Railway facility health state prediction method and device

Abstract

The application is applied to the field of railway facility maintenance, and provides a health state prediction method and device for railway facilities, which comprises the following steps: acquiring multi-modal data of railway facilities; updating a corresponding historical space-time knowledge graph of the railway facilities by using the multi-modal data to obtain a target space-time knowledge graph, wherein the space-time knowledge graph comprises a plurality of nodes and multi-modal features of the nodes, and the types of the nodes are entity nodes, process nodes or event nodes; constructing a graph adjacency bias matrix according to the connection relationship between each entity node, process node and / or event node in the target space-time knowledge graph, and constructing fusion features according to the multi-modal features of each node; processing the fusion features and the graph adjacency bias matrix by using a prediction model to obtain the health state of the railway facilities. The space-time knowledge graph containing entity nodes, process nodes and event nodes is dynamically maintained, so that the health state predicted according to the space-time knowledge graph is more accurate.

Description

Technical Field

[0001] This invention relates to the field of railway facility maintenance technology, and in particular to a method and apparatus for predicting the health status of railway facilities. Background Technology

[0002] The health status of railway facilities is directly related to operational safety. With the enrichment of monitoring methods, railway operation and maintenance systems can acquire a large amount of sensor data. In order to achieve the transformation from passive maintenance to predictive maintenance, predictive maintenance is related to the effective use of this multi-source heterogeneous spatiotemporal data to accurately predict the future health status and degradation trends of railway facilities and equipment.

[0003] To meet the needs of predictive maintenance, existing technical solutions typically employ graph-based prediction methods. For example, this involves first collecting monitoring data from sensors and image data from manual inspections, and then constructing a traditional knowledge graph based on the physical topology of the railway equipment. In this knowledge graph, physical devices are abstracted as nodes, and spatial connections or hierarchical relationships between devices are abstracted as edges. Then, multimodal monitoring data is assigned as attribute features to the corresponding physical nodes and input into a graph neural network or time-series prediction model. By aggregating the features of neighboring nodes and analyzing historical time-series sequences, the model outputs an assessment of the physical node's health status over a future period.

[0004] The problems with this approach include: when dealing with complex railway scenarios, it is unable to cope with the frequent discrete emergencies and gradual changes with continuous evolution characteristics in the real world, and it is easy to miss key disaster-causing information, ultimately resulting in inaccurate health status prediction results. Summary of the Invention

[0005] This invention provides a method and apparatus for predicting the health status of railway facilities, which addresses the shortcomings of inaccurate health status prediction results in the prior art and improves the accuracy of health status prediction results for railway facilities.

[0006] This invention provides a method for predicting the health status of railway facilities, comprising the following steps: Acquire multimodal data of railway facilities; The historical spatiotemporal knowledge graph corresponding to railway facilities is updated using multimodal data to obtain the target spatiotemporal knowledge graph. The target spatiotemporal knowledge graph includes several nodes and their multimodal features. The types of each node are entity nodes, process nodes, or event nodes. Based on the connection relationships between entity nodes, process nodes and / or event nodes in the target spatiotemporal knowledge graph, a graph adjacency bias matrix is ​​constructed, and fusion features are constructed based on the multimodal features of each node; The health status of railway facilities is obtained by processing the fusion features and graph adjacency bias matrix using a prediction model. The prediction model is trained using the sample graph adjacency bias matrix, sample fusion features and corresponding health status labels.

[0007] According to the present invention, a method for predicting the health status of railway facilities is provided, which uses a prediction model to process fused features and a graph adjacency bias matrix to obtain the health status of the railway facilities, including: The fusion features are linearly mapped using a prediction model to obtain the query matrix, key matrix, and value matrix; The query matrix and the transpose of the key matrix are multiplied by a dot product. The result of the dot product is then fused with the graph adjacency bias matrix and normalized to obtain the attention weight matrix. Multiplying the attention weight matrix by the value matrix yields the self-attention features; The health status of railway facilities is obtained based on self-attention characteristics.

[0008] According to the method for predicting the health status of railway facilities provided by the present invention, the multimodal features of each node include sub-multimodal features at multiple time points, and a fused feature is constructed based on the multimodal features of each node, including: The embedding vector is obtained by acquiring the temporal position encoding of each sub-multimodal feature of each node at the corresponding time point and the node type encoding of each node. For each node, the spatiotemporal joint features of the node are obtained by adding the node's sub-multimodal features, the temporal position encoding of the time point corresponding to each sub-multimodal feature, and the node's embedding vector. The spatiotemporal joint features of each node are fused to obtain the fused features.

[0009] According to the present invention, a method for predicting the health status of railway facilities is provided, wherein each process node in the target spatiotemporal knowledge graph is used to record the dynamic change process of the state of one of the entity nodes over time, and each event node is used to record an event that can cause the state of at least some entity nodes to change.

[0010] According to the present invention, a method for predicting the health status of railway facilities is provided, which updates the historical spatiotemporal knowledge graph corresponding to the railway facilities using multimodal data to obtain a target spatiotemporal knowledge graph, including: Entity structure recognition and current state extraction are performed on multimodal data to obtain several entity structures and their current states; events are extracted from multimodal data to obtain several events. For each entity structure, if there is a corresponding entity node in the historical spatiotemporal knowledge graph and the corresponding entity node is connected to a corresponding process node, the multimodal features of the process node are updated according to the current state; or, if there is no corresponding entity node in the historical spatiotemporal knowledge graph, an entity node corresponding to the entity structure is added to the historical spatiotemporal knowledge graph. For each event, if a corresponding event node exists in the historical spatiotemporal knowledge graph, the multimodal features of the corresponding event node are updated according to the event; or if a corresponding event node does not exist in the historical spatiotemporal knowledge graph, a new event node is added to the historical spatiotemporal knowledge graph and a connection relationship is established between the event node and at least one node.

[0011] According to the present invention, a method for predicting the health status of railway facilities, wherein the health status includes a health index, and after processing the fusion features and graph adjacency bias matrix using a prediction model to obtain the health status of the railway facilities, the method further includes: If the health index of the target entity node is determined to be less than the preset health index, multimodal features of process nodes related to the target entity node are extracted from the target spatiotemporal knowledge graph to construct event pairs; The event pairs are used as the problems to be reasoned in the root cause analysis model. The root cause analysis model is used to determine the root cause analysis results of the event pairs based on the target spatiotemporal knowledge graph and the domain causal graph. The root cause analysis model is trained using sample event pairs, sample domain causal graphs and root cause analysis result labels.

[0012] According to the method for predicting the health status of railway facilities provided by the present invention, the domain causal graph is obtained based on the following method: An initial domain causal graph constructed based on prior knowledge; Obtain historical operation and maintenance data of several railway facilities and mine potential causal relationships based on causal discovery algorithms; The initial domain causal graph is fused with the potential causal relationships to obtain the domain causal graph.

[0013] According to the present invention, a method for predicting the health status of railway facilities is provided, which acquires multimodal data of railway facilities, including: Acquire the first data collected by drones on railway facilities; A global risk heat map of the railway facility is determined based on the first data. The global risk heat map is determined based on the defect type, defect density and confidence level of each area in the railway facility. Based on the overall risk heat map, areas to be investigated in detail were identified from the regions. Control the ground monitoring device to collect secondary data from each area to be investigated in detail; Multimodal data is constructed based on the first and second data.

[0014] According to a method for predicting the health status of railway facilities provided by the present invention, a ground monitoring device is controlled to collect second data for each area to be investigated in detail, including: Obtain the current location of the ground monitoring device and the target location of the area to be investigated in detail; The movement path of the ground monitoring device is determined based on the current location, the target location, and the historical spatiotemporal knowledge graph; A data acquisition task is generated and sent to the ground monitoring device, so that the ground monitoring device can move to the target location according to the movement path in the data acquisition task and collect the second data.

[0015] The present invention also provides a health status prediction device for railway facilities, comprising: The multimodal data acquisition module is used to acquire multimodal data of railway facilities; The spatiotemporal knowledge graph update module is used to update the historical spatiotemporal knowledge graph corresponding to railway facilities using multimodal data to obtain the target spatiotemporal knowledge graph. The target spatiotemporal knowledge graph includes several nodes and their multimodal features. The type of each node is an entity node, a process node, or an event node. The data construction module is used to construct a graph adjacency bias matrix based on the connection relationships between entity nodes, process nodes and / or event nodes in the target spatiotemporal knowledge graph, and to construct fusion features based on the multimodal features of each node. The health status prediction module is used to process the fusion features and graph adjacency bias matrix using the prediction model to obtain the health status of railway facilities. The prediction model is trained using the sample graph adjacency bias matrix, sample fusion features and corresponding health status labels.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described methods for predicting the health status of railway facilities.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for predicting the health status of railway facilities.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the above-described methods for predicting the health status of railway facilities.

[0019] This invention provides a method and apparatus for predicting the health status of railway facilities. Compared with existing technologies that only construct static knowledge graphs for the physical structures of railway facilities, this solution first updates the historical spatiotemporal knowledge graph based on multimodal data collected from the railway facilities. Then, it constructs a graph adjacency bias matrix and fusion features based on the updated target spatiotemporal knowledge graph. Next, the prediction model processes the fusion features and the graph adjacency bias matrix to obtain the corresponding health status. In this process, by constructing a spatiotemporal knowledge graph containing entity nodes, process nodes, and event nodes, and dynamically updating the spatiotemporal knowledge graph using multimodal data, the addition of the graph adjacency bias matrix enables the prediction model to directly capture the complex spatiotemporal relationships between entities, events, and processes, thereby improving the accuracy of predicting the health status of railway facilities under complex and ever-changing environments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is one of the flowcharts illustrating the method for predicting the health status of railway facilities provided by the present invention.

[0022] Figure 2 This is the second flowchart illustrating the method for predicting the health status of railway facilities provided by this invention.

[0023] Figure 3 This is a schematic diagram of the health status prediction device for railway facilities provided by the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] The following is combined Figures 1 to 2 This invention describes the method for predicting the health status of railway facilities. Figure 1This is one of the flowcharts illustrating the health status prediction method for railway facilities provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step 101: Obtain multimodal data of railway facilities.

[0027] Railway facilities can include various types of infrastructure within the railway system, such as, but not limited to, bridge piers, bridge decks, tunnels, roadbeds, tracks, overhead contact lines, and signaling equipment. Multimodal data on railway facilities can include multimodal data on all infrastructure within the area requiring monitoring.

[0028] Multimodal data can be acquired through various sensing devices. For example, it can be acquired collaboratively through an aerial wide-area sensing platform and a ground-based fine-sensing platform. The aerial wide-area sensing platform can include unmanned aerial vehicles (UAVs) equipped with multimodal payloads such as high-definition visible light cameras, infrared thermal imagers, and lidar. The ground-based fine-sensing platform can include tracked unmanned ground vehicles equipped with intelligent measurement cameras with millimeter-level measurement accuracy and high-density laser scanners, or fixed non-contact visual monitoring stations deployed along the route for collaborative acquisition.

[0029] For example, multimodal data of bridge piers in railway facilities may include: high-definition visible light images taken by drones, infrared thermal images, lidar point cloud data, and high-resolution images taken at close range by unmanned ground vehicles.

[0030] Step 102: Update the historical spatiotemporal knowledge graph corresponding to the railway facilities using multimodal data to obtain the target spatiotemporal knowledge graph.

[0031] The historical and target spatiotemporal knowledge graphs include several nodes and their multimodal features. Each node can be an entity node, a process node, or an event node. Each entity node is constructed for a railway facility entity or its components. For example, a node might be constructed for a bridge pier, a specific section of track, or a catenary support. Each process node is constructed for the dynamic process of the state evolution of a particular entity node over time. For example, a process node associated with a bridge pier entity node could include a corrosion process node or a settlement process node, used to record the continuous changes in the pier's corrosion or settlement state. Event nodes can be constructed for discrete events that may affect the state of one or more entity nodes. Examples include rainstorm event nodes, overloaded train passage event nodes, and earthquake event nodes. Each node can be associated with a multimodal feature, which can include sub-multimodal data at several time points; for example, the multimodal features of an entity node gradually increase over time.

[0032] For example, a spatiotemporal knowledge graph is a graph-like data structure used to structurally represent railway facilities and their interrelationships and state evolution processes, which can be formally represented as follows: , Represents a set of nodes. Represents an edge set. Represents a set of relation types. Represents a set of timestamps, where, , Represents a set of entity nodes. Represents a set of process nodes. This represents a set of event nodes. The type of relationship between two nodes can be spatial and structural (e.g., containment, adjacency, support), temporal and process (e.g., occurs before, precedes / follows), causal and influence (e.g., induces, accelerates degradation), attribute and characteristic (e.g., exhibits), etc.

[0033] Historical spatiotemporal knowledge graphs refer to knowledge graphs that existed before the acquisition of the current round of multimodal data and recorded the historical state of railway facilities. For example, the target spatiotemporal knowledge graph determined in the previous health status prediction process. In other words, the target spatiotemporal knowledge graph determined in the current health status prediction process serves as the historical spatiotemporal knowledge graph for the next round of health status prediction process.

[0034] Step 102 can be implemented by integrating the information parsed from the newly acquired multimodal data into the historical spatiotemporal knowledge graph. For example, if the new multimodal data detects an increase in the width of a crack in a bridge pier, the multimodal features of the bridge pier entity node and its associated crack development process nodes are updated. If a rainstorm is detected, a rainstorm event node can be added to the historical spatiotemporal knowledge graph, and connections can be established between it and entity nodes such as bridge piers and paths within the affected area. After the update, the resulting target spatiotemporal knowledge graph contains the latest multimodal data.

[0035] Step 103: Based on the connection relationships between entity nodes, process nodes and / or event nodes in the target spatiotemporal knowledge graph, construct a graph adjacency bias matrix, and construct fusion features based on the multimodal features of each node.

[0036] The graph adjacency bias matrix can be constructed as follows: Based on the connection relationships between nodes in the target spatiotemporal knowledge graph, a bias matrix is ​​built. For example, if two nodes have a direct connection in the knowledge graph, the value of their corresponding element in the graph adjacency bias matrix can be set to 0; if two nodes do not have a direct connection, the value of the corresponding element can be set to a preset large negative number (e.g., -). This allows the prediction model to focus on nodes with direct relationships and ignore nodes without direct relationships.

[0037] One way to construct fusion features based on the multimodal features of each node is to fuse the multimodal features of all nodes in the target spatiotemporal knowledge graph to obtain fusion features.

[0038] Alternatively, the fusion feature can be constructed based on the multimodal features of each node as follows: Using one entity node as the target entity node, select nodes with connections to the target entity node from the target knowledge graph as reference nodes. Then, fuse the multimodal features of the target entity node and each reference node to obtain the fusion feature. The nodes with connections to the target entity node can be nodes with direct or indirect connections. In this case, the graph adjacency bias matrix is ​​either the graph adjacency bias matrix of the entire target spatiotemporal knowledge graph, or it can be the graph adjacency bias matrix of the local knowledge graph formed by the target entity node and each reference node.

[0039] In other words, this solution can determine the health status of all entity nodes at once based on the target spatiotemporal knowledge graph, or predict the corresponding health status for a specific entity node.

[0040] Step 104: The fusion features and graph adjacency bias matrix are processed using the prediction model to obtain the health status of the railway facilities.

[0041] A predictive model is a trained machine learning model that takes fused features and a graph adjacency bias matrix as input and outputs a health status. For example, such a predictive model could be a combination of a Spatiotemporal Graph Transformer (STGT) model and linear layers such as a Multi-Layer Perceptron (MLP).

[0042] The multilayer perceptron may include an input layer, a first hidden layer, a second hidden layer, and an output layer. The first hidden layer, the second hidden layer, and the output layer can all be fully connected layers. The first hidden layer and the second hidden layer may use the ReLU activation function, and the output layer may use the Sigmoid activation function.

[0043] For example, the fused features and graph adjacency bias matrix corresponding to the target entity node are used as inputs to the STGT model, and the hidden state output by the STGT model is used as input to the multilayer perceptron. The multilayer perceptron outputs the health status. The health status can include the predictive health index (PHI) of the entity infrastructure corresponding to the target entity node at the time of the aforementioned multimodal data acquisition, or it can also include the health index at multiple future times, i.e., the future health trend. The health index value ranges between [0,1], where 1 represents complete health and 0 represents severe failure.

[0044] The prediction model is trained using the sample graph adjacency bias matrix, sample fusion features, and corresponding health status labels. For example, the prediction model can be trained using supervised learning. The training dataset includes the sample graph adjacency bias matrix, sample fusion features, and corresponding health status labels from historical data.

[0045] In this embodiment, compared to the prior art which only constructs a static knowledge graph for the physical structures of railway facilities, this solution first updates the historical spatiotemporal knowledge graph based on multimodal data collected from railway facilities. Then, it constructs a graph adjacency bias matrix and fusion features based on the updated target spatiotemporal knowledge graph. Next, the prediction model processes the fusion features and the graph adjacency bias matrix to obtain the corresponding health status. In this process, by constructing a spatiotemporal knowledge graph containing entity nodes, process nodes, and event nodes, and dynamically updating the spatiotemporal knowledge graph using multimodal data, the addition of the graph adjacency bias matrix enables the prediction model to directly capture the complex spatiotemporal relationships between entities, events, and processes, thereby improving the accuracy of predicting the health status of railway facilities under complex and ever-changing environments.

[0046] In some embodiments, step 101 described above may include, for example: Figure 2 The following steps are shown: Step 201: Obtain the first data collected by the drone on the railway facilities.

[0047] In this embodiment, the first data is data with a wider coverage than the second data described below, such as image maps, heat maps and preliminary point cloud models of the entire railway facility.

[0048] Step 202: Determine the global risk heat map of the railway facilities based on the first data.

[0049] The global risk heat map is determined based on the defect type, defect density, and confidence level of each area in the railway facility.

[0050] For example, defects are identified based on the first data, and then a global heatmap is constructed based on the defect type, defect density, and confidence level of each region. For example, a global risk heatmap. , This refers to the number of grid regions in the global heatmap, where each grid region corresponds to an area within a railway facility. (Global Risk Heatmap) The feature vector for each region contains the defect type, defect density, and confidence level.

[0051] Step 203: Identify the areas to be investigated in detail from the regions based on the global risk heat map.

[0052] A global risk heatmap based on self-attention can be used to determine the probability of each region being selected as a region for detailed investigation. Specifically, a multi-head self-attention mechanism can be used to weight different regions in the global risk heatmap to obtain the attention score of each region, so as to capture the spatial correlation between risk regions (for example, the settlement risk of upstream bridge piers may affect downstream areas).

[0053] For example, calculating the region in the global risk heatmap Attention Score The calculation is as follows: In this formula and The weight matrix is ​​a learnable matrix. For the region eigenvectors, This represents a global risk heatmap. Indicates the region The corresponding query matrix, This indicates that based on the global risk heatmap The determined key matrix, This indicates that based on the global risk heatmap The dimension of each key vector in the determined key matrix. Represents a global risk heatmap Central region Attention score.

[0054] Then, the top n regions with the highest attention scores can be selected as the regions to be investigated in detail.

[0055] Step 204: Control the ground monitoring device to collect second data for each area to be investigated in detail.

[0056] The ground monitoring device can be a tracked unmanned ground vehicle equipped with intelligent measuring cameras with millimeter-level measurement accuracy and a high-density laser scanner. The second set of data can be data from the intelligent measuring cameras and / or the high-density laser scanner of the area to be investigated in detail.

[0057] In some embodiments, step 204 may include the following steps: First, obtain the current location of the ground monitoring device and the target location of the area to be investigated in detail.

[0058] Secondly, the movement path of the ground monitoring device is determined based on the current location, the target location, and the historical spatiotemporal knowledge graph.

[0059] For example, path risk information, such as historical fault points, steep slopes, and narrow passes, can be extracted from the historical spatiotemporal knowledge graph.

[0060] The current location, target location, and path risk information are used as an attention-based path planning network to generate a movement path.

[0061] For example, to integrate the costs of multiple paths, an attention layer is used to dynamically learn the weights of each cost. At each step of the path search, the overall cost is calculated as follows: In this formula, the weights The path planning network can dynamically determine the path based on environmental data such as the remaining battery power of the ground monitoring device and weather conditions. This represents cost. Distance can be considered as the distance to move from the current position to the next step, slope is the gradient between the current position and the next step, and risk is the risk value of moving from the current position to the next step.

[0062] Then, a data acquisition task is generated and sent to the ground monitoring device, so that the ground monitoring device can move to the target location according to the movement path in the data acquisition task and collect the second data. That is, the attention-based hierarchical reinforcement learning (A-HRL) scheduling algorithm realizes the decision-making of the area to be investigated at the task level and the path planning at the execution level.

[0063] Step 205: Construct multimodal data based on the first and second data.

[0064] For example, the first data and the second data are spatiotemporally aligned and fused to obtain multimodal data.

[0065] In this embodiment, a closed-loop method is proposed, which first uses UAV observation, then analyzes the first data collected by the UAV, and then uses ground monitoring devices to collect and verify the data. This method can improve the efficiency and purposefulness of data collection.

[0066] In some embodiments, each process node in the target spatiotemporal knowledge graph is used to record the dynamic change process of the state of one of the entity nodes over time, and each event node is used to record an event that can cause a change in the state of at least some entity nodes.

[0067] As mentioned above, each entity node corresponds to an infrastructure unit, while process nodes record changes in its vital signs over time. For example, for an entity node named Pier B03, a process node named Pier B03 - Settlement Process can be associated with it. The multimodal features of this process node can be time-series data, recording the millimeter-level settlement values ​​of the pier obtained through lidar or visual measurements over the past few years. Similarly, a process node named Pier B03 - Corrosion Process can be associated, whose multimodal features can record images of the pier's steel reinforcement corrosion area, color, etc., changing over time.

[0068] As mentioned above, event nodes are used to represent discrete events that may impact or affect railway facilities. An event node can be connected to one or more entity nodes simultaneously. For example, an event node representing a rainstorm event on day C of month B of year A in railway section D may have multimodal features including data such as rainfall, duration, and affected area. In the spatiotemporal knowledge graph, this event node may establish connections with all entity nodes (such as multiple bridge pier nodes and roadbed nodes) in railway section D.

[0069] In this embodiment, by constructing process nodes and event nodes, it is possible not only to statically describe what infrastructure exists, but also to dynamically describe what changes are occurring in the infrastructure and what external events are affecting these changes, thereby enhancing the accuracy of predicting the health status of railway facilities.

[0070] In some embodiments, the above-mentioned updating of the historical spatiotemporal knowledge graph corresponding to railway facilities using multimodal data to obtain the target spatiotemporal knowledge graph includes: First, entity structure identification and current state extraction are performed on the multimodal data to obtain several entity structures and their current states. Event extraction is then performed on the multimodal data to obtain several events. For example, using computer vision algorithms to process drone images, the entity structure of bridge pier B03 can be identified, and its current state, such as the main crack width being 5 mm, can be extracted. Simultaneously, by analyzing the accessed meteorological data, an event can be extracted, such as a rainstorm occurring on day C of month B of year A, with a daily rainfall of 120 mm. The extracted entity structures, states, and events can then be semantically correlated with a structured railway operation and maintenance knowledge base, allowing for the filtering of each entity, event, and state.

[0071] Secondly, for each entity structure, node update or addition operations are performed. For example, if a corresponding entity node exists in the historical spatiotemporal knowledge graph and is connected to a corresponding process node, the multimodal features of the process node are updated according to the current state; or, if a corresponding entity node does not exist in the historical spatiotemporal knowledge graph, a new entity node corresponding to the entity structure is added to the historical spatiotemporal knowledge graph. Alternatively, if a corresponding entity node exists in the historical spatiotemporal knowledge graph but is not connected to a corresponding process node, the corresponding process node can be created based on the current state.

[0072] Then, for each event, a node update or addition operation is performed. For example, if a corresponding event node exists in the historical spatiotemporal knowledge graph, the multimodal features of the corresponding event node are updated according to the event; or if a corresponding event node does not exist in the historical spatiotemporal knowledge graph, a new event node is added to the historical spatiotemporal knowledge graph, and a connection relationship is established between the event node and at least one other node. For example, establishing the connection relationship between the event node and at least one other node can be done by establishing a connection relationship between the event node and each entity node within the scope of its influence, based on the scope of influence of the event node.

[0073] In this embodiment, by continuously transforming fragmented multimodal data streams into structured knowledge, the accuracy of subsequent health status prediction is improved.

[0074] In some embodiments, the multimodal features of each node include sub-multimodal features at multiple time points. For example, the first... Each node at a given time point The sub-multimodal features are The above method of constructing fused features based on the multimodal features of each node may include the following steps: First, obtain the temporal location encoding of each sub-multimodal feature of each node at each time point, and the embedding vector obtained by encoding the node type of each node. For example, time points Time position encoding , No. The node type of each node is encoded as follows: .

[0075] Secondly, for each node, the spatiotemporal joint features of the node are obtained by summing the sub-multimodal features of the node, the temporal location encoding of the corresponding time point of each sub-multimodal feature, and the embedding vector of the node. .

[0076] Then, the spatiotemporal joint features of each node are fused to obtain the fused features.

[0077] For example, this embodiment uses one of the entity nodes. As the target entity node, nodes with connections to the target entity node are selected from the target knowledge graph as reference nodes. The spatiotemporal joint features of the target entity node and each reference node are fused to obtain the fused features. For example, each target entity node can be the entity node corresponding to one of the regions to be investigated in detail mentioned above.

[0078] In some application scenarios, environmental covariates can also be obtained. The spatiotemporal joint features of each node are fused to obtain the fused features. Environmental covariates It can include environmental monitoring data such as ambient temperature and humidity.

[0079] In this embodiment, by incorporating time sequence information and node type into the input data of the prediction model, the prediction model can better understand and distinguish the state of different nodes at different points in time, thereby improving the accuracy of railway facility health status prediction.

[0080] In some embodiments, step 104 above may include the following steps: The fusion features are linearly mapped using a prediction model to obtain a query matrix, a key matrix, and a value matrix. The query matrix and the transpose of the key matrix are then subjected to a dot product operation. The result of the dot product operation is fused with the graph adjacency bias matrix and normalized to obtain an attention weight matrix. The attention weight matrix is ​​multiplied by the value matrix to obtain self-attention features. Based on the self-attention features, the health status of the railway facilities is obtained.

[0081] For example, the attention weight matrix is ​​calculated. The method can be: In this formula This represents the query matrix obtained by linearly mapping the fused features using a prediction model. This represents the key matrix obtained by linearly mapping the fused features using a prediction model. Indicates transpose. This represents the value matrix obtained by linearly mapping the fused features using a prediction model. This formula represents the graph adjacency bias matrix. This represents the dimension of each key vector in the key matrix obtained by linearly mapping the fused features using a prediction model.

[0082] Because the value corresponding to the unconnected node pair in the graph adjacency bias matrix is ​​a large negative number, the attention weight matrix acquisition process can reduce the correlation score between nodes that are not directly related. This is equivalent to using the structural information of the target spatiotemporal knowledge graph to guide the correlation between nodes, forcing the prediction model to focus on nodes that have connections in the target spatiotemporal knowledge graph.

[0083] After processing by the Softmax function, the correlation scores that were previously given large negative numbers will approach 0, while other scores will be normalized to a probability distribution that sums to 1. The final matrix is ​​the attention weight matrix.

[0084] Among them, the self-attention feature can be the hidden state output by the STGT model mentioned above. Based on the self-attention feature, the health status of railway facilities can be obtained by using the self-attention feature as the input of the multilayer perceptron and the multilayer perceptron outputting the health status.

[0085] In this embodiment, by incorporating the structural information of the target spatiotemporal knowledge graph into the self-attention mechanism of the prediction model, the prediction model can refer to the latest state of the infrastructure when making health status predictions, thereby improving the accuracy of health status predictions.

[0086] In some embodiments, the health status includes a health index. After processing the fused features and the graph adjacency bias matrix using a prediction model to obtain the health status of the railway facility, the method further includes: If the health index of a target entity node is determined to be less than a preset health index, multimodal features of process nodes related to the target entity node are extracted from the target spatiotemporal knowledge graph to construct event pairs. For example, an early warning is triggered when the health index is less than 0.3, and emergency maintenance is triggered when it is less than 0.15. The preset health index can be 0.3. For example, the event pair can include multimodal features of the entity node and its process nodes, such as the final crack width of a bridge pier increasing from 2 mm to 5 mm within one month. The entity node can be the aforementioned target entity node.

[0087] The event pairs are used as the problems to be reasoned in the root cause analysis model. The root cause analysis model is used to determine the root cause analysis results of the event pairs based on the target spatiotemporal knowledge graph and the domain causal graph.

[0088] The root cause analysis model can be a large language model. The root cause analysis results include causal paths and the contribution of each causal factor within those paths to the health index. The root cause analysis model is trained using sample event pairs, a causal graph of the sample domain, and labels from the root cause analysis results.

[0089] For example, sample event pairs can be automatically extracted from a knowledge graph. The sample domain causal graph can be the same as or different from the aforementioned domain causal graph. The causal path labels in the root cause analysis result labels can be the causal transmission paths marked by experts for the sample event pairs in the sample domain causal graph. The contribution values ​​in the root cause analysis result labels can be the quantitative contribution of each causal factor in the marked causal transmission path to the sample event pair, automatically calculated based on a causal inference toolkit. Combining the above sample event pairs, sample domain causal graph, and root cause analysis result labels yields a structured question-and-answer sample.

[0090] In this embodiment, the root cause analysis model is given the ability to perform structured causal reasoning by fine-tuning the Causal Chain-of-Thought (CCoT) method.

[0091] In some embodiments, the domain causal graph is obtained by: constructing an initial domain causal graph based on prior knowledge; acquiring historical operation and maintenance data of several railway facilities and mining potential causal relationships based on causal discovery algorithms; and fusing the initial domain causal graph with the potential causal relationships to obtain the domain causal graph.

[0092] Prior knowledge can come from industry standards, technical manuals, academic research, and the practical experience of senior engineers. For example, prior knowledge can be normative documents such as the "Railway Technical Management Regulations". An initial domain causal graph can be constructed based on the "Railway Technical Management Regulations". For example, this initial domain causal graph can include 25 core variables, such as foundation moisture content, uneven settlement, and track irregularity, and 48 basic causal edges.

[0093] Historical maintenance data can be records of malfunctions and incidents, while causal discovery algorithms are statistical or machine learning algorithms that can infer causal structures between variables from observed data. For example, the classic PC causal discovery algorithm can be used. After setting a significance level, the causal discovery algorithm can be used to mine causal relationships from historical maintenance data through conditional independence tests, thus constructing a data-driven graph.

[0094] One method for fusing the initial domain causal graph with potential causal relationships to obtain the domain causal graph is as follows: Based on the initial domain causal graph, the following fusion rules are applied: all edges in the initial domain causal graph are retained as mandatory priors; then, new causal relationships discovered in the data-driven graph are added to the initial domain causal graph only if their confidence level is higher than a pre-set confidence level. Conflicting edges in the data-driven graph with directions inconsistent with those in the initial domain causal graph can be submitted to an expert panel for manual voting to determine the final domain causal graph.

[0095] In this embodiment, a domain causal graph that combines the rigor of expert knowledge with data-driven new discoveries can be generated, greatly improving the accuracy of causal reasoning.

[0096] This method can be applied to electronic devices in railway facility management systems, which also include drones and ground monitoring devices that communicate with these devices. Drones can form a wide-area aerial sensing platform, while ground monitoring devices can form a fine-grained ground sensing platform. The railway facility management system can be considered a data twin system of railway facilities. By incorporating dynamic spatiotemporal knowledge graphs and domain causal graphs, it is not only a mirror of the physical world but also capable of discerning the essence of faults and predicting future trends.

[0097] The following describes the railway facility health status prediction device provided by the present invention. The railway facility health status prediction device described below can be referred to in correspondence with the railway facility health status prediction method described above. For example... Figure 3 As shown, the railway facility health status prediction device 300 may include the following steps: The multimodal data acquisition module 301 is used to acquire multimodal data of railway facilities; The spatiotemporal knowledge graph update module 302 is used to update the historical spatiotemporal knowledge graph corresponding to railway facilities using multimodal data to obtain the target spatiotemporal knowledge graph. The target spatiotemporal knowledge graph includes several nodes and their multimodal features. The type of each node is an entity node, a process node, or an event node. The data construction module 303 is used to construct a graph adjacency bias matrix based on the connection relationships between entity nodes, process nodes and / or event nodes in the target spatiotemporal knowledge graph, and to construct fusion features based on the multimodal features of each node. The health status prediction module 304 is used to process the fusion features and graph adjacency bias matrix using the prediction model to obtain the health status of railway facilities. The prediction model is trained using the sample graph adjacency bias matrix, sample fusion features and corresponding health status labels.

[0098] According to the present invention, a health status prediction device 300 for railway facilities includes a health status prediction module 304 that processes fused features and a graph adjacency bias matrix using a prediction model to obtain the health status of the railway facilities, including: The fusion features are linearly mapped using a prediction model to obtain the query matrix, key matrix, and value matrix; The query matrix and the transpose of the key matrix are multiplied by a dot product. The result of the dot product is then fused with the graph adjacency bias matrix and normalized to obtain the attention weight matrix. Multiplying the attention weight matrix by the value matrix yields the self-attention features; The health status of railway facilities is obtained based on self-attention characteristics.

[0099] According to the present invention, a railway facility health status prediction device 300 includes multiple time-point sub-multimodal features in the multimodal features of each node. A data construction module 303 constructs fused features based on the multimodal features of each node, including: The embedding vector is obtained by acquiring the temporal position encoding of each sub-multimodal feature of each node at the corresponding time point and the node type encoding of each node. For each node, the spatiotemporal joint features of the node are obtained by adding the node's sub-multimodal features, the temporal position encoding of the time point corresponding to each sub-multimodal feature, and the node's embedding vector. The spatiotemporal joint features of each node are fused to obtain the fused features.

[0100] According to the present invention, a railway facility health status prediction device 300 is provided, wherein each process node in the target spatiotemporal knowledge graph is used to record the dynamic change process of the state of one of the entity nodes over time, and each event node is used to record an event that can cause the state of at least some entity nodes to change.

[0101] According to the present invention, a railway facility health status prediction device 300 includes a spatiotemporal knowledge graph update module 302 that uses multimodal data to update the historical spatiotemporal knowledge graph corresponding to the railway facility to obtain a target spatiotemporal knowledge graph, comprising: Entity structure recognition and current state extraction are performed on multimodal data to obtain several entity structures and their current states; events are extracted from multimodal data to obtain several events. For each entity structure, if there is a corresponding entity node in the historical spatiotemporal knowledge graph and the corresponding entity node is connected to a corresponding process node, the multimodal features of the process node are updated according to the current state; or, if there is no corresponding entity node in the historical spatiotemporal knowledge graph, an entity node corresponding to the entity structure is added to the historical spatiotemporal knowledge graph. For each event, if a corresponding event node exists in the historical spatiotemporal knowledge graph, the multimodal features of the corresponding event node are updated according to the event; or if a corresponding event node does not exist in the historical spatiotemporal knowledge graph, a new event node is added to the historical spatiotemporal knowledge graph and a connection relationship is established between the event node and at least one node.

[0102] According to the present invention, a health status prediction device 300 for railway facilities is provided. The health status includes a health index. After processing the fusion features and graph adjacency bias matrix using a prediction model to obtain the health status of the railway facilities, the health status prediction module 304 is further configured to: If the health index of the target entity node is determined to be less than the preset health index, multimodal features of process nodes related to the target entity node are extracted from the target spatiotemporal knowledge graph to construct event pairs; The event pairs are used as the problems to be reasoned in the root cause analysis model. The root cause analysis model is used to determine the root cause analysis results of the event pairs based on the target spatiotemporal knowledge graph and the domain causal graph. The root cause analysis model is trained using sample event pairs, sample domain causal graphs and root cause analysis result labels.

[0103] According to the present invention, a health status prediction device 300 for railway facilities is provided, wherein the domain causality graph is obtained based on the following method: An initial domain causal graph constructed based on prior knowledge; Obtain historical operation and maintenance data of several railway facilities and mine potential causal relationships based on causal discovery algorithms; The initial domain causal graph is fused with the potential causal relationships to obtain the domain causal graph.

[0104] According to the present invention, a railway facility health status prediction device 300 includes a multimodal data acquisition module 301 that acquires multimodal data of the railway facility, comprising: Acquire the first data collected by drones on railway facilities; A global risk heat map of the railway facility is determined based on the first data. The global risk heat map is determined based on the defect type, defect density and confidence level of each area in the railway facility. Based on the overall risk heat map, areas to be investigated in detail were identified from the regions. Control the ground monitoring device to collect secondary data from each area to be investigated in detail; Multimodal data is constructed based on the first and second data.

[0105] According to the present invention, a railway facility health status prediction device 300 includes a multimodal data acquisition module 301 that controls a ground monitoring device to collect second data from each area to be investigated in detail, comprising: Obtain the current location of the ground monitoring device and the target location of the area to be investigated in detail; The movement path of the ground monitoring device is determined based on the current location, the target location, and the historical spatiotemporal knowledge graph; A data acquisition task is generated and sent to the ground monitoring device, so that the ground monitoring device can move to the target location according to the movement path in the data acquisition task and collect the second data.

[0106] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communications bus 440. The processor 410 can call logical instructions in the memory 430 to execute a method for predicting the health status of railway facilities. This method includes: acquiring multimodal data of the railway facilities; updating the historical spatiotemporal knowledge graph corresponding to the railway facilities using the multimodal data to obtain a target spatiotemporal knowledge graph, wherein the target spatiotemporal knowledge graph includes several nodes and their multimodal features, and each node is of the type of entity node, process node, or event node; constructing a graph adjacency bias matrix based on the connection relationships between entity nodes, process nodes, and / or event nodes in the target spatiotemporal knowledge graph, and constructing fusion features based on the multimodal features of each node; processing the fusion features and the graph adjacency bias matrix using a prediction model to obtain the health status of the railway facilities, wherein the prediction model is trained using the sample graph adjacency bias matrix, sample fusion features, and corresponding health status labels.

[0107] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the railway facility health status prediction method provided by the above methods. The method includes: acquiring multimodal data of the railway facility; updating the historical spatiotemporal knowledge graph corresponding to the railway facility using the multimodal data to obtain a target spatiotemporal knowledge graph, wherein the target spatiotemporal knowledge graph includes several nodes and their multimodal features, and the type of each node is an entity node, a process node, or an event node; constructing a graph adjacency bias matrix according to the connection relationship between each entity node, process node, and / or event node in the target spatiotemporal knowledge graph, and constructing fusion features according to the multimodal features of each node; processing the fusion features and the graph adjacency bias matrix using a prediction model to obtain the health status of the railway facility, wherein the prediction model is trained using the sample graph adjacency bias matrix, the sample fusion features, and the corresponding health status labels.

[0109] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the method for predicting the health status of railway facilities provided by the above methods. The method includes: acquiring multimodal data of the railway facilities; updating the historical spatiotemporal knowledge graph corresponding to the railway facilities using the multimodal data to obtain a target spatiotemporal knowledge graph, wherein the target spatiotemporal knowledge graph includes several nodes and their multimodal features, and the type of each node is an entity node, a process node, or an event node; constructing a graph adjacency bias matrix based on the connection relationships between each entity node, process node, and / or event node in the target spatiotemporal knowledge graph, and constructing fusion features based on the multimodal features of each node; processing the fusion features and the graph adjacency bias matrix using a prediction model to obtain the health status of the railway facilities, wherein the prediction model is trained using the sample graph adjacency bias matrix, the sample fusion features, and the corresponding health status labels.

[0110] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the health status of railway facilities, characterized in that, include: Acquire multimodal data of railway facilities; The historical spatiotemporal knowledge graph corresponding to the railway facilities is updated using the multimodal data to obtain a target spatiotemporal knowledge graph, wherein the target spatiotemporal knowledge graph includes several nodes and their multimodal features, and the type of each node is an entity node, a process node, or an event node; Based on the connection relationships between the entity nodes, process nodes and / or event nodes in the target spatiotemporal knowledge graph, a graph adjacency bias matrix is ​​constructed, and fusion features are constructed based on the multimodal features of each node; The health status of the railway facility is obtained by processing the fusion features and the graph adjacency bias matrix using a prediction model. The prediction model is trained using the sample graph adjacency bias matrix, sample fusion features, and corresponding health status labels.

2. The method for predicting the health status of railway facilities according to claim 1, characterized in that, The health status of the railway facility is obtained by processing the fused features and the graph adjacency bias matrix using a prediction model, including: The fusion features are linearly mapped using the prediction model to obtain a query matrix, a key matrix, and a value matrix. The query matrix and the transpose of the key matrix are subjected to a dot product operation, and the result of the dot product operation is fused with the graph adjacency bias matrix and then normalized to obtain the attention weight matrix. Multiplying the attention weight matrix by the value matrix yields the self-attention features; The health status of the railway facility is obtained based on the self-attention characteristics.

3. The method for predicting the health status of railway facilities according to claim 1, characterized in that, The multimodal features of each node include sub-multimodal features at multiple time points. A fused feature is constructed based on the multimodal features of each node, including: The embedding vector is obtained by acquiring the temporal position encoding of each sub-multimodal feature of each node at the corresponding time point and the node type encoding of each node; For each node, the spatiotemporal joint features of the node are obtained by adding the sub-multimodal features of the node, the temporal position encoding of the time point corresponding to each sub-multimodal feature, and the embedding vector of the node. The spatiotemporal joint features of each node are fused to obtain the fused features.

4. The method for predicting the health status of railway facilities according to any one of claims 1 to 3, characterized in that, Each process node in the target spatiotemporal knowledge graph is used to record the dynamic change of the state of one of the entity nodes over time, and each event node is used to record events that can cause changes in the state of at least some entity nodes.

5. The method for predicting the health status of railway facilities according to claim 4, characterized in that, The historical spatiotemporal knowledge graph corresponding to the railway facilities is updated using the multimodal data to obtain the target spatiotemporal knowledge graph, including: The multimodal data is subjected to entity structure recognition and current state extraction to obtain several entity structures and their current states. Event extraction is performed on the multimodal data to obtain several events. For each entity structure, there exists a corresponding entity node in the historical spatiotemporal knowledge graph and the corresponding entity node is connected to a corresponding process node. The multimodal features of the process node are updated according to the current state. Alternatively, if there is no corresponding entity node in the historical spatiotemporal knowledge graph, an entity node corresponding to the entity structure is added to the historical spatiotemporal knowledge graph. For each event, if a corresponding event node exists in the historical spatiotemporal knowledge graph, the multimodal features of the corresponding event node are updated according to the event; or if no corresponding event node exists in the historical spatiotemporal knowledge graph, a new event node is added to the historical spatiotemporal knowledge graph and a connection relationship is established between the event node and at least one node.

6. The method for predicting the health status of railway facilities according to any one of claims 1 to 3, characterized in that, The health status includes a health index. After processing the fused features and the graph adjacency bias matrix using a prediction model to obtain the health status of the railway facility, the method further includes: If it is determined that the health index of the target entity node is less than the preset health index, multimodal features of process nodes related to the target entity node are extracted from the target spatiotemporal knowledge graph to construct event pairs; The event pairs are used as the problems to be reasoned in the root cause analysis model. The root cause analysis model is used to determine the root cause analysis results of the event pairs based on the target spatiotemporal knowledge graph and the domain causal graph. The root cause analysis model is trained using sample event pairs, sample domain causal graphs and root cause analysis result labels.

7. The method for predicting the health status of railway facilities according to claim 6, characterized in that, The domain cause-effect graph is obtained based on the following method: An initial domain causal graph constructed based on prior knowledge; Obtain historical operation and maintenance data of several railway facilities and mine potential causal relationships based on causal discovery algorithms; The initial domain causal graph is fused with the potential causal relationships to obtain the domain causal graph.

8. The method for predicting the health status of railway facilities according to any one of claims 1 to 3, characterized in that, Acquire multimodal data of railway facilities, including: Acquire the first data collected by the drone on the railway facilities; A global risk heat map of the railway facility is determined based on the first data. The global risk heat map is obtained based on the defect type, defect density and confidence level of each area in the railway facility. Based on the global risk heat map, areas to be investigated in detail are determined from the region. Control the ground monitoring device to collect second data from each of the areas to be investigated in detail; The multimodal data is constructed based on the first data and the second data.

9. The method for predicting the health status of railway facilities according to claim 8, characterized in that, The control ground monitoring device collects second data from each of the areas to be investigated in detail, including: Obtain the current location of the ground monitoring device and the target location of the area to be investigated in detail; The movement path of the ground monitoring device is determined based on the current location, the target location, and the historical spatiotemporal knowledge graph. A data acquisition task is generated and sent to the ground monitoring device so that the ground monitoring device can move to the target location according to the movement path in the data acquisition task and acquire the second data.

10. A health status prediction device for railway facilities, characterized in that, include: The multimodal data acquisition module is used to acquire multimodal data of railway facilities; The spatiotemporal knowledge graph update module is used to update the historical spatiotemporal knowledge graph corresponding to the railway facilities using the multimodal data to obtain a target spatiotemporal knowledge graph. The target spatiotemporal knowledge graph includes several nodes and their multimodal features, and the type of each node is an entity node, a process node, or an event node. The data construction module is used to construct a graph adjacency bias matrix based on the connection relationships between the entity nodes, process nodes and / or event nodes in the target spatiotemporal knowledge graph, and to construct fusion features based on the multimodal features of each node. The health status prediction module is used to process the fused features and the graph adjacency bias matrix using a prediction model to obtain the health status of the railway facility. The prediction model is trained using the sample graph adjacency bias matrix, sample fused features and corresponding health status labels.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the railway facility health status prediction method as described in any one of claims 1 to 9.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the health status prediction method for railway facilities as described in any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the health status prediction method for railway facilities as described in any one of claims 1 to 9.

Images