Railway safety risk identification-oriented domain semantic enhancement relation extraction model

Abstract

Recognition of railway safety risks is crucial for ensuring continuous and stable operation of a railway, but the relation between entities to be detected is not considered in most of work at present. In addition, the railway text is difficult to understand, so that the final recognition accuracy is directly reduced by poorer domain semantics. In order to solve the challenging problems, a triple in a knowledge graph is introduced, and railway safety risks are modeled in a knowledge interconnection mode. Then, railway safety risk identification is redefined as a relation extraction task, and a new and effective domain semantic enhancement relation extraction model is provided; specifically, the invention designs a domain semantic enhancement Transform mechanism, and the mechanism can automatically enhance railway semantics from a special railway dictionary. And then, further introducing a segmented convolutional neural network to explore fine-grained features contained in the triple structure. By utilizing domain semantics and fine-grained features, the model can fully understand domain texts, so that the performance of relation classification is improved.

Description

[0001] Technical Space

[0002] The present invention relates to the fields of machine learning and natural language processing, and in particular to a method for using a domain semantics enhanced relationship extraction model to identify railway safety risks. Technical Background

[0003] Railway safety risk identification is a necessary condition to ensure the continuous and stable operation of railways. The key is how to automatically identify the safety risks of objects from the "three-in-one" scenario of human defense, physical defense, and technical defense. Traditional railway safety risk identification mainly relies on expert experience to mine safety risks from heterogeneous data. Construct fault trees and event trees to infer the evolution process of safety risks, so as to achieve the purpose of moving the safety risk checkpoint forward and changing post-event management to pre-event prevention. Faced with the massive data scale of intelligent rail transit systems, manual identification methods cannot meet the needs of big data analysis. At the same time, differences in human subjectivity lead to different identification standards, which affects the reasoning process of safety risk evolution. Relying on expert experience to construct fault trees and event trees cannot automatically infer the causal relationship between safety risks. The development of smart railways urgently needs to automatically identify railway safety risks and intelligently reason about their evolution process.

[0004] At present, deep neural networks driven by massive data have become the main method for identifying railway safety risks. DNN uses single neurons and combined neurons to simulate the abstract mechanism of human cognition, and imitates the iterative mechanism of human cognition through feedforward neural networks and back propagation, thereby acquiring the ability to self-learn and adapt to railway field data. However, deep neural networks face two difficulties in identifying railway safety risks from text.

[0005] (1) Ignoring the importance of interconnection between railway objects. These existing models fail to explore the relationship between railway objects and must therefore rely on additional experience or prior knowledge to infer incidents / accidents. In addition, due to differences in human subjectivity, inconsistent results may be produced when additional experience or prior knowledge is introduced to explore such relationships.

[0006] (2) Lack of railway domain semantics. Almost all of these DNN-based models face a serious semantic gap between the railway domain and the general domain, especially in the field of text mining. Existing models usually use pre-trained models trained on open corpora (including Wikipedia, Wall Street Journal, etc.) to initialize data representation, while open corpora lack domain corpora, especially the Chinese Railway corpus, which is detrimental to understanding railway texts and thus affects the final performance. This shows that enhancing domain semantics is very important for understanding railway texts. Summary of the invention

[0007] In view of the above problems, the present invention discloses a domain semantics enhanced relationship extraction model for identifying railway safety risks, which includes four modules, a domain semantic feature module, an enhanced semantic feature module, a knowledge structure semantic feature module and a safety risk relationship classification module. For railway safety risk identification, the concept of triples in the knowledge graph is introduced to explore the relationship between the railway entities to be detected, and the railway safety risk identification is redefined as a relationship extraction task. First, the railway domain terms are constructed as a domain dictionary, and the N-gram mechanism is used to mine word boundary information from the dictionary from a statistical perspective to learn the semantic features of domain vocabulary; then the N-gram mechanism and the multi-head attention mechanism are fused into a domain semantics enhanced Transformer network to encode the input text, integrate railway and general semantics, and realize automatic recognition of railway text semantic information; on the basis of improving the ability to understand domain text, the segmented maximum pooling operation is further introduced in each segment through the segmented convolutional neural network to capture fine-grained features, instead of implementing a single maximum pooling operation in the entire sentence; the output of the segmented convolutional neural network is a score matrix O associated with the relationship type, and then, for each sentence S, the probability of the relationship type can be calculated by the soft-max operation, and finally the cross entropy is used as the objective function.

[0008] Compared with the prior art, the effective effects of the present invention are:

[0009] (1) This paper analyzes railway safety risks from the perspective of knowledge interconnection for the first time, using triple knowledge to model the subject and object of safety risks and their relationships. Compared with existing methods, knowledge interconnection not only provides a new way to identify railway safety risks, but also lays the foundation for knowledge reasoning of railway safety risk evolution.

[0010] (2) A safety risk identification model based on domain semantic enhancement is proposed. This model continuously learns the semantic features of railway text data by enhancing the Transformer mechanism and segmented convolutional neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0012] Figure 1 It is a schematic diagram of identifying railway safety risks implemented by the present invention;

[0013] Figure 2 It is a schematic diagram of the DSERE model structure implemented in the present invention;

[0014] Figure 3 It is a schematic diagram of performance comparison between the DSERE model implemented in the present invention and different classical models;

[0015] Figure 4 is a schematic diagram of performance comparison between the present invention and the comparison algorithm in enhanced domain semantics and general semantics respectively;

[0016] Figure 5 is a schematic diagram of performance comparison between the present invention and the comparison algorithm in fine-grained feature extraction;

[0017] Figure 6 It is a schematic diagram of the ablation results of the DSERE model implemented in the present invention on a real word dataset;

[0018] Figure 7 is a schematic diagram comparing the performance of the DSERE model implemented by the present invention and the recently derived model;

[0019] Figure 8 It is a schematic diagram of example comparison between the present invention and the comparison algorithm on the test set.

[0020] Specific implementation

[0021] Railway safety risk identification is a necessary condition for ensuring the continuous and stable operation of railways. The key is how to automatically identify the safety risks of objects from the "three-in-one" scenario of human defense, physical defense, and technical defense. Construct fault trees and event trees to infer the evolution of safety risks, so as to achieve the purpose of moving the safety risk checkpoint forward and changing post-event management to pre-event prevention.

[0022] Inspired by the semantic interconnection in the knowledge graph, railway safety risk is modeled as triple knowledge, and the subject and object of safety risk are interconnected through the safety risk semantic relationship. Figure 1 As shown in the figure, the deep neural network perceives the abstract semantics of "…the 3# switch east side tip joint bolt is missing a…" through the context, and automatically extracts the triple knowledge of (3# switch, equipment safety risk relationship, joint bolt missing), which expresses that the phenomenon of missing joint bolts on the switch will induce train driving safety risk accidents. Knowledge interconnection forms a structured semantic knowledge base, laying the foundation for knowledge reasoning for downstream railway safety risk evolution tasks. For example, to analyze "a 3# switch train derailed, resulting in the injury of operator Zhang XX", first search the graph to locate the node "3#

[0023] By searching the node "Zhang XX" and "Zhang XX", we found that the knowledge (3# switch, equipment safety risk relationship, lack of joint bolts) is the direct cause of train derailment, and (Zhang XX, personnel safety risk relationship, walking on the switch) is the direct cause of personnel injury; further reverse search of the node "Zhang XX" found that "management safety risk relationship" connects "Zhang XX" and "unqualified training", indicating that the fundamental reason for Zhang XX walking on the switch is that management has not strictly constrained it. Identifying railway safety risks from the perspective of knowledge interconnection can effectively model and deduce railway safety risks, and promote the transition of safety risk management from perceptual intelligence to cognitive intelligence.

[0024] 1. Problem definition,

[0025] Railway safety risk refers to the unsafe state of people or equipment caused by environmental changes in the scenario of railway operation. In order to improve the generalization of railway safety risk identification, the knowledge-oriented railway safety risk is defined as follows:

[0026] Definition 1 The core element of railway safety risk is a triple ∑(E 1 ,R,E 2 ),in:

[0027] It is a set of railway safety risk subjects, representing the set of human, machine, environment and management safety risk subjects, among which:

[0028] Including line workers, bridge and tunnel workers, guards, crossing workers, flaw detectors, etc.

[0029] Mainly refers to rails, ballast, roadbed, turnouts, bridges, etc.;

[0030] Refers to environmental factors that seriously affect the normal operation of railways, such as strong winds, heavy rains, floods, earthquakes, etc.;

[0031] Refers to various safety management systems, such as "Rail Transit Reliability, Availability, Maintainability and Safety Specifications and Examples", "Railway Roadbed Major Maintenance Rules", etc.

[0032] It is a collection of railway safety risk objects, including unsafe conditions caused by human, machine, environment and management safety risk factors, including:

[0033] For example, unsafe behaviors such as not wearing reflective tape, not "signaling with hands", and not leaving the road in time can easily induce personnel safety risk events;

[0034] It is the equipment status that may induce hidden dangers to train operation safety, such as cracks, subsidence, siltation, collapse, etc.;

[0035] Mainly refers to the safety hazards induced by lightning, earthquakes, etc. to personnel or equipment;

[0036] For example, grassroots workers or equipment.

[0037] R = {r per ,r dev ,r evn ,r mg} is a set of railway safety risk relationship types, where r per represents the personnel safety risk relationship, r dev represents the equipment safety risk relationship, r evn represents the environmental risk safety relationship, r mg Represents the Manage Security Risk relationship.

[0038] Definition 2: Knowledge-oriented railway safety risk is a tuple ∑(S,T), where:

[0039] 1) S = {s i |s i It is marked with e j and e k The statement, e j ∈E 1 ,e k ∈E 2 ,i,j,k∈N +} is the annotated railway safety risk text set.

[0040] 2) T 1 ={T per ,T dev ,T evn ,T mg} is a collection of railway safety risk knowledge, where:

[0041] It is a collection of personnel safety risk knowledge, expressing the semantics of personal safety risk events that may be induced by unsafe behaviors of personnel;

[0042] It is a collection of equipment safety risk knowledge, expressing the semantics of how an unsafe state of equipment may lead to safety hazards;

[0043] It is a collection of environmental safety risk knowledge, expressing the semantics of personnel or equipment being affected by extreme weather, thus affecting the safe operation of railways;

[0044] It is a collection of knowledge on managing safety risks, expressing the semantics of safety hazards caused by the lack of strict constraints on people and things by rules and regulations.

[0045] The railway safety risk identification for knowledge interconnection is to use DNN to identify the type of knowledge T from S, that is, It means that the deep neural network is trained on the text corpus S to automatically extract the railway safety risk knowledge contained in each text sentence.

[0046] 2. Overall framework of the model

[0047] Figure 2The overall architecture of the DSERE model is presented. First, railway domain terms are constructed as a domain dictionary. Then, the N-gram mechanism and multi-head attention mechanism are integrated into an enhanced Transformer network to realize automatic recognition of railway text semantic information. Based on the improvement of the model's ability to represent domain text, the knowledge structure features are further learned from a fine-grained perspective through a segmented convolutional neural network. Finally, railway safety risks are identified through a softmax classifier.

[0048] 2.1 Domain Semantic Features

[0049] In order to understand the semantic information of the railway field from text data, DSERE learns the domain semantic features through N-gram. It mines word boundary information from a statistical perspective in the context and learns the semantic features of domain vocabulary. Let the N-gram sequence of the training sample s∈S be in L is a dictionary consisting of railway terms, and N-gram can be expressed as:

[0050]

[0051] 2.2 Enhanced semantic features

[0052] The enhanced Transformer mechanism fuses statistical semantic features and network self-perception semantic features to generate enhanced semantic features. The domain semantic information and shallow semantic information obtained by multi-head self-attention mechanism (MhA) network encoding N-gram are as follows:

[0053]

[0054] in The attention of other N-gram elements in layer l is calculated for the query vector Q, U l is the matrix representing the features of all N-gram elements in layer l, providing information for K and V in MhA; v l and Respectively represent the i-th word in layer l and the k-th N-gram element associated with the word; V l is the embedding matrix of all words, M is the N-gram matching matrix, V l* is the input to the next layer.

[0055] 2.3 Semantic Features of Knowledge Structure

[0056] Railway safety risks are expressed through triple knowledge, and the semantic features of its knowledge structure are the semantic features of the relationship between the subject and the object. Convolutional neural networks can mine the implicit semantic features of the relationship in the context in the nonlinearity of distributed representation and feature fusion. The DSERE model will enhance the semantic features X = {x1, x2, ..., x |x|} Input the convolutional neural network. The convolution operation is performed through matrix dot multiplication. The i-th convolution operation is as follows:

[0057] cov i =[tanh(Wq+b1)] i

[0058] where q = x i-l′+1:i (1≤i≤|X|+l′-1, l′ is the size of the convolution kernel) is the concatenation of the word vector sequence in the i-th sliding window, W={w i ,w2,…,w a}(w i ∈R l′×d ) is the convolution kernel matrix, b1 is the bias vector, and tanh is the activation function. After a convolutional encoding, we get Con = {con1, con2, ..., con a}, and then fully explore the semantic features of knowledge structure from the fine-grained level of text. Each convolution layer is divided into three parts according to the position of entity pairs. i1 ,con i2 ,con i3}, and then concatenate the results of the three parts after performing the pooling operation, namely:

[0059] p={p i1 ,p i2 ,p i3}

[0060] where p ij =max(cov ij ),1≤j≤3, to conduct in-depth analysis of relational semantics.

[0061] 2.4 Classification of security risk relationships

[0062] The eigenvector g = tanh(p 1:a ) is input into the soft-max classifier to calculate the confidence of each security risk relationship type. The security risk relationship is classified according to the confidence. The final output of the DSERE model is:

[0063] o=W1g+b2

[0064] Where W1 is the matrix representation of the relationship and b2 is the bias term.

[0065] A soft-max operation is used on all security risk relationship types to calculate the probability of relationship labels:

[0066]

[0067] Where s is an input text sample, r i ∈R represents the i-th security risk knowledge.

[0068] Using cross entropy as the objective function, it is defined as follows:

[0069]

[0070] Where θ is the set of parameters of the neural network, Represents the training examples s i The true label of is the j-th security risk relationship, and the objective function is optimized by the gradient descent algorithm.

[0071] Figure 3 It is a schematic diagram of the performance comparison between the DSERE model implemented in the present invention and different classical models, in which three indicators and some comparison results are marked to better demonstrate the performance of the DSERE model. Figure 4 The influence of the present invention and the comparison algorithm on the experimental dataset in terms of enhanced domain semantics and general semantics is shown, and the precision / recall curve is used for a more intuitive representation. Figure 5 Three classic neural networks are compared based on the same semantics to verify the effectiveness of fine-grained feature extraction. Figure 6 The ablation results of the DSERE model implemented in the present invention on a real word dataset are shown, demonstrating the important role of the semantic feature enhancement mechanism in the model. Figure 7 It is a schematic diagram of the performance comparison of the DSERE model implemented by the present invention and the recently derived models, summarizing the AUC and F1 scores, semantic, solution, metrics to evaluate the overall performance. Figure 8 It is a schematic diagram of example comparison between the present invention and the comparison algorithm on the test set. Examples from the test set are compared, and each sentence is annotated by entity pairs and relationship types. Blue words indicate the existence of security risks, and red words indicate objects of security risks.

[0072] Those skilled in the art will readily understand the embodiments of the present invention. The above description is merely an example of the basic implementation of the present invention and is not intended to limit the present invention.

Claims

1. A domain semantics-enhanced relation extraction model for identifying railway safety risks, characterized by: (1) Introducing the concept of triples in the knowledge graph, exploring the relationship between the railway entities to be detected, redefining railway safety risk identification as a relationship extraction task, and projecting railway safety risks into a knowledge interconnected space that can be used for accident inference; (2) A domain semantic enhancement Transformer mechanism is designed, which can automatically enhance the railway domain semantic information from the railway dictionary through the N-gram mechanism; (3) A piecewise convolutional neural network is introduced to explore the fine-grained features contained in the triple structure and improve the performance of relation classification.

2. The domain semantics enhanced relationship extraction model for identifying railway safety risks according to claim 1 is characterized in that: Extract the relationship between railway entities. Through the investigation and statistics of railway safety risk events, 8 entities and 4 relationships are defined to describe the railway safety risk subjects, objects and their relationships. In order to interconnect railway objects, we model railway safety risks as triples: ∑(E 1 ,KING 2 )。 3. The domain semantics enhanced relationship extraction model for identifying railway safety risks according to claim 1 is characterized in that: The railway domain semantic information is enhanced from a dedicated railway dictionary. The N-gram mechanism is introduced from a statistical perspective to explore the word boundary features in the railway dictionary. is a railway terminology dictionary, the joint probability of N-gram can be expressed as:

4. The domain semantics enhanced relationship extraction model for identifying railway safety risks according to claims 1 and 3, characterized in that: Enhance the railway domain semantic information. The enhanced Transformer mechanism fuses statistical semantic features and general semantic features to generate enhanced semantic features. The domain semantic information and sentence input obtained by encoding N-gram through a multi-head attention network are encoded as follows:

5. The domain semantics enhanced relationship extraction model for identifying railway safety risks according to claim 1, characterized in that: Relation extraction for railway text. Based on enhanced semantic features, segmented convolutional neural network is used for relation extraction. The triple structure divides the sentence into three segments, and a segmented maximum pooling operation is introduced in each segment, instead of implementing a single maximum pool in the entire sentence, to capture these fine-grained features. Formally, the i-th convolution operation is defined as: The output is a score matrix O associated with the relation type. Then, for each sentence S, the probability of the relation type is calculated by the soft-max operation as:

6. The domain semantics enhanced relationship extraction model for identifying railway safety risks according to claims 1 and 5, characterized in that: Automatically enhance domain semantics and capture fine-grained features to extract relationships. The model objective function is: 。

Images