reports covered in the analysis includes modules of Factual information, Incident description, Incident analysis, Conclusions, and Key lessons, while the appendix and recommendations in each report are excluded from this analysis. The results show that 3558 causal relation mentions are extracted using 12 causal connectives. The distribution of the extracted causal relation mentions is shown in Table 3.

[0120] Table 2 CSB reports collected

[0121]

[0122] Table 3 Distribution of extracted causal relation mentions

[0123]

[0124] Using the three causal relation templates of verb + noun set, <subject-verb-object> tuple and noun phrase set, the 3558 causal relation mentions are further parsed into causal events. Finally, 5855 causal events and 3558 causal relation pairs are obtained. These causal events can be stored in the information database of causal events, preparing for the subsequent generalization and prediction of causal events. The storage format of the causal events is shown in Table 4-4, including the causal relation templates, for example, the causal event "Managers ordered a total refinery evacuation, after the fire erupted" can be represented as: (1) {fire, erupted}→{managers, order, refinery, evacuation}; (2) {fire, erupted}→{managers, ordered, evacuation}; (3) {the fire}→{managers, a total refinery evacuation}. However, different representation methods are suitable for different events, so the next step is to optimize the event representation.

[0125] Table 4 Examples of causal event representation

[0126]

[0127] Generalized representation of all causal events. First, we determine whether the original text of each causal event contains a <subject-verb-object> structure; if so, we use the <subject-verb-object> template to abstractly represent the event; otherwise, we represent it using a verb + noun set. Second, all words are root-processed to avoid the impact of different forms of the same word on the performance of the model (e.g., “ignition”, “ignited” and “ignites” are all replaced by “ignite”). Third, we calculate the similarity between two causal events using their feature vectors, and further integrate events with a similarity greater than 0.95. As a result, we obtain 4982 generalized causal events and 3249 causal event pairs. In addition, we map these generalized causal events and causal relationships into a vector space to form a generalized causal event network. We use the generalized causal event network to reveal some key event categories and their distribution. In this network, all data are stored in the form of nodes (i.e., causal events) and directed edges (i.e., causal relationships). Table 5 shows some high-frequency event chains.

[0128] Table 5. High-frequency event chains

[0129]

[0130] We train the prediction model from the established generalized causal event network. The generalized causal event network consists of nodes and edges with feature vectors. In short, the prediction model aims to learn the impact mechanism or information propagation pattern between these nodes. Therefore, when a new event is input into the trained model, the most likely subsequent events caused by the event can be inferred. In this empirical study, the dataset is divided into training and test sets in a 9:1 ratio. Using the evaluation index shown in Equation 8, the performance of the model based on the cross-entropy loss function and the marginal loss function is shown in Equations Figure 6a and 6b The higher the AUC value, the better the performance of the model.

[0131] Since the model is built based on the English accident reports, to verify the practicability and generalizability of the reasoning model, an explosion accident is taken as an example to verify the reasoning process and accuracy of the model. According to the Chinese record of the explosion accident, first, it is translated into English expression, and the event chain (composed of seven basic events) of the explosion accident is built by manual analysis. Then, the initial event "In the clethodim workshop, an operator did not check and confirm the materials (an operator in the clethodim workshop did not check and confirm the materials)" of the explosion accident is input into the model to predict the five most likely subsequent events. Subsequently, the five predicted events are compared with the true accident chain, and the event closest to the true situation is taken as the input of the next prediction. After six predictions, the reasoning model generates a set of potential accident scenarios from the initial event. Among the constructed potential accident scenarios, there is an event chain consistent with the true accident. As expected, after six predictions using the prediction model, each prediction can accurately predict an event close to the actual situation (i.e., one of the five most likely events). More importantly, the prediction results of the model give more potential accident paths, which has important enlightenment significance for safety management and accident prevention practice.

[0132] The embodiment of the present application further provides a computer readable storage medium, the computer readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method described in any one of the above.

[0133] It should be understood that, in various embodiments herein, the size of the sequence number of each process described above does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments herein.

[0134] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can adopt a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer usable program codes.

[0135] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0136] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0137] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0138] The foregoing description of specific exemplary embodiments of the application has been presented for the purposes of illustration and description. It is not intended to be a limitation on the broad concepts of the application. Obviously, many modifications and variations of the specific exemplary embodiments described herein are possible in light of this disclosure, which is intended to be protected not only by the forms described but also by any alternative forms to which the principles disclosed in this document can be applied. The embodiments were chosen and described in order to explain the principles of the application and its practical application and to allow others skilled in the art to understand the application for various embodiments with various modifications as are suited to the particular use contemplated. It is intended that the scope of the application be defined by the claims and their equivalents.

Claims

1. A reasoning method for accident scenarios, characterized in that, The method includes: Causal references are extracted from single complete sentences in historical data to construct a causal reference dataset; the historical data records the causal events and result events of historical accidents; the causal references represent text containing causal pairs; the causal pairs include: related causal events and result events, and the historical data is derived from English accident reports; The causal references in the causal reference dataset are parsed as causal events; Based on the causal events, at least two causal events with common characteristics are integrated to form a generalized causal event network. The generalized causal event network includes graph structure data. Integrating at least two causal events with common characteristics to form a generalized causal event network includes: vectorizing each causal event to generate a feature vector of the causal event; calculating the similarity between at least two causal events; when the similarity is greater than a preset threshold, integrating the two sets of causal events into a generalized causal event; mapping the generalized causal event to the generalized causal event network, where the feature vector of the causal event is a word vector. Reasoning based on the generalized causal event network includes: constructing a graph neural network based on the generalized causal event network; using the feature vector of each causal event as the node attribute of each node in the graph neural network; using the number of times two nodes appear simultaneously as the edge attribute in the graph neural network; constructing an accident prediction model based on the node attribute and edge attribute to reason about the accident to be analyzed; the accident prediction model is used to predict whether there is a causal relationship between two events, and the prediction result of the accident prediction model includes multiple potential accident paths; and constructing the accident prediction model based on the node attribute and edge attribute to reason about the accident to be analyzed includes: inputting a new event into the accident prediction model and reasoning to obtain the most likely subsequent event caused by the new event.

2. The reasoning method for accident scenarios according to claim 1, characterized in that, Causal relationships extracted from historical data include: Identify causal conjunctions in historical data; Based on the causal relationship connectors, extract the associated causal events and result events; Causal references are obtained based on the causal events, causal conjunctions, and result events.

3. The reasoning method for accident scenarios according to claim 1, characterized in that, The causal relationship reference is resolved to causal events including: Based on the data features used to determine the causal relationship, a causal relationship template is selected; the causal relationship template represents the grammatical format for describing the causal relationship. Based on the causal relationship template, the causal relationship reference is parsed as the causal event.

4. The reasoning method for accident scenarios according to claim 1, characterized in that, The method further includes: Training and test sets were constructed based on the first graph structure data; The edges in the first graph structure data are considered as positive samples; For each cause node of an edge, multiple non-existent edges are sampled as negative samples in both the training and test sets; A model consisting of two layers of GraphSAGE is built based on the first graph structure data corresponding to the positive samples in the training set, and the first probability score of the existence of edges between positive sample nodes is calculated. Calculate the second probability score of the absence of edges between negative sample nodes; The accident prediction model is evaluated based on the first probability score and the second probability score.

5. The reasoning method for accident scenarios according to claim 4, characterized in that, The method further includes: The accident prediction model is optimized using the cross-entropy loss function and the marginal loss function.

6. A reasoning device for an accident scenario, characterized in that, The device includes: A construction unit is used to extract causal references from a single complete sentence in historical data and construct a causal reference dataset; the historical data records the causal events and result events of historical accidents; the causal references represent text containing causal pairs; the causal pairs include: related causal events and result events, and the historical data comes from English accident reports; The parsing unit is used to parse the causal relationship mentions in the causal relationship mention dataset into causal events; A generalization unit is used to integrate at least two causal events with common characteristics to form a generalized causal event network based on the causal events. The generalized causal event network includes graph structure data. Integrating at least two causal events with common characteristics to form a generalized causal event network includes: vectorizing each causal event to generate a feature vector of the causal event; calculating the similarity between at least two causal events; when the similarity is greater than a preset threshold, integrating the two sets of causal events into a generalized causal event; and mapping the generalized causal event to the generalized causal event network, wherein the feature vector of the causal event is a word vector. The inference unit is used to perform inference on the accident to be analyzed based on the generalized causal event network, including: constructing a graph neural network based on the generalized causal event network; using the feature vector of each causal event as the node attribute of each node in the graph neural network; using the number of times two nodes appear simultaneously as the edge attribute in the graph neural network; constructing an accident prediction model based on the node attribute and edge attribute to perform inference on the accident to be analyzed; the accident prediction model is used to predict whether there is a causal relationship between two events, and the prediction result of the accident prediction model includes multiple potential accident paths; and constructing the accident prediction model based on the node attribute and edge attribute to perform inference on the accident to be analyzed includes: inputting a new event into the accident prediction model and inferring the most likely subsequent event caused by the new event.

7. A computer-readable storage medium storing one or more programs, said one or more programs being executable by one or more processors to implement the method as described in any one of claims 1-5.

Images