A catenary hot defect evolution prediction and work order pre-generation method

Abstract

The application discloses a catenary hot defect evolution prediction and work order pre-generation method, and belongs to the technical field of electrified railway catenary safety monitoring. The application integrates related multi-source data of the catenary, constructs a component-level time series data set according to the unique identifier of the key components of the catenary, extracts statistical features and static features, processes the component-level time series data set through a grey system model, amplifies the temperature evolution trend and identifies potential defect components, inputs the processed time series data and static features into a time series Transformer model, outputs temperature prediction results and a critical failure time interval, evaluates operation and maintenance resources based on the prediction results, pre-generates alternative disposal work orders, activates corresponding work orders according to the actual state of the components, and completes dispatching. The application realizes preventive whole-process management of catenary hot defects, improves the reliability of defect identification and trend prediction, optimizes operation and maintenance resource allocation, and improves the overall operation and maintenance efficiency and decision-making scientificity of the system.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology for overhead contact lines of electrified railways, and in particular to a method for predicting the evolution of thermal defects in overhead contact lines and pre-generating work orders. Background Technology

[0002] With the continuous increase in the operational mileage of electrified railways, the overhead contact system, as the sole power source for electric locomotives, directly impacts the safety and efficiency of railway transportation. Key components such as overhead contact system dropper clamps, electrical connection clamps, insulators, and disconnecting switches operate under complex environments of high current, high vibration, and wide temperature ranges, making them prone to thermal defects due to poor contact and material aging. These defects are major causes of pantograph-catenary line failures and train service interruptions. Currently, overhead contact system thermal defect detection technology has developed into three mainstream approaches: manual inspection, fixed online monitoring, and onboard detection, each suitable for different line conditions and maintenance needs. Simultaneously, the railway industry has gradually established independent power supply maintenance management systems, work order management systems, and geographic information systems, enabling basic work order dispatch, progress tracking, and archiving management. This provides information support for the daily maintenance of the overhead contact system, and these technologies are widely used in electrified railway lines nationwide.

[0003] Existing overhead contact line thermal defect maintenance technologies can only detect and perform basic handling of existing thermal defects. They cannot effectively integrate multi-source data such as infrared detection, visible light imaging, environmental sensing, line load, and equipment basic information. Furthermore, they cannot construct cross-cycle component-level time-series datasets based on the unique identifiers of key overhead contact line components, resulting in a lack of effective correlation between data from different sources and at different times. Current technologies cannot identify potential thermal defects in their early stages, nor can they accurately predict the future evolution trend and critical failure time of discovered defects, making it difficult to grasp the development trend of defects in advance. Simultaneously, existing technologies do not integrate defect prediction results with maintenance resource management. They cannot assess maintenance resource needs in advance based on defect development trends and pre-generate handling work orders. Instead, manual scheduling and work order creation are only possible after a defect occurs, resulting in maintenance work remaining in a reactive state and failing to meet the ever-increasing safety maintenance needs of electrified railways. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the evolution of thermal defects in overhead contact lines and pre-generating work orders.

[0005] The objective of this invention is achieved through the following technical solution: A method for predicting the evolution of thermal defects in overhead contact lines and pre-generating work orders is provided. This method includes the following steps: S1. Integrate multi-source data related to the overhead contact system, construct a component-level time-series dataset according to the unique identifier of the key components of the overhead contact system, and extract the statistical and static features of the key components of the overhead contact system; S2. Based on the component-level time series dataset, the gray system model is used to amplify the temperature evolution trend of key components of the overhead contact system and identify potentially defective components; S3. Input the time-series data processed by the grey system model and the static characteristics of the key components of the overhead contact system into the time-series Transformer model, and output the temperature prediction results and critical failure time intervals of the key components of the overhead contact system. S4. Based on temperature prediction results and critical failure time intervals, conduct operation and maintenance resource assessment, pre-generate alternative disposal work orders, activate the corresponding disposal work orders according to the actual status of key components of the overhead contact system, and complete the dispatch.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. Acquire contact network infrared temperature data, visible light image data, positioning data, environmental sensor data, line load data, equipment basic information, and historical work order processing data; S1.2. The acquired multi-source data are time-aligned and spatial-aligned according to the unique identifiers of the key components of the overhead contact system to generate a cross-cycle component-level time-series dataset; S1.3. Extract statistical and static features of key overhead contact system components from the component-level time-series dataset. Time alignment is based on a unified timestamp, and spatial alignment is based on location data and overhead contact system support numbers. Statistical features include mean temperature, standard deviation of temperature, rate of temperature rise, and temperature distribution entropy. Static features include component type, installation location, years of operation, and maintenance records. The component-level time-series dataset is stored in chronological order and supports retrieval by the unique identifier of the key overhead contact system component.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. Perform cumulative generation operation on the temperature time series data in the component-level time series dataset to reduce the randomness of the temperature time series data; S2.2. The gray GM(1,1) model is used to fit the processed temperature time series data to generate a temperature evolution sequence; S2.3. Calculate the residual between the temperature evolution sequence and the actual temperature sequence, identify potential defective components, and perform cumulative generation operation to accumulate the temperature time series data item by item in chronological order. The gray GM(1,1) model obtains the fitting curve of temperature evolution by solving the differential equation. The residual is the difference between the temperature evolution sequence and the actual temperature sequence. When the residual continuously exceeds the set value, the corresponding contact wire key component is determined to be a potential defective component. The information of the identified potential defective components is input into the subsequent detection frequency adjustment step.

[0008] Furthermore, step S3 includes the following sub-steps: S3.1. Construct a temporal Transformer model, which includes an input layer, a multi-head self-attention layer, a feedforward neural network layer, and an output layer; S3.2. Input the temperature time series data, environmental parameter time series data, line load time series data and static characteristics of key components of the catenary after processing by the gray system model into the input layer of the time series Transformer model; S3.3. Capture feature dependencies at different time scales through a multi-head self-attention layer, and perform feature transformation through a feedforward neural network layer; S3.4. The output layer outputs the temperature prediction results and critical failure time intervals of key components of the overhead contact system. The input layer performs dimensional transformation and normalization on the input data. The multi-head self-attention layer contains multiple parallel attention heads, each of which independently calculates the attention weights of the features. The feedforward neural network layer contains two linear transformation layers and one activation function layer. The output layer maps the features to the temperature prediction results and critical failure time intervals through linear transformation.

[0009] Furthermore, step S4 includes the following sub-steps: S4.1. Obtain usage status data of operation and maintenance resources and railway track maintenance window plan data, and construct a spatiotemporal map of available operation and maintenance resources; S4.2. Based on the temperature prediction results and the critical failure time interval, the early warning level is divided, and alternative disposal work orders are pre-generated for different early warning levels; S4.3. Compare the actual temperature of key components of the overhead contact system with the set threshold. When the actual temperature reaches the set threshold, activate the corresponding alternative handling work order. S4.4. The activated handling work order is dispatched to the corresponding handling entity. The spatiotemporal map of operation and maintenance resources can be used to include the spatiotemporal distribution information of operation and maintenance personnel, operation and maintenance vehicles and operation and maintenance tools. The alternative handling work order includes the handling time limit, responsible entity, required resources, handling steps and acceptance criteria. The threshold is set according to the warning level. The handling work order is dispatched to the terminal device of the corresponding handling entity through a standardized communication interface.

[0010] Furthermore, in step S1.3, after extracting the statistical and static features of the key components of the overhead contact system, the grey relational analysis method is used to calculate the correlation between the statistical and static features and the defect evolution. Features with a correlation greater than a set value are retained to generate a feature subset. The grey relational analysis method calculates the grey relational coefficient between each feature and the defect evolution. The correlation is obtained by the average value of the correlation coefficients. The set value is set according to the statistical results of historical data. The feature subset contains all statistical and static features with a correlation greater than the set value. The feature subset is input into the grey system model and the time-series Transformer model for subsequent defect evolution prediction.

[0011] Furthermore, in step S2.3, after identifying potentially defective components, the detection frequency of key components of the overhead contact system is adjusted, dense time-series data of potentially defective components is obtained, and the component-level time-series dataset is updated. The detection frequency adjustment is set according to the residual size of the potentially defective components. The dense time-series data includes infrared temperature data, visible light image data, and environmental sensing data with higher time resolution. When updating the component-level time-series dataset, the dense time-series data is inserted into the original data according to the unique identifier and timestamp of the key components of the overhead contact system. The updated component-level time-series dataset is re-input into the gray system model for a new round of defect evolution analysis.

[0012] Furthermore, in step S3.4, when outputting the temperature prediction results and critical failure time intervals of key components of the overhead contact system, the Monte Carlo dropout method is used to calculate the confidence level of the temperature prediction results and critical failure time intervals to generate risk assessment data. The Monte Carlo dropout method randomly discards some neurons in the feedforward neural network layer and performs multiple forward inferences. The confidence level interval of the temperature prediction results and critical failure time intervals is calculated through the statistical distribution of the results of multiple inferences. The risk assessment data is classified according to the confidence level intervals and input into the work order pre-generation step to adjust the disposal priority and resource allocation of alternative disposal work orders.

[0013] Furthermore, in step S4.4, after the activated disposal work order is dispatched to the corresponding disposal entity, the temperature data of the key components of the overhead contact system and the status data of the maintenance resources are acquired in real time. The content and dispatch object of the disposal work order are adjusted according to the temperature data and the status data of the maintenance resources. The temperature data is collected at a set time interval, and the status data of the maintenance resources is updated at a set period. The adjustment of the content of the disposal work order includes the modification of the disposal time limit, required resources and disposal steps. The adjustment of the dispatch object is based on the real-time load of the maintenance resources. The adjusted disposal work order is re-dispatched to the terminal equipment of the corresponding disposal entity through the standardized communication interface.

[0014] Furthermore, after step S4, the execution result data of the handling work order is obtained, and the execution result data is fed back to the gray system model and the time-series Transformer model to update the parameters of the gray system model and the time-series Transformer model. The execution result data includes the defect elimination status, actual handling time and resource consumption. The parameter update is carried out in an incremental learning manner. When updating the parameters of the gray system model, the fitting curve of the temperature evolution sequence is corrected using the execution result data. When updating the parameters of the time-series Transformer model, the weights of the model are adjusted using the execution result data. The updated gray system model and time-series Transformer model are used for subsequent prediction of the evolution of thermal defects in the overhead contact system.

[0015] The beneficial effects of this invention are: (1) By integrating multi-source data to construct a component-level time series dataset, and combining the gray system and time series Transformer model to complete defect identification and evolution prediction, the operation and maintenance resources are evaluated based on the prediction results and work orders are pre-generated to realize the preventive full-process management of contact network thermal defects; (2) By screening multi-dimensional features and optimizing model structure, and integrating multiple influencing factors to carry out defect evolution analysis, we can effectively capture early potential defect signals and improve the reliability of defect identification and trend prediction. (3) By constructing a spatiotemporal map of operation and maintenance resources and a dynamic scheduling mechanism for work orders, and combining incremental learning feedback of disposal results, the allocation of operation and maintenance resources is optimized, and the overall operation and maintenance efficiency and scientific decision-making of the system are continuously improved. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a method for predicting the evolution of thermal defects in overhead contact lines and pre-generating work orders. Figure 2 The flowchart illustrates the specific steps of a method for predicting the evolution of thermal defects in overhead contact lines and pre-generating work orders, as provided in this embodiment. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 See Figure 1 This embodiment provides a method for predicting the evolution of thermal defects in overhead contact lines and pre-generating work orders. The method includes the following steps: S1. Integrate multi-source data related to the overhead contact system, construct a component-level time-series dataset according to the unique identifier of the key components of the overhead contact system, and extract the statistical and static features of the key components of the overhead contact system; S2. Based on the component-level time series dataset, the gray system model is used to amplify the temperature evolution trend of key components of the overhead contact system and identify potentially defective components; S3. Input the time-series data processed by the grey system model and the static characteristics of the key components of the overhead contact system into the time-series Transformer model, and output the temperature prediction results and critical failure time intervals of the key components of the overhead contact system. S4. Based on temperature prediction results and critical failure time intervals, conduct operation and maintenance resource assessment, pre-generate alternative disposal work orders, activate the corresponding disposal work orders according to the actual status of key components of the overhead contact system, and complete the dispatch.

[0019] In some embodiments, step S1 includes the following sub-steps: S1.1. Acquire contact network infrared temperature data, visible light image data, positioning data, environmental sensor data, line load data, equipment basic information, and historical work order processing data; S1.2. The acquired multi-source data are time-aligned and spatial-aligned according to the unique identifiers of the key components of the overhead contact system to generate a cross-cycle component-level time-series dataset; S1.3. Extract statistical and static features of key overhead contact system components from the component-level time-series dataset. Time alignment is based on a unified timestamp, and spatial alignment is based on location data and overhead contact system support numbers. Statistical features include mean temperature, standard deviation of temperature, rate of temperature rise, and temperature distribution entropy. Static features include component type, installation location, years of operation, and maintenance records. The component-level time-series dataset is stored in chronological order and supports retrieval by the unique identifier of the key overhead contact system component.

[0020] In some embodiments, step S2 includes the following sub-steps: S2.1. Perform cumulative generation operation on the temperature time series data in the component-level time series dataset to reduce the randomness of the temperature time series data; S2.2. The gray GM(1,1) model is used to fit the processed temperature time series data to generate a temperature evolution sequence; S2.3. Calculate the residual between the temperature evolution sequence and the actual temperature sequence, identify potential defective components, and perform cumulative generation operation to accumulate the temperature time series data item by item in chronological order. The gray GM(1,1) model obtains the fitting curve of temperature evolution by solving the differential equation. The residual is the difference between the temperature evolution sequence and the actual temperature sequence. When the residual continuously exceeds the set value, the corresponding contact wire key component is determined to be a potential defective component. The information of the identified potential defective components is input into the subsequent detection frequency adjustment step.

[0021] In some embodiments, step S3 includes the following sub-steps: S3.1. Construct a temporal Transformer model, which includes an input layer, a multi-head self-attention layer, a feedforward neural network layer, and an output layer; S3.2. Input the temperature time series data, environmental parameter time series data, line load time series data and static characteristics of key components of the catenary after processing by the gray system model into the input layer of the time series Transformer model; S3.3. Capture feature dependencies at different time scales through a multi-head self-attention layer, and perform feature transformation through a feedforward neural network layer; S3.4. The output layer outputs the temperature prediction results and critical failure time intervals of key components of the overhead contact system. The input layer performs dimensional transformation and normalization on the input data. The multi-head self-attention layer contains multiple parallel attention heads, each of which independently calculates the attention weights of the features. The feedforward neural network layer contains two linear transformation layers and one activation function layer. The output layer maps the features to the temperature prediction results and critical failure time intervals through linear transformation.

[0022] In some embodiments, step S4 includes the following sub-steps: S4.1. Obtain usage status data of operation and maintenance resources and railway track maintenance window plan data, and construct a spatiotemporal map of available operation and maintenance resources; S4.2. Based on the temperature prediction results and the critical failure time interval, the early warning level is divided, and alternative disposal work orders are pre-generated for different early warning levels; S4.3. Compare the actual temperature of key components of the overhead contact system with the set threshold. When the actual temperature reaches the set threshold, activate the corresponding alternative handling work order. S4.4. The activated handling work order is dispatched to the corresponding handling entity. The spatiotemporal map of operation and maintenance resources can be used to include the spatiotemporal distribution information of operation and maintenance personnel, operation and maintenance vehicles and operation and maintenance tools. The alternative handling work order includes the handling time limit, responsible entity, required resources, handling steps and acceptance criteria. The threshold is set according to the warning level. The handling work order is dispatched to the terminal device of the corresponding handling entity through a standardized communication interface.

[0023] In some embodiments, in step S1.3, after extracting the statistical and static features of the key components of the overhead contact system, the grey relational analysis method is used to calculate the correlation between the statistical and static features and the defect evolution. Features with a correlation greater than a set value are retained to generate a feature subset. The grey relational analysis method calculates the grey relational coefficient between each feature and the defect evolution. The correlation is obtained by the average value of the correlation coefficients. The set value is set according to the statistical results of historical data. The feature subset contains all statistical and static features with a correlation greater than the set value. The feature subset is input into the grey system model and the time-series Transformer model for subsequent defect evolution prediction.

[0024] In some embodiments, in step S2.3, after identifying potentially defective components, the detection frequency of key components of the overhead contact system is adjusted, dense time-series data of the potentially defective components are obtained, and the component-level time-series dataset is updated. The detection frequency adjustment is set according to the residual size of the potentially defective components. The dense time-series data includes infrared temperature data, visible light image data, and environmental sensing data with higher time resolution. When updating the component-level time-series dataset, the dense time-series data is inserted into the original data according to the unique identifier and timestamp of the key components of the overhead contact system. The updated component-level time-series dataset is re-input into the gray system model for a new round of defect evolution analysis.

[0025] In some embodiments, in step S3.4, when outputting the temperature prediction results and critical failure time intervals of key components of the overhead contact system, the Monte Carlo dropout method is used to calculate the confidence level of the temperature prediction results and critical failure time intervals to generate risk assessment data. The Monte Carlo dropout method randomly discards some neurons in the feedforward neural network layer and performs multiple forward inferences. The confidence level interval of the temperature prediction results and critical failure time intervals is calculated by the statistical distribution of the results of multiple inferences. The risk assessment data is classified according to the confidence level intervals. The risk assessment data is input to the work order pre-generation step to adjust the disposal priority and resource allocation of alternative disposal work orders.

[0026] In some embodiments, in step S4.4, after the activated disposal work order is dispatched to the corresponding disposal entity, the temperature data of the key components of the overhead contact line and the status data of the operation and maintenance resources are acquired in real time. The content and dispatch object of the disposal work order are adjusted according to the temperature data and the status data of the operation and maintenance resources. The temperature data is collected at a set time interval, and the status data of the operation and maintenance resources is updated at a set period. The adjustment of the content of the disposal work order includes the modification of the disposal time limit, required resources and disposal steps. The adjustment of the dispatch object is based on the real-time load of the operation and maintenance resources. The adjusted disposal work order is re-dispatched to the terminal equipment of the corresponding disposal entity through the standardized communication interface.

[0027] In some embodiments, after step S4, the execution result data of the handling work order is obtained, and the execution result data is fed back to the gray system model and the time-series Transformer model to update the parameters of the gray system model and the time-series Transformer model. The execution result data includes the defect elimination status, actual handling time and resource consumption. The parameter update is performed using incremental learning. When updating the parameters of the gray system model, the fitting curve of the temperature evolution sequence is corrected using the execution result data. When updating the parameters of the time-series Transformer model, the weights of the model are adjusted using the execution result data. The updated gray system model and time-series Transformer model are used for subsequent prediction of the evolution of thermal defects in the overhead contact system.

[0028] Example 2 This embodiment provides a specific implementation process for a method of predicting the evolution of thermal defects in overhead contact lines and pre-generating work orders. Based on the architecture of an electrified railway overhead contact line operation and maintenance system, this embodiment integrates multi-source detection data and operation and maintenance data to achieve early identification of thermal defects in the overhead contact line, prediction of their evolution trends, and preventative work order management. It is applicable to the daily operation and maintenance and safety assurance work of various electrified railway overhead contact lines. Figure 2 As shown, the specific implementation process is as follows: Step 1. Multi-source data integration and feature extraction: Step 1.1. Acquisition of multi-source data: The system acquires contact network infrared temperature data, visible light image data, positioning data, environmental sensor data, line load data, equipment basic information, and historical work order processing data. Contact network infrared temperature data refers to the surface temperature distribution data of various components of the contact network collected by infrared thermal imaging equipment, reflecting the thermal state of the components. In this embodiment, it is collected by an infrared thermal imaging device installed on the top of the inspection vehicle, with the acquisition frequency matched to the vehicle's operating speed. Visible light image data refers to the appearance image data of the contact network components collected by a visible light camera, used to assist in component identification and defect location. It is collected synchronously with the infrared thermal imaging equipment to ensure a one-to-one correspondence between infrared and visible light data at the same time. Positioning data refers to the geographical location data of the inspection equipment obtained through a satellite positioning system, used to determine the spatial location of the contact network components. The acquisition frequency is consistent with that of the infrared thermal imaging equipment. Environmental sensor data refers to the temperature, humidity, and wind speed data of the inspection site collected by environmental sensors. Environmental parameter data, such as speed, is used to correct temperature measurement results and analyze the impact of the environment on defect evolution. Environmental sensors are installed on the outside of the inspection vehicle and can collect environmental parameters at the inspection site in real time. Line load data refers to load parameter data such as current and voltage of the catenary power supply line, which is used to analyze the impact of load changes on the thermal state of components and is obtained from the database of the railway power supply system. Equipment basic information refers to static information such as the model, installation time, and manufacturer of each component of the catenary, which is used to assess the aging degree and failure probability of the components and is obtained from the database of the railway equipment management system. Historical work order handling data refers to maintenance record data such as the discovery time, handling process, and handling results of past catenary defects, which is used to analyze the evolution pattern of defects and the handling effect and is obtained from the database of the railway work order management system.

[0029] Step 1.2. Construction of Time Series Dataset: The acquired multi-source data is time-aligned and spatially aligned according to the unique identifiers of key components of the overhead contact system, generating a cross-cycle component-level time-series dataset. Time alignment involves matching data from different sources using a unified timestamp to ensure that different types of data collected at the same time correspond to the same overhead contact system component. In this embodiment, the timestamp of data collected by the infrared thermal imaging device is used as a benchmark, and visible light image data, positioning data, and environmental sensor data are interpolated and matched according to their timestamps to ensure consistency across all data. Spatial alignment involves matching positioning data with overhead contact system support numbers to determine the specific location of the overhead contact system component corresponding to each data point. A pre-established database of overhead contact system support locations is used to convert the positioning data into corresponding line sections and support numbers, thereby determining the overhead contact system component corresponding to each data point. The component-level time-series dataset is organized according to the unique identifiers of key overhead contact system components. Each component corresponds to one time-series data record, which includes information such as temperature data, environmental data, and load data for that component at different time points. The dataset is stored in chronological order and supports retrieval by the unique identifier of the key overhead contact system component.

[0030] In some specific implementations, to address the feature misalignment problem caused by insufficient time synchronization accuracy of multi-source sensor data, a high-precision synchronous acquisition board is used to achieve microsecond-level time synchronization of multi-source data. The infrared thermal imaging device is set to a 100fps acquisition frame rate, the visible light camera is set to the same 100fps acquisition frame rate, the satellite positioning module is set to a 10Hz acquisition frequency, and the environmental sensor is set to a 50Hz acquisition frequency. All sensors are triggered for acquisition by a 1MHz synchronization clock signal output from the high-precision synchronous acquisition board. The synchronous acquisition board adds a uniform timestamp with a resolution of 1 microsecond to each frame of acquired data, controlling the time synchronization accuracy within ±1 microsecond. For sensor data with different acquisition frequencies, a linear interpolation method is used to resample them to the same 100Hz time resolution as the infrared thermal imaging device. During the interpolation process, the timestamp information of the original data is retained, ensuring that the time error between the interpolated data and the original data does not exceed 1 millisecond. This method ensures that infrared temperature data, visible light image data, positioning data, and environmental sensor data acquired at the same time are accurately mapped to the same contact wire component, avoiding feature extraction errors caused by time misalignment.

[0031] Step 1.3. Feature Extraction and Filtering: Statistical and static features of key overhead contact system components are extracted from component-level time-series datasets. Statistical features include mean temperature, standard deviation temperature, rate of temperature rise, and temperature distribution entropy. The mean temperature refers to the average temperature of a component over a period of time, reflecting its overall thermal state; it is obtained by calculating the average of all temperature samples within a set time window. The standard deviation temperature refers to the degree of temperature fluctuation over a period of time, reflecting the stability of the component's thermal state; it is obtained by calculating the standard deviation of all temperature samples within a set time window. The rate of temperature rise refers to the rate of temperature change over time, reflecting the speed of thermal defect development; it is obtained by calculating the ratio of the temperature difference between two adjacent time points to the time interval. The temperature distribution entropy refers to the degree of disorder in the surface temperature distribution of a component, reflecting the distribution characteristics of thermal defects; it is obtained by calculating the entropy of the probability density function of the surface temperature distribution. Static characteristics include component type, installation location, service life, and maintenance records. Component type refers to the type of overhead contact line component, such as dropper clamps, electrical connection clamps, insulators, and disconnectors. Installation location refers to the specific location of the component in the overhead contact line, such as section or support post number. Service life refers to the time from installation to current use. Maintenance records refer to information such as past maintenance times and contents. Time alignment is based on a unified timestamp, and spatial alignment is based on location data and overhead contact line support post numbers. Component-level time-series datasets are stored in chronological order and support retrieval by the unique identifier of key overhead contact line components.

[0032] In some embodiments, after extracting the statistical and static features of key components of the overhead contact system, a grey relational analysis method is used to calculate the correlation degree between the statistical and static features and defect evolution. Features with a correlation degree greater than a set value are retained to generate a feature subset. Grey relational analysis is a method used to analyze the degree of correlation between various factors in a system. By calculating the grey relational coefficient between each factor and system behavior features, the influence of each factor on system behavior is determined. In this embodiment, the grey relational analysis method calculates the grey relational coefficient between each feature and defect evolution, and obtains the correlation degree through the average value of the correlation coefficients. The set value is set based on the statistical results of historical data. The feature subset includes all statistical and static features with a correlation degree greater than the set value. The feature subset is input into the grey system model and the time-series Transformer model for subsequent defect evolution prediction.

[0033] In some specific implementations, to address the issue that a single time window feature cannot comprehensively reflect the evolutionary characteristics of defects at different stages, a multi-scale sliding window method is employed to extract multi-dimensional temporal features during the feature extraction process. First, three sliding time windows of different lengths are set: a short-term window of length 10, a medium-term window of length 50, and a long-term window of length 200, with a sliding step size of 1 for each window. Then, corresponding statistical features are extracted within each sliding time window, including mean temperature, standard deviation of temperature, rate of temperature rise, and temperature distribution entropy. Next, the statistical features extracted from different sliding time windows are concatenated to form a multi-scale statistical feature vector containing 12 dimensions. Finally, the multi-scale statistical feature vector is fused with static features containing 4 dimensions to generate a final feature subset containing 16 dimensions. This method can simultaneously capture the evolutionary patterns of defects at hourly, daily, and weekly time scales, avoiding the loss of feature information due to improper setting of a single time window and improving the accuracy of subsequent defect evolution prediction.

[0034] Step 2. Potential Defect Identification and Trend Analysis: Step 2.1. Temperature data preprocessing: An accumulation operation is performed on the temperature time-series data in the component-level time-series dataset to reduce the randomness of the temperature time-series data. The accumulation operation adds the temperature time-series data item by item in chronological order, transforming the random fluctuations in the original data into a smooth growth trend, which facilitates subsequent model fitting and trend analysis. In this embodiment, for the original temperature time-series data sequence... The sequence obtained by the cumulative generation operation: ; in , .

[0035] Step 2.2. Generation of temperature evolution sequence: A grey GM(1,1) model is used to fit the processed temperature time-series data to generate a temperature evolution sequence. The grey GM(1,1) model is a prediction model based on grey system theory, suitable for system prediction problems with limited data and incomplete information. In this embodiment, the grey GM(1,1) model obtains the fitted curve of temperature evolution by solving a first-order linear differential equation. This curve reflects the temperature change trend of the contact wire components over time. The differential equation of the grey GM(1,1) model is: ; in, and The parameters of the model are obtained by solving using the least squares method. and Then, the prediction formula for the temperature evolution sequence is obtained as follows: , .

[0036] In some specific implementations, to address the insufficient fitting accuracy of the original grey GM(1,1) model for nonlinear temperature evolution trends, a residual-corrected grey GM(1,1) model is constructed to optimize the temperature evolution sequence. First, the original grey GM(1,1) model is used to obtain the initial temperature evolution sequence and the corresponding residual sequence. Then, the residual sequence undergoes sign testing, and continuous portions with consistent signs are extracted to construct residual subsequences, with a minimum length of 10 sampling points. Next, the grey GM(1,1) model is used to fit the residual subsequences, generating a residual prediction sequence. Finally, the residual prediction sequence is superimposed on the initial temperature evolution sequence to obtain the corrected temperature evolution sequence. This method can reduce the average fitting error of the original grey GM(1,1) model to within 2℃, effectively compensating for the fitting error of the original model and improving the predictive ability for nonlinear temperature evolution processes. It is particularly suitable for predicting trends after thermal defects in contact wire components enter a rapid development stage.

[0037] Step 2.3. Potential Defect Identification: The residual between the temperature evolution sequence and the actual temperature sequence is calculated to identify potentially defective components. The residual is the difference between the temperature evolution sequence and the actual temperature sequence. When the residual continuously exceeds a set value, the corresponding key contact wire component is determined to be a potentially defective component. In this embodiment, the residual sequence is: ; in, , , This is a predicted value from the original temperature sequence obtained based on the temperature evolution sequence. When multiple consecutive values ​​in the residual sequence exceed the set value, it indicates that the temperature change of the component deviates from the normal evolution trend, indicating a potential thermal defect. The information of the identified potentially defective components is input into the subsequent detection frequency adjustment step.

[0038] In some specific implementations, to address the issues of missed detections and false alarms caused by ambiguity in residual judgment thresholds, a tiered residual judgment threshold and a continuous anomaly point judgment rule are established. The residual judgment threshold is divided into three levels: 2℃, 3℃, and 4℃, corresponding to different potential defect risk levels. The number of continuous anomaly points is set to five; that is, when the residual values ​​of five consecutive sampling points in the residual sequence exceed the threshold of the corresponding level, the component is judged to have a potential defect of the corresponding level. Simultaneously, the threshold is adjusted according to the type of contact network component: a first-level threshold of 3℃ is used for dropper clamps and electrical connection clamps, while a first-level threshold of 2℃ is used for insulators and disconnectors. This method can control the missed detection rate of early potential defects to within 5% and the false alarm rate to within 3%, balancing the contradiction between missed detections and false alarms and improving the reliability of early defect identification.

[0039] In some embodiments, after identifying potentially defective components, the detection frequency of key components of the overhead contact system is adjusted to obtain dense time-series data of the potentially defective components, and the component-level time-series dataset is updated. The detection frequency adjustment is set according to the residual size of the potentially defective components; the larger the residual, the higher the detection frequency. The dense time-series data includes infrared temperature data, visible light image data, and environmental sensor data with higher temporal resolution. When updating the component-level time-series dataset, the dense time-series data is inserted into the original data according to the unique identifier and timestamp of the key components of the overhead contact system. The updated component-level time-series dataset is then re-input into the gray system model for a new round of defect evolution analysis.

[0040] Step 3. Defect Evolution Prediction and Risk Assessment: Step 3.1. Constructing the Temporal Transformer Model: A temporal Transformer model is constructed, comprising an input layer, a multi-head self-attention layer, a feedforward neural network layer, and an output layer. The input layer receives and preprocesses the input data; the multi-head self-attention layer captures feature dependencies at different time scales within the input data; the feedforward neural network layer performs a non-linear transformation on the features output by the attention layer; and the output layer maps the transformed features to the prediction result. All layers are fully connected, with the output of the input layer serving as the input to the multi-head self-attention layer, the output of the multi-head self-attention layer serving as the input to the feedforward neural network layer, and the output of the feedforward neural network layer serving as the input to the output layer.

[0041] In some embodiments, the training process of the time-series Transformer model is as follows: First, historical contact network thermal defect data and corresponding evolution results data are collected to construct a training dataset. The training dataset contains input features and corresponding labels. The input features are time-series temperature data of components, time-series environmental parameter data, time-series line load data, and static features. The labels are the future temperature values ​​and critical failure times of the components. Second, the training dataset is input into the time-series Transformer model, and the backpropagation algorithm is used to adjust the model's weight parameters. The loss function is the mean squared error function, and the optimizer is the Adam optimizer. Then, the trained model is validated using a validation dataset, and the model's hyperparameters, such as the learning rate, batch size, and number of training epochs, are adjusted to prevent overfitting. Finally, the predictive performance of the model is evaluated using a test dataset to ensure that the model meets the prediction accuracy requirements.

[0042] Step 3.2. Model Input Data Preparation: The temperature time-series data, environmental parameter time-series data, line load time-series data, and static features of key components of the overhead contact system, processed by the grey system model, are input into the input layer of the time-series Transformer model. The input layer performs dimensionality transformation and normalization on the input data, converting input data of different dimensions and magnitudes into feature vectors of a unified format for easier subsequent model processing. In this embodiment, the input layer converts the temperature time-series data, environmental parameter time-series data, and line load time-series data into fixed-length sequence vectors, converts the static features into fixed-length vectors, and then concatenates all vectors into a single complete feature vector, which is then input into the multi-head self-attention layer.

[0043] Step 3.3. Feature Extraction and Transformation: A multi-head self-attention layer captures feature dependencies at different time scales, and a feedforward neural network layer performs feature transformation. The multi-head self-attention layer contains multiple parallel attention heads, each independently calculating attention weights for features. Different attention heads can capture different types of feature dependencies in the input data, such as short-term and long-term dependencies. In this embodiment, each attention head obtains its attention weight by calculating the dot product of the query vector, key vector, and value vector. Then, the value vectors are weighted and summed according to the attention weights to obtain the output of that attention head. The outputs of multiple attention heads are concatenated and then transformed through a linear transformation layer to obtain the final output of the multi-head self-attention layer. The feedforward neural network layer contains two linear transformation layers and one activation function layer. The first linear transformation layer maps the features output by the attention layer to a high-dimensional space, and the activation function layer performs a non-linear transformation on the high-dimensional features using the ReLU function. The second linear transformation layer maps the high-dimensional features back to a low-dimensional space, obtaining the transformed feature vector.

[0044] In some specific implementations, to improve the ability of the temporal Transformer model to capture the temporal features of overhead contact lines, relative position encoding is used instead of traditional absolute position encoding. First, the relative position offset between any two time steps in the input sequence is calculated, with the value ranged from -32 to +32; offsets exceeding this range are truncated. Then, the relative position offset is converted into a 64-dimensional relative position encoding vector, the dimension of which is consistent with the dimension of each attention head in the multi-head self-attention layer. Next, when calculating the attention weights, the relative position encoding vector is incorporated into the dot product operation of the query vector and the key vector to obtain attention weights that include relative position information. Finally, the value vectors are weighted and summed according to the adjusted attention weights to obtain the feature output incorporating relative position information. This method can effectively capture the relative dependencies between different time steps in the input sequence, solve the problem of absolute position encoding information decay in long sequence prediction, and improve the accuracy of the model's long sequence prediction.

[0045] Step 3.4. Prediction result output: The output layer outputs the temperature prediction results and critical failure time intervals for key components of the overhead contact system. The output layer maps the feature vectors output by the feedforward neural network layer to the temperature prediction results and critical failure time intervals through a linear transformation. The input layer performs dimensionality transformation and normalization on the input data. The multi-head self-attention layer contains multiple parallel attention heads, each independently calculating the attention weights of the features. The feedforward neural network layer contains two linear transformation layers and one activation function layer. The output layer maps the features to the temperature prediction results and critical failure time intervals through a linear transformation. In this embodiment, the output layer contains two linear transformation branches: one branch predicts the temperature values ​​of the components at different future time points, and the other branch predicts the critical failure time intervals of the components. The outputs of the two branches together constitute the final prediction result of the model.

[0046] In some embodiments, when outputting temperature prediction results and critical failure time intervals for key components of the overhead contact system, the Monte Carlo dropout method is used to calculate the confidence levels of the temperature prediction results and critical failure time intervals to generate risk assessment data. The Monte Carlo dropout method is a method for estimating the uncertainty of neural network prediction results. It randomly discards some neurons in the feedforward layer of the neural network, performs multiple forward inferences, obtains multiple prediction results, and calculates the confidence interval of the prediction results by statistically analyzing the distribution of these multiple prediction results. In this embodiment, the Monte Carlo dropout method randomly discards some neurons in the feedforward neural network layer, performs multiple forward inferences, and calculates the confidence interval of the temperature prediction results and critical failure time intervals by statistically analyzing the distribution of the multiple inference results. The risk assessment data is classified according to the confidence intervals and input into the work order pre-generation step to adjust the disposal priority and resource allocation of alternative disposal work orders.

[0047] Step 4. Work order pre-generation and dynamic scheduling: Step 4.1. Operation and Maintenance Resource Assessment: This method acquires usage status data of maintenance resources and railway track maintenance window plan data to construct a spatiotemporal map of available maintenance resources. The spatiotemporal map includes the spatiotemporal distribution information of maintenance personnel, vehicles, and tools, providing a clear view of available maintenance resources at different times and locations. Railway track maintenance window plan data refers to the time periods allocated by the railway department for line maintenance, used to determine available time windows for defect handling. In this embodiment, the usage status data of maintenance resources is obtained from the database of the railway maintenance resource management system, including personnel scheduling information, vehicle dispatching information, and tool usage information; the track maintenance window plan data is obtained from the database of the railway dispatching system, including track maintenance window times and types for different lines and sections. When constructing the spatiotemporal map of available maintenance resources, the usage status data and track maintenance window plan data are integrated, and an index of available maintenance resources is established using time and space as dimensions, supporting queries for available maintenance resources by time and space.

[0048] Step 4.2. Pre-generation of alternative work orders: Early warning levels are determined based on temperature prediction results and critical failure time intervals, and alternative handling work orders are pre-generated for each level. Each alternative handling work order includes a handling timeframe, responsible party, required resources, handling steps, and acceptance criteria. Early warning levels are categorized according to the severity and urgency of the defect, with different levels corresponding to different handling timeframes and procedures. In this embodiment, early warning levels are divided into three levels, each corresponding to different handling requirements and resource allocations. When pre-generating alternative handling work orders, a corresponding work order template is selected from the standardized work order template library based on the defect's early warning level, and the defect's location information, temperature information, prediction results, etc., are automatically filled in to generate the alternative handling work order.

[0049] In some specific implementations, to address the issues of operational resource conflicts and unreasonable work order scheduling in scenarios with multiple concurrent defects, a work order priority ranking model based on multiple constraints is constructed. First, five influencing factors for work order priority are identified, including defect risk level, critical failure time, required resources for handling, window time matching degree, and historical handling difficulty. Then, corresponding weight coefficients are assigned to each influencing factor, namely 0.3, 0.25, 0.2, 0.15, and 0.1, determined using the analytic hierarchy process (AHP) combined with historical operational data. Next, based on the specific information of each candidate work order, a score value for each influencing factor is calculated, ranging from 0 to 10 points. Finally, the scores of each influencing factor are multiplied by their corresponding weight coefficients and summed to obtain the comprehensive priority score for each work order. The candidate work orders are sorted in descending order of comprehensive priority score, with priority scores of 7 or higher being given priority. At the same time, the optimal combination of operation and maintenance resources is matched for each work order by combining the spatiotemporal map of available operation and maintenance resources to avoid resource conflicts and idleness, which can shorten the average processing time of work orders in multi-defect concurrent scenarios.

[0050] Step 4.3. Work Order Activation: By comparing the actual temperature of key components of the overhead contact system with a set threshold, the corresponding alternative handling work order is activated when the actual temperature reaches the set threshold. The set threshold is set according to the warning level; different warning levels correspond to different set thresholds. When the actual temperature of a component reaches the set threshold for the corresponding warning level, a pre-generated alternative handling work order is automatically activated. In this embodiment, the set threshold is obtained through statistical analysis of historical defect data; different types of components and different environmental conditions correspond to different set thresholds. When the actual temperature of a component reaches the set threshold, the system automatically marks the corresponding alternative handling work order as active, ready for dispatch.

[0051] Step 4.4. Work Order Dispatch and Adjustment: Activated work orders are dispatched to the corresponding handling entities. The spatiotemporal map of operational resources, containing the spatiotemporal distribution information of operational personnel, vehicles, and tools, can be utilized. Candidate work orders include handling time limits, responsible entities, required resources, handling steps, and acceptance criteria. Thresholds are set according to the warning level. Work orders are dispatched to the terminal devices of the corresponding handling entities via a standardized communication interface. In this embodiment, the standardized communication interface uses an HTTP interface, supports JSON data format, and can interact with various terminal devices. After work orders are dispatched, the system tracks the execution status of the work orders in real time, including the receipt status, handling progress, and handling results.

[0052] In some embodiments, after an activated work order is dispatched to the corresponding handling entity, temperature data of key components of the overhead contact system and status data of maintenance resources are acquired in real time. The content and dispatch recipients of the work order are adjusted based on the temperature data and the status data of the maintenance resources. Temperature data is collected at set time intervals, and the status data of maintenance resources is updated according to a set cycle. The adjustment of the work order content includes modifications to the handling time limit, required resources, and handling steps. The adjustment of the dispatch recipients is based on the real-time load of the maintenance resources. The adjusted work order is then re-dispatched to the terminal equipment of the corresponding handling entity through a standardized communication interface.

[0053] In some embodiments, after step 4, the execution result data of the handling work order is obtained and fed back to the gray system model and the time-series Transformer model to update the parameters of the gray system model and the time-series Transformer model. The execution result data includes the defect elimination status, actual handling time, and resource consumption. The parameter updates are performed using incremental learning. When updating the parameters of the gray system model, the fitting curve of the temperature evolution sequence is corrected using the execution result data. When updating the parameters of the time-series Transformer model, the model weights are adjusted using the execution result data. The updated gray system model and time-series Transformer model are used for subsequent prediction of the evolution of thermal defects in the overhead contact system.

[0054] This embodiment integrates multi-source detection data and operation and maintenance data of the overhead contact system to construct a complete technical system for predicting the evolution of thermal defects and pre-generating work orders. It achieves fully automated processing of thermal defects from early identification and evolution prediction to preventative work order management. This embodiment uses a grey system model to process the time-series temperature data of overhead contact system components, effectively identifying potential early thermal defects and detecting fault hazards in advance. It uses a time-series Transformer model to predict the evolution trend of defects, accurately predicting component temperature changes and critical failure times, providing a basis for operation and maintenance decisions. By pre-generating alternative disposal work orders and dynamically activating and adjusting them according to actual conditions, it optimizes the allocation of operation and maintenance resources and improves the efficiency and timeliness of defect disposal. This embodiment can be seamlessly integrated with existing overhead contact system operation and maintenance systems without large-scale modifications, exhibiting good compatibility and scalability. By establishing a closed-loop feedback mechanism and continuously optimizing model parameters using disposal result data, this embodiment can continuously improve the system's prediction accuracy and the scientific nature of operation and maintenance decisions, helping to reduce the incidence of overhead contact system failures and ensure the safe and stable operation of electrified railways.

[0055] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting the evolution of thermal defects in overhead contact lines and pre-generating work orders, characterized in that, Includes the following steps: S1. Integrate multi-source data related to the overhead contact system, construct a component-level time-series dataset according to the unique identifier of the key components of the overhead contact system, and extract the statistical and static features of the key components of the overhead contact system; S2. Based on the component-level time series dataset, the gray system model is used to amplify the temperature evolution trend of key components of the overhead contact system and identify potentially defective components; S3. Input the time-series data processed by the grey system model and the static characteristics of the key components of the overhead contact system into the time-series Transformer model, and output the temperature prediction results and critical failure time intervals of the key components of the overhead contact system. S4. Based on temperature prediction results and critical failure time intervals, conduct operation and maintenance resource assessment, pre-generate alternative disposal work orders, activate the corresponding disposal work orders according to the actual status of key components of the overhead contact system, and complete the dispatch.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Acquire contact network infrared temperature data, visible light image data, positioning data, environmental sensor data, line load data, equipment basic information, and historical work order processing data; S1.

2. The acquired multi-source data are time-aligned and spatial-aligned according to the unique identifiers of the key components of the overhead contact system to generate a cross-cycle component-level time-series dataset; S1.

3. Extract statistical and static features of key overhead contact system components from the component-level time-series dataset. Time alignment is based on a unified timestamp, and spatial alignment is based on location data and overhead contact system support numbers. Statistical features include mean temperature, standard deviation of temperature, rate of temperature rise, and temperature distribution entropy. Static features include component type, installation location, years of operation, and maintenance records. The component-level time-series dataset is stored in chronological order and supports retrieval by the unique identifier of the key overhead contact system component.

3. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.

1. Perform cumulative generation operation on the temperature time series data in the component-level time series dataset to reduce the randomness of the temperature time series data; S2.

2. The gray GM(1,1) model is used to fit the processed temperature time series data to generate a temperature evolution sequence; S2.

3. Calculate the residual between the temperature evolution sequence and the actual temperature sequence, identify potential defective components, and perform cumulative generation operation to accumulate the temperature time series data item by item in chronological order. The gray GM(1,1) model obtains the fitting curve of temperature evolution by solving the differential equation. The residual is the difference between the temperature evolution sequence and the actual temperature sequence. When the residual continuously exceeds the set value, the corresponding contact wire key component is determined to be a potential defective component. The information of the identified potential defective components is input into the subsequent detection frequency adjustment step.

4. The method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Construct a temporal Transformer model, which includes an input layer, a multi-head self-attention layer, a feedforward neural network layer, and an output layer; S3.

2. Input the temperature time series data, environmental parameter time series data, line load time series data and static characteristics of key components of the catenary after processing by the gray system model into the input layer of the time series Transformer model; S3.

3. Capture feature dependencies at different time scales through a multi-head self-attention layer, and perform feature transformation through a feedforward neural network layer; S3.

4. The output layer outputs the temperature prediction results and critical failure time intervals of key components of the overhead contact system. The input layer performs dimensional transformation and normalization on the input data. The multi-head self-attention layer contains multiple parallel attention heads, each of which independently calculates the attention weights of the features. The feedforward neural network layer contains two linear transformation layers and one activation function layer. The output layer maps the features to the temperature prediction results and critical failure time intervals through linear transformation.

5. The method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. Obtain usage status data of operation and maintenance resources and railway track maintenance window plan data, and construct a spatiotemporal map of available operation and maintenance resources; S4.

2. Based on the temperature prediction results and the critical failure time interval, the early warning level is divided, and alternative disposal work orders are pre-generated for different early warning levels; S4.

3. Compare the actual temperature of key components of the overhead contact system with the set threshold. When the actual temperature reaches the set threshold, activate the corresponding alternative handling work order. S4.

4. The activated handling work order is dispatched to the corresponding handling entity. The spatiotemporal map of operation and maintenance resources can be used to include the spatiotemporal distribution information of operation and maintenance personnel, operation and maintenance vehicles and operation and maintenance tools. The alternative handling work order includes the handling time limit, responsible entity, required resources, handling steps and acceptance criteria. The threshold is set according to the warning level. The handling work order is dispatched to the terminal device of the corresponding handling entity through a standardized communication interface.

6. The method according to claim 2, characterized in that, In step S1.3, after extracting the statistical and static features of the key components of the overhead contact system, the grey relational analysis method is used to calculate the correlation between the statistical and static features and the defect evolution. Features with a correlation greater than a set value are retained to generate a feature subset. The grey relational analysis method calculates the grey relational coefficient between each feature and the defect evolution. The correlation is obtained by the average value of the correlation coefficients. The set value is set according to the statistical results of historical data. The feature subset contains all statistical and static features with a correlation greater than the set value. The feature subset is input into the grey system model and the time series Transformer model for subsequent defect evolution prediction.

7. The method according to claim 3, characterized in that, In step S2.3, after identifying potentially defective components, the detection frequency of key components of the overhead contact system is adjusted, dense time-series data of the potentially defective components are obtained, and the component-level time-series dataset is updated. The detection frequency adjustment is set according to the residual size of the potentially defective components. The dense time-series data includes infrared temperature data, visible light image data, and environmental sensor data with higher time resolution. When updating the component-level time-series dataset, the dense time-series data is inserted into the original data according to the unique identifier and timestamp of the key components of the overhead contact system. The updated component-level time-series dataset is re-input into the gray system model for a new round of defect evolution analysis.

8. The method according to claim 4, characterized in that, In step S3.4, when outputting the temperature prediction results and critical failure time intervals of key components of the overhead contact system, the Monte Carlo dropout method is used to calculate the confidence level of the temperature prediction results and critical failure time intervals to generate risk assessment data. The Monte Carlo dropout method randomly discards some neurons in the feedforward neural network layer and performs multiple forward inferences. The confidence level interval of the temperature prediction results and critical failure time intervals is calculated by the statistical distribution of the results of multiple inferences. The risk assessment data is classified according to the confidence level intervals and input into the work order pre-generation step to adjust the disposal priority and resource allocation of alternative disposal work orders.

9. The method according to claim 5, characterized in that, In step S4.4, after the activated disposal work order is dispatched to the corresponding disposal entity, the temperature data of the key components of the overhead contact system and the status data of the maintenance resources are acquired in real time. The content and dispatch object of the disposal work order are adjusted according to the temperature data and the status data of the maintenance resources. The temperature data is collected at a set time interval, and the status data of the maintenance resources is updated at a set period. The adjustment of the content of the disposal work order includes the modification of the disposal time limit, required resources and disposal steps. The dispatch object is adjusted according to the real-time load of the maintenance resources. The adjusted disposal work order is re-dispatched to the terminal equipment of the corresponding disposal entity through the standardized communication interface.

10. The method according to claim 1, characterized in that, After step S4, the execution result data of the handling work order is obtained and fed back to the gray system model and the time-series Transformer model. The parameters of the gray system model and the time-series Transformer model are updated. The execution result data includes the defect elimination status, actual handling time and resource consumption. The parameter update is carried out in an incremental learning manner. When updating the parameters of the gray system model, the fitting curve of the temperature evolution sequence is corrected using the execution result data. When updating the parameters of the time-series Transformer model, the weights of the model are adjusted using the execution result data. The updated gray system model and time-series Transformer model are used for subsequent prediction of the evolution of thermal defects in the overhead contact system.

Images