Train fault analysis and decision method fusing dynamic knowledge and affair logic

Abstract

The application relates to a train fault analysis decision method fusing dynamic knowledge and matter logic, and belongs to the field of natural language processing and machine learning. Firstly, fault events in a train fault phenomenon description text are extracted, time sequence dynamic knowledge and matter logic relations are fused, and a knowledge graph is constructed. Then, a graph conversion network combining a multi-head attention mechanism is used to encode the knowledge graph, and a priority vector representation of the knowledge graph is generated. Finally, a decoder combining an attention mechanism is used to select a key operation sequence from a solution word table constructed in advance according to the encoding result to form a solution. Since the application adds modeling of time sequence dynamic knowledge and matter logic relations of fault events, the effectiveness and accuracy of the analyzed and generated solution to the fault phenomenon are improved.

Description

TECHNICAL FIELD

[0001] The application relates to a train fault analysis and decision-making method fusing dynamic knowledge and matter logic, and belongs to the field of natural language processing and machine learning. BACKGROUND

[0002] Train fault analysis and decision-making is an important part of train safety operation and maintenance. When dealing with train faults, artificial needs to rely on a large amount of professional knowledge support, historical and real-time rail transit network situation awareness, and make accurate judgments in a short time according to the situation, in this process, it is easy to be negligent and missed. With the wide application of knowledge graph technology, a fault decision-making method based on knowledge graph is generated, but it mainly stays in using simple entity relationship triplets to construct a knowledge graph, dividing according to fault core components and fault detection core elements, focusing on entities and the relationship between entities, and lacking the mining of train fault matter logic knowledge. The evolution law of the fault event and its successive occurrence in time and the dynamic knowledge and matter logic relationship of the fault event are not fully modeled. At the same time, this kind of method uses traditional graph neural network to learn the knowledge graph, which cannot well capture the global semantic features, so that the matter logic semantic features such as continuity, causality and condition in the train fault event are diluted.

[0003] Therefore, the application provides a train fault analysis and decision-making method fusing time sequence dynamic knowledge and matter logic, to solve the problem that the evolution law of the fault event and the matter logic are not effectively utilized, so as to improve the effect of train fault phenomenon analysis and auxiliary decision-making. SUMMARY

[0004] The purpose of the application is to fuse the time sequence characteristics of the fault event to dynamically model the knowledge, to reason the fault event chain, and to use the graph conversion network based on the multi-head attention mechanism to learn the knowledge graph, to better capture the continuity, causality and condition features of the fault event from the global, so as to effectively solve the problem that the dynamic space-time evolution law and the logical relationship of the train fault event are not effectively modeled and utilized.

[0005] The design principle of the application is as follows: first, a solution meta-operation word table is constructed, and a train fault description text is subjected to named entity recognition and relation extraction by using an information extraction tool; second, the extracted fault event, minimum maintainable unit, solution and logical relationship are modeled, and a global node is added to construct a knowledge graph fusing dynamic knowledge and matter logic; then, a graph conversion network (Graph Transformer) based on a multi-head attention mechanism is used to learn and encode the graph structure, to obtain a priority vector representation of the constructed knowledge graph; finally, a decoder combining the multi-head attention mechanism is used to output the solution in the priority sequence.

[0006] The technical scheme of the present application is realized by the following steps:

[0007] Step 1, construct the solution meta-operation word table. Extract all solutions from the train fault handling log data set, divide the key operations therein, define them as meta-operations, and form a meta-operation word table.

[0008] Step 2, use information extraction tools to perform named entity recognition and relationship extraction on train fault phenomenon description text.

[0009] Step 2.1, introduce expert knowledge and use information extraction tools to extract fault events and the smallest maintainable units involved in each log.

[0010] Step 2.2, construct a keyword dictionary, and calculate the logical relationship of fault event affairs by combining keyword matching and fault event timing information.

[0011] Step 3, use the fault events, smallest maintainable units, and logical relationships of affairs obtained in the above steps to model, add global nodes, and construct a knowledge graph that integrates timing dynamic knowledge and logical relationships of affairs.

[0012] Step 3.1, construct fault events as P1 class nodes, and the logical relationships of affairs of fault events as R1 class edges, each R1 class edge connecting two P1 class nodes. For P1 class nodes, different weights are assigned to the nodes according to the timing of event occurrence.

[0013] Step 3.2, construct the smallest maintainable units as P2 class nodes, and construct a special edge R2 class edge. P2 class nodes are child nodes of P1 class nodes. Each P2 class node is connected to the corresponding P1 node through the R2 class edge.

[0014] Step 3.3, construct a global node P0, and construct a R3 class edge. R3 class edge is used to connect all P1, P2 class nodes with global node P0.

[0015] Step 4, encode the knowledge graph using a graph transformer network based on a multi-head attention mechanism.

[0016] Step 4.1, restructure the graph obtained in the above steps into an unlabeled bipartite graph to facilitate encoding by the graph transformer network.

[0017] Step 4.2, the graph transformer network focuses on each node and its adjacent nodes in the graph, and follows a multi-head attention mechanism to calculate the hidden representation of each node in the graph.

[0018] Step 4.3, obtain the contextualized encoding of the entire knowledge graph by concatenating the hidden representations of each node.

[0019] Step 5, output the solution in priority sequence by using the decoder combined with the multi-head attention mechanism.

[0020] Advantages

[0021] Compared with the general train fault analysis and decision method, the application can fully consider the dynamic knowledge and logical relationship of the evolution rule of the fault event and its successive occurrence in time, better model the fault event and the solution, improve the logicality and accuracy of the fault analysis and decision, and use the graph conversion network based on the multi-head attention mechanism to process the knowledge graph, so that the global features and hidden semantic features of the knowledge graph can be better captured, and the effect of the train fault analysis and decision is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 Principle diagram of the train fault analysis and decision method of the application fusing dynamic knowledge and logic of things.

[0023] Figure 2 Knowledge graph instance diagram constructed by the application. DETAILED DESCRIPTION

[0024] In order to better illustrate the purpose and advantages of the application, the embodiments of the method of the application will be further described in detail below in combination with examples.

[0025] The specific process is as follows:

[0026] Step 1, build a solution meta-operation word table. Extract all solutions from the train fault handling log data set, divide the key operations therein, define them as meta-operations, and form a meta-operation word table.

[0027] Step 2, use an information extraction tool to perform named entity recognition and relationship extraction on the input train fault phenomenon description text.

[0028] Step 2.1, introduce expert knowledge, and use an information extraction tool to extract the fault event and the minimum maintainable unit involved in the fault in each log.

[0029] Step 2.2, construct a keyword dictionary, integrate the time sequence features, and calculate the time logical relationship of the fault event by combining keyword matching and context description sequence. The relationship includes: cause and effect, condition, sequence, upper and lower, and concurrency.

[0030] Step 3, model the fault event, the minimum maintainable unit and the logical relationship obtained in the above steps, add global nodes, and construct a knowledge graph fusing time sequence dynamic knowledge and logic of things.

[0031] Step 3.1, construct the fault events as P1 class nodes, and the spatio-temporal logical relationship of the fault events as R1 class edges, each R1 class edge connecting two P1 class nodes. Denoted as (P1-a, R1, P1-b). Different weights are given to the nodes according to the time sequence of the event occurrence.

[0032] Step 3.2, construct the minimum maintainable unit as P2 class nodes, and construct a special edge R2 class edge at the same time. The P2 class nodes are the child nodes of the P1 class nodes. Each P2 class node is connected to the corresponding P1 node by means of the R2 class edge. Denoted as (P2-a, R2, P1-a). There is only one kind of R2 class.

[0033] Step 3.3, construct a global node P0, and construct a R3 class edge to connect all P1, P2 class nodes with the global node P0.

[0034] Step 4, encode the knowledge graph obtained in the above steps by using a graph transformer network based on a multi-head attention mechanism.

[0035] Step 4.1, reconstruct the graph obtained in the above steps into an unlabeled bipartite graph in order to facilitate the encoding of the graph transformer network. The specific method is as follows: if there is (P1-a, R1-[a->b], P1-b), replace the labeled edge R1 with two vertices: make such vertices into edge nodes, one representing the positive direction of the relationship NR-[a->b], and the other representing the reverse direction of the relationship NR-[b->a]. The label type of the edge is converted to the weight of the edge node by calculation. Connect the two edge nodes to the nodes P1-a and P1-b connected by the original edge. The same operation is performed for (P2-a, R2, P1-a), and since R2 has only one class, the edge node converted from R2 also has only one weight. The final structure of these reconstruction operations is an unlabeled directed graph. Denoted as G=(V, E). Where V is the list of all nodes, and E is the adjacency matrix describing the directed edges.

[0036] Step 4.2, the graph transformer network focuses on each node in the graph and its adjacent nodes, and follows a multi-head attention mechanism to calculate the hidden representation of each node in the graph. Specifically, V is embedded into a dense continuous space, resulting in a matrix, as shown in equation (1).

[0037]

[0038] Each node is represented as v i by focusing on v iThe other nodes connected to G are contextualized. Each node computes the attention weight as shown in equation (2). A self-attention layer with N taps is used, where N independent attention layers are involved in the calculation and connection after the connection to other connections, and the calculation method is shown in equation (3).

[0039]

[0040]

[0041] Let K represent the series connection of N attention heads, N i represent the field of v i in G where a n is the attention mechanism parameter of each attention head, and the attention calculation method is shown in equation (4).

[0042]

[0043] Each a learns the independent transformation W k of k and q q , and the resulting product is normalized on all connected edges, and in order to weaken the weakening of these dot products to gradient descent, its gradient is marked as

[0044] Step 4.3, construct a block network and stack it L layers to enhance the multi-head attention calculation layer. Specifically, add a regularization layer and a feedforward layer after the multi-head attention layer to form a block network, and stack the block network, allowing information to propagate in the graph. The block network is stacked L times, and the output of layer L-1 is used as the input of layer L, that is, as shown in equation (5).

[0045]

[0046] Step 4.4, the resulting vertex encoding is called the knowledge graph contextualized node encoding.

[0047] Step 5, use the decoder combined with the multi-head attention mechanism to output the solution in the priority sequence.

[0048] Step 5.1, use the decoder to decode the priority vector encoding obtained by the graph conversion network encoder, and use the multi-head attention mechanism to calculate and construct the final fault event chain vector C g , and the calculation method is shown in equation (6).

[0049]

[0050] Use the input feedback decoder to get h t and C gAs input is passed to the next RNN timestamp.

[0051] h t and C g Calculate the probability p, which determines the selection of the meta operation from the solution meta operation word table to form a meta operation sequence, that is, the solution.

[0052] In the training phase, the model learns by minimizing the negative joint log-likelihood of the output meta operation sequence and the corresponding meta operation sequence in the original training set. In the application phase, the fault phenomenon description text is input into the trained model, and the corresponding solution is output.

[0053] In the test, 200 fault phenomenon description texts are analyzed, and the solutions are output. Three experts in the field manually score to determine whether the problem can be accurately solved. If a solution is given by two experts, it is considered accurate. After testing, the solution generated by the method has a 92% pass rate, of which the proportion of three experts scoring results is 87%.

[0054] The above specific description further details the purpose, technical solutions and benefits of the application. It should be understood that the above description is only a specific embodiment of the application and is not intended to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application should be included in the protection scope of the application.

Claims

1. A train fault analysis decision-making method that fuses dynamic knowledge and case-based reasoning logic, characterized by The method comprises the following steps: Step 1, constructing a solution meta-operation word table; extracting all solutions from the train fault handling log data set, dividing the key operations therein, defining them as meta-operations, and forming a meta-operation word table; Step 2, based on the train fault handling text, adopting a task-driven named entity recognition and relationship extraction method, jointly recognizing semantic units, the semantic units including fault events, minimum maintainable units, device components, time markers, operation verbs, operation objects and fault positions; combining the time sequence information and causal trigger words, constructing an event causal relationship set with semantic types and time constraints; Step 3, based on the extracted fault events, minimum maintainable units and the logical relationship therebetween, constructing a knowledge graph structure for train fault analysis and decision generation: modeling the fault events with clear time sequence as P1 type nodes and assigning time perception weights to support causal reasoning ordering; modeling the minimum maintainable units as P2 type nodes as functional sub-nodes of the P1 nodes, connected through R2 type edges; modeling the logical reasoning relationship between events as R1 type edges, embedding semantic information such as event type, operation condition and action sequence; introducing a global control node P0 and connecting all P1 and P2 nodes through R3 type edges; Step 4, encoding the knowledge graph obtained in the above steps: first, converting the multi-class semantic node graph structure into a label-free directed connection graph, and introducing a structure reconstruction strategy to maintain the propagation separation and semantic control boundary of the P1, P2 and P0 nodes in the connection topology; on this basis, a graph Transformer model based on a multi-head attention mechanism is used to encode the graph, and in the encoding process, prior structural information such as the semantic category of the node, the time weight, the connection type, etc. is introduced to realize selective aggregation of the upper and lower cultural representations; finally, the node representation is guided by introducing a global control node and a multi-head attention path; Step 5, decoding by using a decoder combined with a multi-head attention mechanism to select a meta-operation sequence to constitute a solution.

Images