Rail transit fault recovery time prediction method, system and equipment
By constructing a rail transit fault recovery event prediction model, using feature fusion in historical fault information and neural network training, the problem of inaccurate prediction of rail transit fault recovery time in the existing technology is solved, and fast and accurate prediction of fault recovery time is achieved.
Patent Information
- Application Number
- CN202411855414.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-17
- Publication Date
- 2025-05-06
AI Technical Summary
The prior art is difficult to accurately predict the recovery time of rail transit failures. Due to many factors, the estimate deviation is large.
By obtaining historical fault information, the event embedding vector, category embedding vector, event position feature matrix and influencing factor characteristics are extracted, and these features are fused to build a rail transit fault recovery event prediction model, and use neural networks for training and prediction.
It realizes rapid and accurate prediction of the fault recovery time of the rail transit system, reduces prediction deviations, and improves the efficiency of fault recovery management.
Smart Images
Figure CN119940599A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of fault time prediction, and in particular to a method, system and device for predicting rail transit fault recovery events. Background Art
[0002] For the daily operation of urban rail transit systems, it is inevitable that emergencies or accidents will cause the entire system to be unable to operate normally. In the existing technology, the fault recovery time is mostly estimated in a single type and single professional form, and the personnel judge the recovery time based on their processing experience. Since each professional system is relatively independent, it is also affected by internal factors such as passenger flow in the station, train load rate, emergency response capabilities of maintenance and passenger transport personnel, and external factors such as passenger flow around the station and real-time weather. Taking into account many factors, the estimation of rail transit fault recovery time cannot be relatively accurate. The more factors a fault involves, the greater the estimated deviation value will be. Summary of the invention
[0003] In order to overcome the above technical defects, the present invention provides a rail transit fault recovery event prediction method, system and device, which can quickly predict the time required for the rail transit system to resume normal operation.
[0004] In order to solve the above problems, the present invention is implemented according to the following technical solutions:
[0005] A method for predicting rail transit fault recovery time comprises the following steps:
[0006] Get historical fault information;
[0007] According to each fault information event in the historical fault information, the event embedding vector, the category embedding vector, the event position feature matrix, and the influencing factor features are obtained;
[0008] The event embedding vector, category embedding vector, and event position feature matrix are fused to obtain fault information features;
[0009] Add the fault information features and the influencing factor features to obtain the information aggregation features;
[0010] Using information aggregation features as input and actual recovery time as output, a rail transit fault recovery event prediction model for predicting recovery time is built;
[0011] Use historical fault information to train the rail transit fault recovery event prediction model;
[0012] The real-time fault information is input into the rail transit fault recovery event prediction model to predict the recovery time.
[0013] As a further improvement of the present invention, the step of obtaining an event embedding vector according to each fault information event in the historical fault information includes:
[0014] Perform one-hot encoding on each fault information event to obtain an event vector;
[0015] The fault event embedding matrix is used to embed the event vector to obtain the event embedding vector.
[0016] As a further improvement of the present invention, a category embedding vector is obtained according to each fault information event in the historical fault information, including:
[0017] For each category of the fault information event, a category embedding vector with the same dimension as the event embedding vector is randomly initialized.
[0018] As a further improvement of the present invention, the present invention further comprises the steps of:
[0019] The rail transit system stations and line networks are characterized by matrixing, and each element in the two-dimensional matrix is used to quantitatively store the information in the corresponding map grid.
[0020] As a further improvement of the present invention, according to each fault information event in the historical fault information, an event location feature is obtained, including:
[0021] According to the actual coordinate position of the site in the historical fault information, query the position of the corresponding element in the two-dimensional matrix;
[0022] The events occurring at the corresponding element positions in the two-dimensional matrix are represented as event position features.
[0023] As a further improvement of the present invention, according to each fault information event in the historical fault information, the influencing factor characteristics are obtained, including:
[0024] According to the time when the historical fault events occurred in the historical fault information, the passenger flow of the station where the fault occurred, the number of stations connected to the current station, the number of lines connected to the current station, and the network public opinion with positioning are collected;
[0025] According to the passenger flow of the station, the number of stations connected to the current station, the number of routes connected to the current station, and the network public opinion with positioning, the passenger flow feature matrix, the connected station feature matrix, the connected route feature matrix, the optimistic public opinion feature matrix, and the pessimistic public opinion feature matrix are constructed;
[0026] Connect the passenger flow feature matrix, the connected station feature matrix, the connected line feature matrix, the optimistic public opinion feature matrix, and the pessimistic public opinion feature matrix according to the channel direction to obtain the influencing factor feature matrix;
[0027] The impression factor neural network is used to extract the influencing factor features from the influencing factor feature matrix.
[0028] As a further improvement of the present invention, the step of fusing the event embedding vector, the category embedding vector, and the event position feature matrix to obtain the fault information feature includes:
[0029] Randomly sample a weight from a 0-1 uniform distribution;
[0030] Multiply the event embedding vector and the category embedding vector by the weights respectively and then add them together to obtain the fusion vector;
[0031] The event position feature matrix is converted into Gaussian prior features;
[0032] Based on Gaussian prior features, position information features are output through a position information neural network;
[0033] The fusion vector is added to the position information feature to obtain the fault information feature.
[0034] As a further improvement of the present invention, cross entropy loss is used as the loss function of training.
[0035] The present invention also provides a rail transit fault recovery event prediction system, which is used to implement the above rail transit fault recovery time prediction method, comprising:
[0036] A historical fault information acquisition module is used to acquire historical fault information;
[0037] The fault information learning module is used to obtain an event embedding vector, a category embedding vector, and an event position feature matrix according to each fault information event in the historical fault information, and fuse the event embedding vector, the category embedding vector, and the event position feature matrix to obtain the fault information feature;
[0038] An influencing factor learning module is used to obtain influencing factor characteristics based on each fault information event in the historical fault information;
[0039] The prediction module is used to add the fault information features and the influencing factor features to obtain the information aggregation features, and build a rail transit fault recovery event prediction model for predicting the recovery time based on the information aggregation features and the prediction head. At the same time, real-time fault information is obtained and input into the rail transit fault recovery event prediction model to predict the predicted recovery time.
[0040] The present invention also provides a computer device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;
[0041] The memory is used to store computer programs;
[0042] The processor is used to implement the above-mentioned rail transit fault recovery time prediction method when executing the program stored in the memory.
[0043] Compared with the prior art, the present invention has the following beneficial effects: by quantifying historical fault information and characterizing it as the input of the rail transit fault recovery event prediction model, the rail transit fault recovery event prediction model is trained using large-scale historical fault information, and the real-time fault information can be input into the rail transit fault recovery event prediction model according to different information and real-time conditions of the emergency event to predict the recovery time, thereby being able to quickly predict the time required for the rail transit system to resume normal operation. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] The specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings, wherein:
[0045] Figure 1 The hardware conditions on which the rail transit fault recovery time prediction method of the present invention relies;
[0046] Figure 2 This is a flow chart of the rail transit fault recovery time prediction method of the present invention;
[0047] Figure 3 This is a schematic diagram of the training data of the prediction model;
[0048] Figure 4 It is a structural schematic diagram of the rail transit fault recovery time prediction system of the present invention. DETAILED DESCRIPTION
[0049] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0050] For the daily operation of urban rail transit systems, it is inevitable that emergencies or accidents will cause the entire system to be unable to operate normally. Figure 1As shown in the figure, a neural network-based artificial intelligence model is used to open a variety of multi-modal data interfaces at the input end. For the large-scale raw data detected related to events, environment, weather and other information, these multi-modal big data can be quantified, modeled, and characterized as the input of the neural network model. Thanks to the proposed new characterization data modeling method and neural network architecture, the prediction model will output a prediction of the time required for the rail transit system to resume normal operation. This prediction method is based on the training and learning of the designed model on large-scale historical data, and has sufficient robustness to cope with any unexpected events occurring at any station in an urban rail transit system in any environment.
[0051] Based on the above invention concept, the present invention provides a method for predicting rail transit fault recovery time, such as Figure 2 As shown, the steps include:
[0052] S1. Obtaining historical fault information. The rail transit fault recovery event prediction model of the present invention adopts a supervised learning strategy. Therefore, during the model training stage, a large amount of historical fault information of rail transit fault suspension will be counted and processed, such as Figure 3 As shown, the historical fault information includes: fault information, influencing factors and actual recovery time.
[0053] For fault information, each piece of information contains the name of the event, its category and the site where it occurred. Among them, the categories are divided into track foreign objects (such as objects that affect the normal operation of the track falling into the track), signal failures (such as central control system network anomalies), social factors (such as sudden social events near the station), and natural factors (such as extreme weather). At the same time, for each recorded fault information, the station passenger flow at that moment, the number of stations and lines connected to the station, and the network social media text information near the station location are collected as factors affecting the predicted recovery time. In addition, the actual recovery time (in minutes) of each historical fault information is collected and stored.
[0054] For the actual recovery time of each fault event in the historical record, it is discretized into units of five minutes to convert it into categories, and the cases with a recovery time of more than 90 minutes are classified into the same category. In this way, a total of 19 categories are divided to represent different actual recovery times.
[0055] S2. Perform feature matrixing on the rail transit system stations and lines, and use each element in the two-dimensional matrix to quantitatively store the information in the corresponding map grid.
[0056] Specifically, for the entire urban rail transit system, firstly, the station is characterized in two dimensions from a bird's-eye view according to its actual location in the entire urban space. For the entire urban area map, the longitude and latitude are divided at equal intervals, so that the entire map is divided into H×W grids. Correspondingly, the grids belonging to R are used. H×W Each element in the two-dimensional matrix quantitatively stores the information in the corresponding map grid. For each element in the matrix, it can accurately correspond to a grid area in the city in terms of geographical location. At the same time, the capacity of the information stored in the matrix can be expanded in the channel dimension through a multi-channel matrix. It should be noted that when the map is segmented, in order to ensure a balance between precision and computational efficiency, it is set that no two subway stations are located in the same map grid. On this basis, the grid size is as large as possible to control the data quantization level.
[0057] S3. According to each fault information event in the historical fault information, an event embedding vector, a category embedding vector, an event position feature matrix, and an influencing factor feature are obtained.
[0058] S31, the step of obtaining an event embedding vector according to each fault information event in the historical fault information, comprising:
[0059] Perform unique hot encoding on each fault information event e to obtain an event vector. Assume that the total number of events in the event set ε is N, and use a 0-1 vector to represent the nth event:
[0060]
[0061] Among them, x e Represents such an event vector, for All elements in are zero except the nth element.
[0062] Based on the quantized encoding vector of the event, the fault event embedding matrix θ is adopted e ∈R N×C The event is vectorized. Through matrix multiplication, the event embedding vector representing the event can be obtained and expressed as:
[0063]
[0064] For the fault event embedding matrix, its parameters are randomly initialized and updated through iterative optimization during training.
[0065] S32. For the four categories of fault information (track foreign matter, signal failure, social factors, and natural factors), a category embedding vector is obtained according to each fault information event in the historical fault information, including:
[0066] For each category of the fault information event, a category embedding vector with the same dimension as the event embedding vector is randomly initialized. The event embedding vector is used to characterize the category to which the currently selected fault event belongs. Events of the same category share the same category embedding vector, and the category embedding vector is updated through iterative optimization during training.
[0067] S33. According to each fault information event in the historical fault information, an event location feature is obtained, including:
[0068] For the site where the fault event occurred, based on the matrixing of the grid map, the position of the corresponding element in the two-dimensional matrix can be queried according to the actual coordinate position of the site in the historical fault information;
[0069] Use a 0-1 matrix to represent the site where an event occurs, and represent the event that occurs at the corresponding element position (h,w) of the two-dimensional matrix as the event location feature:
[0070]
[0071] Among them, (h,w)=1, and the rest of the elements are 0.
[0072] Through the above steps S33-S34, for a fault information event, a related event vector x can be quantified and obtained. e , a category embedding vector x c , an event location feature x s For fault event embedding, dimension C is a hyperparameter that can be adjusted during training to facilitate model training. The above three features will be used as part of the input of the rail transit fault recovery event prediction model to help model training and prediction of fault recovery time.
[0073] S34. According to each fault information event in the historical fault information, the influencing factor characteristics are obtained, including:
[0074] The influencing factors are classified into: station passenger flow, the number of stations connected to the current station, the number of lines connected to the current station, and network public opinion with positioning. For the time when the historical fault event occurred in the historical fault information, the station passenger flow, the number of stations connected to the current station, the number of lines connected to the current station, and the network public opinion with positioning at the fault site are collected and quantified;
[0075] According to the passenger flow of the station, the number of stations connected to the current station, the number of routes connected to the current station, and the network public opinion with positioning, we construct the passenger flow feature matrix, the connected station feature matrix, the connected route feature matrix, the optimistic public opinion feature matrix, and the pessimistic public opinion feature matrix.
[0076] Specifically, for the passenger flow at the station, according to the passenger flow density shown by the image recording system of the station, it is discretized and defined as four states: very crowded, crowded, not crowded, and sparsely populated. For these four states, a four-channel grid matrix is used to store the quantitative information of the passenger flow. The passenger flow feature matrix is recorded as x human ∈R 4×H×W Therefore, for the matrix element position (h, w) corresponding to the event site, the traffic feature matrix can extract a four-dimensional feature vector along the channel dimension:
[0077]
[0078] For this four-dimensional feature vector, the corresponding passenger flow state is set to 1, and the rest are assigned 0. Therefore, the passenger flow feature matrix can be used to represent the passenger flow state of the subway station when a fault event occurs.
[0079] For the connectivity of the station where the fault event occurred in the entire rail transit system, two matrices are used to represent the number of connected stations and lines. Specifically, for the number of connected stations, the connected station feature matrix x station ∈R 1×H×W For the elements corresponding to the station locations, their values are set to the number of connected stations, and the remaining elements are set to zero. Similarly, the connected line feature matrix x can be defined as line ∈R 1×H×W , and store the number of connected stations in the corresponding element position.
[0080] The network public opinions with location are all in the form of natural language text (such as microblog content). Therefore, the pre-trained sentiment classifier is used to pre-process the data of these texts. Specifically, each text is input into the sentiment classifier to obtain the binary classification results of the sentiment of the text: optimistic and pessimistic. For each area in the gridded map, an uncertain number of texts will be published, and the optimism and pessimism of these texts are also different. Therefore, for all the texts in an area, the number of texts containing optimism and pessimism is N respectively. positive and N negative , and calculate an optimistic score s positive and a pessimistic score s negative :
[0081]
[0082] By obtaining these scores, two matrices are designed for optimistic and pessimistic public opinion respectively. And store the corresponding scores in the corresponding element positions of the corresponding matrix.
[0083] For the factors affecting the recovery time, a total of five feature matrices can be obtained: the traffic feature matrix x human , connected station feature matrix x station , connected line feature matrix x line , optimistic public opinion feature matrix x emo-p , pessimistic public opinion feature matrix x emo-n , these feature channels have different numbers, but the dimensions other than the channel dimension are the same, and they maintain consistency in the actual positioning information representation. Therefore, these feature matrices are connected together in the channel direction to obtain the influencing factor feature matrix:
[0084] x f =concat[x human ,x station ,x line ,x emo-p ,x emo-n ]∈R 8×H×W
[0085] The influencing factor feature matrix fully characterizes the impact of different factors on fault recovery time. As part of the input of the entire system model, it participates in the training optimization of the model and the prediction of recovery time.
[0086] Influencing factor feature matrix x f It contains very rich multimodal information and is represented by a low number of channels. Therefore, the present invention designs an impression factor neural network based on two-dimensional convolution. f , extracting high-dimensional features of multimodal information as influencing factor features z f ∈R C×H×W :
[0087] z f =f f (x f θ f )
[0088] where θ f For the neural network f f The parameters to be optimized. The influencing factor characteristics here are f Channel number and fault information characteristics m The number of channels of the influencing factor feature matrix remains consistent and much larger than the number of channels of the influencing factor feature matrix to provide a sufficiently robust representation space for information representation. f The output of the influencing factor learning module is used for subsequent predictions.
[0089] S4. Fuse the event embedding vector, category embedding vector, and event position feature matrix to obtain fault information features.
[0090] For the event embedding vector ze and the category embedding vector x c , the fault information it carries has different granularity distinctions. The event embedding vector carries information accurate to each event, while the category embedding vector carries coarse-grained event category information. In order to allow these two types of information to interact better, the present invention uses a random weighting method to fuse the two vectors. Specifically, during training, for each fault event sample, a weight w is randomly sampled on a 0-1 uniform distribution, so that the fused vector z can be obtained. a ∈R 1×C :
[0091] z a = w·z e +(1-w)·x c
[0092] This fusion vector can generalize the fault event information at different granularity levels. Moreover, in the testing phase of the model, it cannot be guaranteed that any fault event is included in the training, so the corresponding z cannot be obtained through embedding. e However, any fault event can be classified into one category. At this time, setting w to 0 can obtain the fusion vector of the event.
[0093] For the event position feature matrix x s , convert the 0-1 matrix into a Gaussian prior feature x g Each element of the matrix is given an activation value, which conforms to a two-dimensional Gaussian distribution with a mean at element 1 as prior information, thus supporting a hypothesis: the closer the fault location is, the more likely it is to be affected. Based on this prior, the learned geographic location likelihood is obtained through a location information neural network f s , the neural network is a lightweight two-dimensional convolutional neural network that outputs the position information feature z s ∈R C×H×W :
[0094] z s =f s (x g θ s )
[0095] Therefore, the fusion vector z f Add the channel dimension vector of the corresponding element position on the position information feature to obtain the fault information feature z m .
[0096] S5. Fault information characteristics m and influencing factors fThe fault information and influencing factors are well represented in high-dimensional space, and the two information are fused by simple and derivable element addition to obtain the information aggregation feature z p ∈R C ×H×W :
[0097] z p =z m +z f
[0098] S6. Using information aggregation features as input and actual recovery time as output, a rail transit fault recovery event prediction model for predicting recovery time is built.
[0099] Information aggregation feature z p In the high-dimensional space, all possible input information for this task is reasonably and efficiently integrated. Therefore, taking it as input, the present invention designs a prediction head f that combines a convolutional neural network and a fully connected neural network. p , with the SoftMax layer as the output layer, the fused information is predicted to obtain the recovery time y required for the input fault event:
[0100] y=f p (z p θ p )
[0101] where θ p For the neural network f p Parameters to be optimized. When obtaining historical fault information, the time has been discretized into units of five minutes. Therefore, the final prediction result of the rail transit fault recovery event prediction model is also discrete (for example: the estimated recovery time for this fault is 20-25 minutes).
[0102] S7. Use historical fault information to train a rail transit fault recovery event prediction model.
[0103] During the training phase, the data used are all historical data with actual recovery time. Here, cross-entropy loss is used as the loss function for training, which can be written as:
[0104]
[0105] At the same time, the rail transit fault recovery event prediction model is recorded as f, and we can get:
[0106] y=f(x e ,x c ,x s ,x f θe ,θ s ,θ f ,θ p )
[0107] θ s It is the parameter to be optimized for the event location feature.
[0108] So use stochastic gradient descent to optimize the loss function of the objective equation to get:
[0109]
[0110] The optimal model is obtained through end-to-end optimization for reasoning.
[0111] S8. Input the real-time fault information into the rail transit fault recovery event prediction model to predict the recovery time. For a new fault, only relevant information (x e ,x s ,x f ) can achieve end-to-end efficient prediction.
[0112] Based on the same inventive concept, the present invention also provides a rail transit fault recovery event prediction system for implementing the rail transit fault recovery time prediction method, such as Figure 4 As shown, including:
[0113] The historical fault information acquisition module is used to acquire historical fault information.
[0114] The fault information learning module is used to obtain an event embedding vector, a category embedding vector, and an event position feature matrix according to each fault information event in the historical fault information, and fuse the event embedding vector, the category embedding vector, and the event position feature matrix to obtain the fault information feature.
[0115] The influencing factor learning module is used to obtain the influencing factor characteristics according to each fault information event in the historical fault information.
[0116] The prediction module is used to add the fault information features and the influencing factor features to obtain the information aggregation features, and build a rail transit fault recovery event prediction model for predicting the recovery time based on the information aggregation features and the prediction head. At the same time, real-time fault information is obtained and input into the rail transit fault recovery event prediction model to predict the predicted recovery time.
[0117] When the model training is completed, the fault information learning module and the influencing factor learning module have very powerful information extraction and aggregation capabilities. The information aggregation features obtained have very high practical value and scalability in high-dimensional space, not only limited to predicting recovery time. Due to the rapid development of large language models in the field of deep learning, the rail transit fault recovery event prediction system can be combined with some natural language descriptions of fault events and information aggregation features obtained from related information, using the neural network model Transformer as a bridge, grafted with the mainstream large language model and performed secondary fine-tuning learning, so that the fault information is highly aligned with the language logic of the large language model. In this way, the rail transit fault recovery event prediction system can not only predict the recovery time when a fault occurs, but also cooperate with the large language model to provide maintenance personnel, station staff, passengers and related personnel with the most direct advice and guidance to deal with sudden faults.
[0118] Please refer to the previous content for the specific implementation process of the rail transit fault recovery event prediction system, which will not be repeated here.
[0119] Based on the same inventive concept, the present invention also provides a computer device, including a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to load the rail transit fault recovery time prediction method as described in the present invention.
[0120] Specifically, the computer device includes: a memory and a processor. The memory and the processor are connected via a bus communication. A computer program is stored in the memory. The computer program can be run on the processor to load the liquidation method of the present invention.
[0121] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.
[0122] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
[0123] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.
Claims
1. A method for predicting rail transit fault recovery time, characterized in that: Includes steps: Get historical fault information; According to each fault information event in the historical fault information, the event embedding vector, the category embedding vector, the event position feature matrix, and the influencing factor features are obtained; The event embedding vector, category embedding vector, and event position feature matrix are fused to obtain fault information features; Add the fault information features and the influencing factor features to obtain the information aggregation features; Using information aggregation features as input and actual recovery time as output, a rail transit fault recovery event prediction model for predicting recovery time is built; Use historical fault information to train the rail transit fault recovery event prediction model; The real-time fault information is input into the rail transit fault recovery event prediction model to predict the recovery time.
2. The rail transit fault recovery time prediction method according to claim 1 is characterized in that: The step of obtaining an event embedding vector according to each fault information event in the historical fault information includes: Perform one-hot encoding on each fault information event to obtain an event vector; The fault event embedding matrix is used to embed the event vector to obtain the event embedding vector.
3. The rail transit fault recovery time prediction method according to claim 2 is characterized in that: According to each fault information event in the historical fault information, a category embedding vector is obtained, including: For each category of the fault information event, a category embedding vector with the same dimension as the event embedding vector is randomly initialized.
4. According to the rail transit fault recovery time prediction method according to claim 1, it is characterized in that: Also includes the steps: The rail transit system stations and line networks are characterized by matrixing, and each element in the two-dimensional matrix is used to quantitatively store the information in the corresponding map grid.
5. The rail transit fault recovery time prediction method according to claim 4 is characterized in that: According to each fault information event in the historical fault information, the event location features are obtained, including: According to the actual coordinate position of the site in the historical fault information, query the position of the corresponding element in the two-dimensional matrix; The events occurring at the corresponding element positions in the two-dimensional matrix are represented as event position features.
6. The rail transit fault recovery time prediction method according to claim 1 is characterized in that: According to each fault information event in the historical fault information, the influencing factor characteristics are obtained, including: According to the time when the historical fault events occurred in the historical fault information, the passenger flow of the station where the fault occurred, the number of stations connected to the current station, the number of lines connected to the current station, and the network public opinion with positioning are collected; According to the passenger flow of the station, the number of stations connected to the current station, the number of routes connected to the current station, and the network public opinion with positioning, the passenger flow feature matrix, the connected station feature matrix, the connected route feature matrix, the optimistic public opinion feature matrix, and the pessimistic public opinion feature matrix are constructed; Connect the passenger flow feature matrix, the connected station feature matrix, the connected line feature matrix, the optimistic public opinion feature matrix, and the pessimistic public opinion feature matrix according to the channel direction to obtain the influencing factor feature matrix; The impression factor neural network is used to extract the influencing factor features from the influencing factor feature matrix.
7. The rail transit fault recovery time prediction method according to claim 1 is characterized in that: The steps of fusing the event embedding vector, the category embedding vector, and the event position feature matrix to obtain the fault information feature include: Randomly sample a weight from a 0-1 uniform distribution; Multiply the event embedding vector and the category embedding vector by the weights respectively and then add them together to obtain the fusion vector; The event position feature matrix is converted into Gaussian prior features; Based on Gaussian prior features, position information features are output through a position information neural network; The fusion vector is added to the position information feature to obtain the fault information feature.
8. The rail transit fault recovery time prediction method according to claim 1 is characterized in that: The cross entropy loss is used as the loss function for training.
9. A rail transit fault recovery event prediction system, characterized in that: A method for predicting rail transit fault restoration time according to any one of claims 1 to 8, comprising: A historical fault information acquisition module is used to acquire historical fault information; The fault information learning module is used to obtain an event embedding vector, a category embedding vector, and an event position feature matrix according to each fault information event in the historical fault information, and fuse the event embedding vector, the category embedding vector, and the event position feature matrix to obtain the fault information feature; An influencing factor learning module is used to obtain influencing factor characteristics based on each fault information event in the historical fault information; The prediction module is used to add the fault information features and the influencing factor features to obtain the information aggregation features, and build a rail transit fault recovery event prediction model for predicting the recovery time based on the information aggregation features and the prediction head. At the same time, real-time fault information is obtained and input into the rail transit fault recovery event prediction model to predict the predicted recovery time.
10. A computer device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the rail transit fault recovery time prediction method as described in any one of claims 1 to 8 when executing the program stored in the memory.