A method and system for early warning of the degree of contamination of a roof high-voltage cable terminal

Abstract

The application provides a kind of roof high-voltage cable terminal contamination degree early warning method and system, belong to rail transit electrical equipment condition monitoring technical field.The method includes: collecting data and calculating static contamination distribution characteristic value, dynamic discharge activity characteristic value and environmental aggravating factor;The static and dynamic characteristic values are fused after being corrected using the environmental aggravating factor, to obtain the comprehensive contamination risk index;The reinforcement learning model is constructed and trained, the historical operation data, the current comprehensive contamination risk index, the environmental parameters and the early warning history are used as the state space, the correction coefficient, the fusion weight and the early warning threshold are used as the action space, the early warning accuracy and the operation and maintenance cost are used as the reward function, and the optimized parameters are output;The system is updated using the optimized parameters, the comprehensive contamination risk index calculated from the subsequent collected data is compared with the updated early warning threshold, and the corresponding level of early warning signal is triggered.The application can effectively improve the accuracy of pollution flashover risk prediction and the reliability of early warning.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring technology for railway electrical equipment, and in particular to a method and system for early warning of pollution levels in high-voltage cable terminals on train roofs. Background Technology

[0002] In electrified railway and urban rail transit systems, trains typically use overhead contact lines for power collection. The high-voltage cable terminals on the roof, as critical electrical connection components, are responsible for transmitting electrical energy from the pantograph to the traction power unit. Because the high-voltage equipment on the roof is constantly exposed to a complex external environment, dust, salt spray, industrial pollutants, bird droppings, and other contaminants easily accumulate on the surface of the cable terminals. When encountering high-humidity weather conditions such as fog, dew, rain, or snow, the soluble salts in the contaminant layer dissolve in the moisture, forming conductive paths on the insulation surface. This significantly reduces insulation performance and, in severe cases, can cause flashover accidents, leading to power outages and affecting train operation safety.

[0003] Currently, the main methods for detecting the pollution status of external insulation equipment include traditional contact detection methods, leakage current monitoring methods, non-contact optical detection methods, and artificial intelligence and multi-source information fusion methods. Traditional contact detection methods require disassembling the external insulation equipment for offline measurement, resulting in poor timeliness and difficulty in meeting real-time monitoring needs. Although leakage current monitoring methods are widely used, they are affected by factors such as material, pollution distribution, and humidity, making it difficult to establish a stable correlation between pollution levels, and fixed threshold alarms are prone to false alarms or missed alarms. Optical detection methods, such as infrared, ultraviolet, and visible light imaging, are easily affected by environmental interference or have limited wavelength information, making it difficult to accurately assess the pollution level. Artificial intelligence fusion methods are mostly aimed at external insulation equipment in power systems and focus on single information sources, lacking a comprehensive early warning mechanism for high-voltage cable terminals on vehicle roofs.

[0004] In summary, the existing technologies have the following shortcomings: First, they rely on a single parameter for pollution assessment, which makes it difficult to fully reflect the complex coupling relationship between pollution status and environmental factors; second, fixed threshold alarm models have poor adaptability and cannot be dynamically adjusted according to operating conditions and environmental changes; and third, they lack a comprehensive consideration of static pollution distribution characteristics and dynamic discharge activity, making it difficult to achieve accurate graded early warning.

[0005] To address the aforementioned issues, there is an urgent need for an intelligent early warning method for the pollution level of rooftop high-voltage cable terminals that can integrate image information, leakage current characteristics, and environmental parameters, and possess adaptive optimization capabilities, in order to improve the accuracy of pollution flashover risk prediction and the reliability of early warning. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for early warning of pollution levels in high-voltage cable terminals on vehicle roofs, which solves the technical problems of existing methods for early warning of pollution levels in high-voltage cable terminals on vehicle roofs, such as reliance on a single parameter for evaluation, poor adaptability of fixed threshold alarm models, and lack of comprehensive consideration of static pollution distribution characteristics and dynamic discharge activity, resulting in insufficient accuracy and reliability of early warning.

[0007] To achieve the above objectives, the present invention provides a method for early warning of the contamination level of a high-voltage cable terminal on a vehicle roof, comprising the following steps: Step S1: Collect surface images, leakage current signals, and environmental parameters of the high-voltage cable terminal on the vehicle roof, and calculate the static pollution distribution characterization value, dynamic discharge activity characterization value, and environmental aggravation factor, respectively. Step S2: Correct the static pollution distribution characterization value and the dynamic discharge activity characterization value using environmental aggravation factors to obtain the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value. Step S3: Combine the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value to obtain the comprehensive pollution risk index; Step S4: Construct and train a reinforcement learning model, using historical operating data, current comprehensive pollution risk index, environmental parameters and warning history as the state space, correction coefficient, fusion weight and warning threshold as the action space, warning accuracy and operation and maintenance cost as the reward function, and output the optimized correction coefficient, fusion weight and warning threshold. Step S5: Update the correction coefficients used in step S2 and the fusion weights used in step S3 using the optimized correction coefficients and fusion weights, respectively. Step S6: Update the preset warning threshold using the optimized warning threshold; Step S7: Compare the comprehensive pollution risk index calculated from the subsequently collected data with the updated warning threshold, and trigger the corresponding level of warning signal based on the comparison result.

[0008] Preferably, the calculation of the static pollution distribution characterization value in step S1 specifically includes: Step S101: Set up two cameras to capture surface images of the high-voltage cable terminal on the roof from different angles; Step S102: Detect the ambient light intensity. When the ambient light intensity is lower than the preset light threshold, activate the supplementary lighting device to provide auxiliary lighting. Step S103: Perform contrast enhancement, noise reduction, and perspective correction on the acquired original image to obtain a preprocessed image; Step S104: Input the preprocessed image into a lightweight semantic segmentation network to identify and segment the dirty areas on the surface of the cable terminal; Step S105: Calculate the area ratio of the contaminated area on the surface of the cable terminal. ,in To reduce the total number of dirty region pixels output by a lightweight semantic segmentation network, This represents the total number of pixels in the cable termination area. Step S106: Divide the cable terminal surface into For each grid sub-region, the proportion of dirty pixels within each sub-region is counted, and the uniformity of dirt distribution is calculated. ,in The variance of the percentage of contamination in each grid sub-region. For the first Line 1 The percentage of dirty pixels in the column grid sub-regions This represents the average percentage of dirty pixels across all grid sub-regions. This represents the theoretical maximum variance. Step S107: Determine the area proportion and uniformity By performing weighted combination, the static pollution distribution characterization value is obtained. ,in and The preset static weight coefficients.

[0009] Preferably, the calculation of the dynamic discharge activity characterization value in step S1 specifically includes: Step S111: Perform 3-level wavelet packet decomposition on the acquired leakage current signal to obtain wavelet packet coefficients for 8 frequency bands. ,in Frequency band number, , The wavelet packet coefficient index, with a value range from 1 to... , For the first The number of wavelet packet coefficients in each frequency band; Step S112: Calculate the energy value of the wavelet packet coefficients for each frequency band. The energy values ​​of each frequency band are normalized to obtain the normalized energy characteristics of each frequency band. ,in Frequency band number, , For the first Energy values ​​of wavelet packet coefficients in each frequency band; Step S113: Assign frequency band weighting coefficients to different frequency bands based on the correlation between each frequency band and historical flashover accidents. ; Step S114: Normalize the energy characteristics of each frequency band. With the corresponding frequency band weighting coefficient By performing weighted summation, the dynamic discharge activity characterization value is obtained. .

[0010] Preferably, the calculation of the environmental aggravation factor in step S1 specifically includes: Step S121: Synchronously collect the temperature around the cable terminal. ,humidity Wind speed and particulate matter concentration ; Step S122: Collect the temperature ,humidity Wind speed and particulate matter concentration Inputting a pre-trained multilayer perceptron model yields environmental aggravation factors. .

[0011] Preferably, step S2 specifically includes: Step S21: The static pollution distribution characterization value... Multiplying by the first correction factor yields the corrected static pollution distribution characterization value. ,in The static environmental impact coefficient has a first correction factor of 1. ; Step S22: The dynamic discharge activity characterization value... Multiplying by the second correction factor yields the corrected dynamic discharge activity characterization value. ,in The dynamic environmental impact coefficient is [value], and the second correction coefficient is [value]. .

[0012] Preferably, step S3 specifically includes: processing the corrected static pollution distribution characterization values. and the corrected dynamic discharge activity characterization values By performing weighted summation, a comprehensive pollution risk index is obtained. ,in For fusion weighting coefficients.

[0013] Preferably, the reinforcement learning model in step S4 is a deep deterministic policy gradient network, specifically including: Step S41: Define the state space ,in The rate of change of the comprehensive pollution risk index over the past hour. This is the train's current operating speed. This represents the cumulative mileage since the last cleaning. This represents the number of alerts triggered in the past 7 days. Step S42: Define the motion space ,in and These are the first warning threshold and the second warning threshold, respectively. Step S43: Define the reward function ; in, To accurately predict reward items, For false alarm penalties, For the penalty items for failure to report, To normalize maintenance costs, , , This is the preset penalty coefficient; Step S44: Construct an experience replay pool using historical running data, based on the current state. Input to maximize cumulative rewards For the goal, among which Using the discount factor, train a deep deterministic policy gradient network to output the optimal action. ,in The optimal static environmental impact coefficient is... The optimal dynamic environmental impact coefficient. To achieve the optimal fusion weight coefficient, The optimal first warning threshold is... The optimal second warning threshold is determined by this threshold.

[0014] Preferably, step S7 specifically includes: Step S71: Preset the first warning threshold Second warning threshold The second warning threshold is greater than the first warning threshold. The first and second warning thresholds are preset empirical values ​​when the system is initially running. After the update in step S6, the updated warning thresholds are used. Step S72, when the comprehensive pollution risk index satisfy At that time, the first-level warning is triggered; Step S73, when the comprehensive pollution risk index satisfy At that time, a level two warning is triggered; The second-level warning level is higher than the first-level warning level.

[0015] This invention also provides a vehicle roof high-voltage cable terminal pollution warning system, used to implement the above-described vehicle roof high-voltage cable terminal pollution warning method, including: The data acquisition module is used to collect surface images, leakage current signals, and environmental parameters of the high-voltage cable terminals on the vehicle roof. The static fouling calculation module is used to calculate the static fouling distribution characterization value based on the surface image; The dynamic discharge calculation module is used to calculate the dynamic discharge activity characterization value based on the leakage current signal; The environmental factor calculation module is used to calculate environmental aggravation factors based on environmental parameters. The correction module is used to correct the static pollution distribution characterization value and the dynamic discharge activity characterization value using environmental aggravation factors, respectively, to obtain the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value. The fusion module is used to fuse the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value to obtain a comprehensive pollution risk index. The reinforcement learning optimization module is used to build and train a reinforcement learning model. It uses historical operating data, the current comprehensive pollution risk index, environmental parameters, and historical warnings as the state space, correction coefficients, fusion weights, and warning thresholds as the action space, and warning accuracy and operation and maintenance costs as the reward functions. The module outputs the optimized correction coefficients, fusion weights, and warning thresholds. The reinforcement learning model is a deep deterministic policy gradient network. The parameter update module is used to update the corresponding parameters used in the correction module and the fusion module using the optimized correction coefficients and fusion weights, and to update the preset warning threshold using the optimized warning threshold. The early warning judgment module is used to compare the comprehensive pollution risk index calculated from subsequent collected data with the updated early warning threshold, and trigger the corresponding level of early warning signal based on the comparison result. The communication module is used to provide local alerts via the vehicle-mounted display terminal after a warning signal is triggered, and simultaneously transmit the warning information to the remote operation and maintenance platform via the wireless communication module.

[0016] Therefore, the present invention employs the above-mentioned method and system for early warning of contamination levels in high-voltage cable terminals on vehicle roofs, and the beneficial technical effects are as follows: (1) This invention overcomes the limitations of existing technologies that rely on a single parameter for pollution assessment by constructing a three-level index system consisting of static pollution distribution characterization value, dynamic discharge activity characterization value, and environmental aggravation factor. Specifically, the static pollution distribution characterization value extracts the area ratio and distribution uniformity of polluted regions based on a semantic segmentation network, comprehensively reflecting the spatial distribution characteristics of pollution; the dynamic discharge activity characterization value extracts multi-frequency band energy features from the leakage current through wavelet packet decomposition, effectively characterizing the discharge activity caused by pollution; and the environmental aggravation factor quantifies the degree to which environmental factors exacerbate the risk of pollution flashover. By organically integrating these three elements, the complex coupling relationship between pollution status and environmental factors can be comprehensively reflected, significantly improving the accuracy of pollution level assessment.

[0017] (2) This invention constructs a deep deterministic policy gradient network, using historical operating data, comprehensive pollution risk index, environmental parameters, and early warning history as the state space, correction coefficients, fusion weights, and early warning thresholds as the action space, and early warning accuracy and operation and maintenance cost as the reward function, to train a reinforcement learning model to output optimal parameters. This mechanism overcomes the poor adaptability of fixed threshold alarm models in existing technologies, enabling the system to dynamically adjust correction coefficients, fusion weights, and early warning thresholds according to the operating conditions of different routes, seasons, and vehicle models, achieving adaptive optimization of the early warning model and effectively reducing false alarm rate and missed alarm rate.

[0018] (3) This invention utilizes environmental aggravation factors to correct the static pollution distribution characterization value and the dynamic discharge activity characterization value, respectively, reflecting the physical coupling mechanism of "environment-pollution-discharge". After correction, the values ​​are then fused to obtain the comprehensive pollution risk index. By setting multi-level early warning thresholds, corresponding early warning signals are triggered when the comprehensive pollution risk index reaches different levels, realizing the leap from "current state judgment" to "future risk prediction". This mechanism can not only assess the current pollution state, but also predict future flashover risks based on environmental change trends, providing maintenance personnel with differentiated response strategies and significantly improving the level of intelligent operation and maintenance of rail transit. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for early warning of the contamination level of a high-voltage cable terminal on a vehicle roof according to the present invention; Figure 2 To enhance learning and training and optimize flowcharts; Figure 3 This is an architectural diagram of a vehicle roof high-voltage cable terminal pollution warning system according to the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] Example 1 like Figures 1-2 As shown, a method for early warning of pollution levels in a high-voltage cable terminal on a vehicle roof includes the following steps: Step S1: Collect surface images, leakage current signals, and environmental parameters of the high-voltage cable terminal on the vehicle roof, and calculate the static pollution distribution characterization value, dynamic discharge activity characterization value, and environmental aggravation factor, respectively.

[0023] (1) Calculation of static pollution distribution characterization value.

[0024] Two industrial-grade high-definition cameras are installed at appropriate locations on the roof, pointing at the cable terminals from different angles to ensure complete coverage of the cable terminal surface. The cameras have built-in light sensors to detect ambient light intensity in real time. When the light intensity falls below a preset threshold (e.g., 50 lux), an auxiliary lighting device (such as an LED array) is automatically activated to ensure image quality.

[0025] The acquired raw images are first preprocessed: histogram equalization is used to enhance contrast, median filtering is used to remove noise, and perspective correction algorithm is used to transform images from different angles to a standard viewpoint (e.g., front view of a cable terminal) to obtain the preprocessed image.

[0026] The preprocessed image is input into a lightweight semantic segmentation network. This embodiment employs a semantic segmentation network based on a MobileNetV3 encoder and a DeepLabV3 decoder. This network has been trained on a training set containing images of various types of contaminated cable terminals. The training set images are pixel-level labeled with contaminated areas, cable terminal background areas, etc. The network outputs the class probability of each pixel, which is then processed by argmax to obtain a segmentation mask, thereby identifying and segmenting the contaminated areas on the surface of the cable terminal.

[0027] Calculate the percentage of the contaminated area on the cable terminal surface. ,in To reduce the total number of dirty region pixels output by a lightweight semantic segmentation network, This represents the total number of pixels in the cable terminal area.

[0028] The cable terminal surface is divided into Each grid sub-region (in this embodiment, we take...) =10, =10), count the proportion of dirty pixels in each sub-region, and calculate the uniformity of dirt distribution. ,in The variance of the percentage of contamination in each grid sub-region. For the first Line 1 The percentage of dirty pixels in the column grid sub-regions This represents the average percentage of dirty pixels across all grid sub-regions. This represents the theoretical maximum variance. The value range is [0,1]. The larger the value, the more evenly the dirt is distributed, and the smaller the value, the more concentrated the dirt is.

[0029] Area proportion and uniformity By performing weighted combination, the static pollution distribution characterization value is obtained. ,in and For preset static weight coefficients, and In this embodiment, based on experience, =0.6, =0.4, which means that more emphasis is placed on the contribution of the dirty area.

[0030] (2) Calculation of dynamic discharge activity characterization value.

[0031] Leakage current signals were collected using a high-frequency current transformer installed on the grounding wire of the cable terminal. The sampling frequency was set to 10 kHz, and each acquisition lasted for 1 second. The collected leakage current signals were then subjected to three-level wavelet packet decomposition to obtain wavelet packet coefficients in eight frequency bands. ,in Frequency band number, , The wavelet packet coefficient index, with a value range from 1 to... , For the first Number of wavelet packet coefficients per frequency band.

[0032] Calculate the energy value of wavelet packet coefficients in each frequency band. The energy values ​​of each frequency band are normalized to obtain the normalized energy characteristics of each frequency band. ,in Frequency band number, , For the first Energy values ​​of wavelet packet coefficients in each frequency band; Based on the correlation between each frequency band and historical flashover accidents, frequency band weighting coefficients are assigned to different frequency bands. Frequency band weighting coefficient satisfy Furthermore, frequency bands with a higher correlation to historical flashover accidents are assigned a larger frequency band weighting coefficient. In this embodiment, the frequency band weighting coefficient... Set according to Table 1 (specific values ​​can be adjusted based on actual data analysis): Table 1 Weighting Coefficient Settings

[0033] Normalized energy characteristics of each frequency band With the corresponding frequency band weighting coefficient By performing weighted summation, the dynamic discharge activity characterization value is obtained. .

[0034] The value ranges from [0,1]. The larger the value, the stronger the discharge activity and the more severe the pollution.

[0035] (3) Calculation of environmental aggravation factors.

[0036] An environmental sensor was installed near the cable terminal on the roof to collect temperature data synchronously. (Unit: °C) Humidity (Unit: %RH), Wind Speed (Unit: m / s) and particulate matter concentration (Unit: μg / m) 3 (Including a weighted or combined value of PM2.5 and PM10). Data is collected every 10 minutes.

[0037] Collected temperature ,humidity Wind speed and particulate matter concentration Inputting a pre-trained multilayer perceptron model yields environmental aggravation factors. .

[0038] The multilayer perceptron model structure is as follows: 4 neurons in the input layer (corresponding to temperature). ,humidity Wind speed and particulate matter concentration The model consists of two hidden layers, each with 64 neurons and ReLU activation function, and an output layer with one neuron and Sigmoid activation function, compressing the output to the [0,1] interval. The model training data comes from historical flashover accident records: environmental parameters from one hour prior to each flashover accident are selected as input, and labels are determined by experts based on a comprehensive assessment of the accident severity and environmental adverseness (scoring range 0-1). Training uses the mean squared error loss function, the Adam optimizer, a learning rate of 0.001, and 200 iterations. The trained model can output an environmental aggravation factor in real time based on the current environmental parameters; the closer this value is to 1, the higher the degree of aggravation of flashover risk by environmental conditions.

[0039] Step S2: Correct the static pollution distribution characterization value and the dynamic discharge activity characterization value using environmental aggravation factors to obtain the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value.

[0040] Step S21: The static pollution distribution characterization value... Multiplying by the first correction factor yields the corrected static pollution distribution characterization value. ,in The static environmental impact coefficient has a first correction factor of 1. .

[0041] Step S22: The dynamic discharge activity characterization value... Multiplying by the second correction factor yields the corrected dynamic discharge activity characterization value. ,in The dynamic environmental impact coefficient is [value], and the second correction coefficient is [value]. .

[0042] Environmental factors (especially humidity and particulate matter) can significantly alter the conductivity and discharge behavior of contaminant layers. For example, under high humidity conditions, dry contaminants are normally non-conductive, but dissolved salts after wetting can form a conductive layer, dramatically increasing the risk of static contamination. Simultaneously, humidity also enhances the amplitude and activity of leakage current, thus increasing dynamic discharge activity. Directly using static contamination distribution characterization values... and dynamic discharge activity characterization value This fails to reflect the amplifying effect of the environment on risk. By introducing environmental aggravating factors... Making corrections is equivalent to dynamically adjusting the basic risks based on current environmental conditions: when environmental conditions are adverse, Close to 1, after correction and Significantly magnified; when environmental conditions are favorable, When the value is close to 0, the corrected value is essentially equal to the original value. This correction mechanism reflects the physical coupling relationship between "environment-pollution-discharge," making the risk assessment more closely reflect the actual physical process.

[0043] During initial system startup, and Set to preset empirical values: based on the cable terminal material (such as silicone rubber, porcelain, etc.) and the characteristics of historical operating line environments, for example, for rainy and humid areas, Take 0.3, Take 0.4; for dry and dust-free areas, Take 0.1, Let's take 0.2. In this embodiment, the initial value is... =0.2, =0.3.

[0044] Step S3: Combine the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value to obtain the comprehensive pollution risk index.

[0045] Corrected static pollution distribution characterization values and the corrected dynamic discharge activity characterization values By performing weighted summation, a comprehensive pollution risk index is obtained. ,in To integrate the weighting coefficients, and ; Fusion weighting coefficient During initial system operation, preset empirical values ​​are used, determined based on the historical contribution rates of static pollution and dynamic discharge to flashover accidents. For example, these values ​​are set by statistically analyzing the correlation between the pollution area and leakage current energy in historical accident images. =0.5 indicates that both are equally important. This embodiment initially takes... =0.5. Comprehensive Pollution Risk Index Used to comprehensively assess the current contamination risk of cable terminals.

[0046] Step S4: Build and train the reinforcement learning model, and output the optimized correction coefficients, fusion weights and warning thresholds.

[0047] This embodiment employs a Deep Deterministic Policy Gradient (DDPG) network as the reinforcement learning model to achieve adaptive parameter optimization. The significance of introducing reinforcement learning lies in the correction coefficients in steps S2 and S3. , and fusion weight and warning thresholds and The initial values ​​are all based on experience, but the optimal parameters may differ for different routes, seasons, and vehicle models. Furthermore, pollution characteristics and environmental features may change over time. Fixed parameters cannot adapt to these dynamic changes, easily leading to a decline in early warning performance. Reinforcement learning, through continuous interaction with the environment, automatically learns the optimal parameter combination with early warning accuracy and operational costs as optimization objectives. This achieves adaptive optimization of the early warning model, maintaining high accuracy and low false alarm / missed alarm rates under different operating conditions.

[0048] The specific settings are as follows: Step S41: Define the state space ,in The rate of change of the comprehensive pollution risk index over the past hour (using the slope of a sliding window linear regression). The current operating speed of the train (obtained from the train control system). This represents the cumulative mileage since the last cleaning. This represents the number of warnings triggered in the past 7 days.

[0049] Step S42: Define the motion space ,in and These are the first and second warning thresholds, respectively. All action variables take values ​​continuously within a reasonable range, and the network output layer is scaled to the corresponding interval after being activated with Sigmoid or Tanh.

[0050] Step S43: Define the reward function ; in, To ensure accurate early warning rewards, when a flashover incident occurs within 24 hours of an early warning, ,otherwise ; As a penalty for false alarms, if no flashover occurs within 24 hours after the warning is triggered and the humidity is below 60%, ,otherwise ; As a penalty for underreporting, if a flashover incident occurs within 24 hours but no warning is triggered. ,otherwise ; To normalize maintenance costs, ,in This represents the cumulative mileage of this cleaning operation. The preset maximum cleaning interval mileage; , , These are preset penalty coefficients used to adjust the relative importance of false alarms, false negatives, and maintenance costs; in this embodiment, we take... =0.5, =0.8, =0.2, meaning that more emphasis is placed on the penalty for underreporting.

[0051] Step S44: Construct an experience replay pool using historical running data, based on the current state. For input (where (Time step number) to maximize cumulative rewards For the goal, among which A discount factor is used to train a deep deterministic policy gradient network (DDPG). DDPG consists of an Actor network (policy network) and a Critic network (value network), both using 3 fully connected neural networks with 128 hidden neurons each, and the activation function is ReLU. The Adam optimizer is used, with an Actor learning rate of 0.0001 and a Critic learning rate of 0.001. =0.99. The experience replay pool has a capacity of 10,000 records, and 64 records are sampled for training each time. The network parameters are updated every 100 time steps. The training data comes from the historical operation database, which includes operation data, environmental parameters, warning records, and flashover accident records from the past year. After training until the reward function converges, the network parameters are fixed for online output of the optimal action. ,in The optimal static environmental impact coefficient is... The optimal dynamic environmental impact coefficient. To achieve the optimal fusion weight coefficient, The optimal first warning threshold is... The optimal second warning threshold is determined by this threshold.

[0052] Step S5: Using the optimized correction coefficients and fusion weights, update the correction coefficients used in step S2 and the fusion weights used in step S3, respectively.

[0053] Every preset update cycle (24 hours in this embodiment), the optimal static environment influence coefficient output in step S4 is updated. Assign the value to the static environmental impact coefficient used in step S21 ; Optimal dynamic environmental impact coefficient Assign the value to the dynamic environmental impact coefficient used in step S22 ; Optimal fusion weight coefficients Assign values ​​to the fusion weight coefficients used in step S3 .

[0054] Step S6: Update the preset warning threshold using the optimized warning threshold.

[0055] Similarly, every 24 hours, the optimal first warning threshold output in step S4 will be... Assign a value to the preset first warning threshold The optimal second early warning threshold output in step S4 Assign a value to the preset second warning threshold .

[0056] The preset update cycle is one of 24 hours, 7 days or 30 days, or it is triggered when the train's cumulative running mileage reaches a preset mileage threshold.

[0057] Step S7: Compare the comprehensive pollution risk index calculated from the subsequently collected data with the updated warning threshold, and trigger the corresponding level of warning signal based on the comparison result.

[0058] Step S71: Preset the first warning threshold Second warning threshold The second warning threshold is higher than the first warning threshold. Both the first and second warning thresholds are preset empirical values ​​during the initial system operation; the updated warning thresholds are used after the update in step S6. The warning level of the second-level warning is higher than that of the first-level warning. Step S72, when the comprehensive pollution risk index satisfy When this happens, the first-level warning is triggered, indicating that "dirt is accumulating rapidly, and cleaning is recommended in the near future."

[0059] Step S73, when the comprehensive pollution risk index satisfy At that time, a Level 2 warning was triggered, indicating "There is a high risk of flashover. It is recommended to stop the vehicle immediately for inspection and take emergency measures."

[0060] Upon triggering the warning, the system will display the warning information locally via the vehicle-mounted display terminal and simultaneously transmit it to the remote operation and maintenance platform via a wireless communication module (such as 4G / 5G). The warning information includes: warning level, current comprehensive pollution risk index, static pollution distribution characterization value, corrected static pollution distribution characterization value, dynamic discharge activity characterization value, corrected dynamic discharge activity characterization value, environmental aggravation factors, and recommended maintenance measures. Maintenance personnel can then promptly arrange repairs based on this information to prevent flashover accidents.

[0061] The invention will be further illustrated below with specific examples.

[0062] A certain type of train is operating on a certain section of a railway line, and the early warning system of this invention is installed on its roof. During a data acquisition cycle, a camera captures images of cable terminals. After processing by a lightweight semantic segmentation network, the system identifies a contaminated area ratio of 18.3% and a contamination distribution uniformity of 0.92, resulting in a static contamination distribution characterization value of 0.48. A high-frequency current transformer collects leakage current signals. After three-layer wavelet packet decomposition and frequency band weighting, a dynamic discharge activity characterization value of 0.15 is calculated. Environmental sensors measure the temperature at 22.3℃, humidity at 85.2%, wind speed at 3.1 m / s, and particulate matter concentration at 75 μg / m³. 3 The input multilayer perceptron model outputs an environmental aggravation factor of 0.72.

[0063] The system uses the current static environmental impact coefficient of 0.20 and the dynamic environmental impact coefficient of 0.30 to correct the environmental aggravation factors, with the static characterization value corrected to 0.54 and the dynamic characterization value corrected to 0.19. Weighted fusion with the current fusion weight of 0.50 yields a comprehensive pollution risk index of 0.36.

[0064] The system completed the reinforcement learning model update in the early hours of the day, outputting optimal parameters based on the data from the past 30 days: static environment influence coefficient 0.18, dynamic environment influence coefficient 0.28, fusion weight 0.55, first warning threshold 0.35, and second warning threshold 0.65. The system automatically updated the above parameters.

[0065] Two hours later, the newly collected data calculated a comprehensive contamination risk index of 0.42. Compared with the updated threshold, 0.42 falls between 0.35 and 0.65, triggering a Level 1 yellow alert. The system pushed the alert information to the driver's cab display screen and the ground maintenance platform, prompting "Contamination is accumulating rapidly; cleaning is recommended soon." Subsequent inspection confirmed significant contamination on the cable terminal surface, verifying the accuracy of the alert.

[0066] To further verify the technical effect of the present invention, three consecutive months of operating data were selected from this line section, and comparative tests were conducted using different methods. The results are shown in Table 2.

[0067] Table 2 Comparison of Early Warning Performance of Different Methods

[0068] As shown in Table 2, the method of the present invention outperforms the comparative methods in terms of early warning accuracy, false alarm rate, and missed alarm rate, which fully demonstrates the effectiveness of multi-source information fusion and environmental correction.

[0069] Table 3 Comparison of Comprehensive Pollution Risk Index under Different Humidity Conditions

[0070] Table 3 shows that, under the same polluted conditions, the environmental aggravation factor increases in a humid environment. This invention improves the comprehensive risk index from 0.32 to 0.42 through correction, thus more accurately reflecting the actual flashover risk.

[0071] Table 4 Comparison of early warning thresholds and effects before and after reinforcement learning optimization

[0072] Table 4 shows that after reinforcement learning optimization, the warning threshold was increased, reducing the number of unnecessary warnings, while the number of missed warnings decreased significantly, indicating that the model found a better threshold balance point.

[0073] Example 2 like Figure 3 As shown, a vehicle roof high-voltage cable terminal pollution warning system includes: The data acquisition module is used to collect surface images, leakage current signals, and environmental parameters of the high-voltage cable terminals on the vehicle roof. The static fouling calculation module is used to calculate the static fouling distribution characterization value based on the surface image; The dynamic discharge calculation module is used to calculate the dynamic discharge activity characterization value based on the leakage current signal; The environmental factor calculation module is used to calculate environmental aggravation factors based on environmental parameters. The correction module is used to correct the static pollution distribution characterization value and the dynamic discharge activity characterization value using environmental aggravation factors, respectively, to obtain the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value. The fusion module is used to fuse the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value to obtain a comprehensive pollution risk index. The reinforcement learning optimization module is used to build and train a reinforcement learning model. It uses historical operating data, the current comprehensive pollution risk index, environmental parameters, and historical warnings as the state space, correction coefficients, fusion weights, and warning thresholds as the action space, and warning accuracy and operation and maintenance costs as the reward functions. The module outputs the optimized correction coefficients, fusion weights, and warning thresholds. The reinforcement learning model is a deep deterministic policy gradient network. The parameter update module is used to update the corresponding parameters used in the correction module and the fusion module using the optimized correction coefficients and fusion weights, and to update the preset warning threshold using the optimized warning threshold. The early warning judgment module is used to compare the comprehensive pollution risk index calculated from subsequent collected data with the updated early warning threshold, and trigger the corresponding level of early warning signal based on the comparison result. The communication module is used to provide local alerts via the vehicle-mounted display terminal after a warning signal is triggered, and simultaneously transmit the warning information to the remote operation and maintenance platform via the wireless communication module.

[0074] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0075] Therefore, this invention adopts the above-mentioned method and system for early warning of pollution levels in high-voltage cable terminals on vehicle roofs. By constructing a three-level index system of static pollution distribution characterization value, dynamic discharge activity characterization value, and environmental aggravation factor, and using the environmental aggravation factor to correct the static and dynamic indicators, a comprehensive pollution risk index is obtained by fusion. Furthermore, a reinforcement learning model is introduced to adaptively optimize the correction coefficient, fusion weight, and early warning threshold, thereby realizing the fusion of multi-source information and the dynamic adjustment of the early warning model. This system can comprehensively reflect the complex coupling relationship between pollution status and environmental factors, effectively overcoming the shortcomings of existing technologies that rely on single-parameter evaluation and have poor adaptability of fixed-threshold alarm models, and improving the accuracy of pollution flashover risk prediction and the reliability of early warning.

[0076] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for early warning of the degree of contamination of a roof high-voltage cable termination, characterized in that, Includes the following steps: Step S1: Collect surface images, leakage current signals, and environmental parameters of the high-voltage cable terminal on the vehicle roof, and calculate the static pollution distribution characterization value, dynamic discharge activity characterization value, and environmental aggravation factor, respectively. Step S2: Correct the static pollution distribution characterization value and the dynamic discharge activity characterization value using environmental aggravation factors to obtain the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value. Step S3: Combine the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value to obtain the comprehensive pollution risk index; Step S4: Construct and train a reinforcement learning model, using historical operating data, current comprehensive pollution risk index, environmental parameters and warning history as the state space, correction coefficient, fusion weight and warning threshold as the action space, warning accuracy and operation and maintenance cost as the reward function, and output the optimized correction coefficient, fusion weight and warning threshold. Step S5: Update the correction coefficients used in step S2 and the fusion weights used in step S3 using the optimized correction coefficients and fusion weights, respectively. Step S6: Update the preset warning threshold using the optimized warning threshold; Step S7: Compare the comprehensive pollution risk index calculated from the subsequently collected data with the updated warning threshold, and trigger the corresponding level of warning signal based on the comparison result; Step S1, which calculates the static pollution distribution characterization value, specifically includes: Step S101: Set up two cameras to capture surface images of the high-voltage cable terminal on the roof from different angles; Step S102: Detect the ambient light intensity. When the ambient light intensity is lower than the preset light threshold, activate the supplementary lighting device to provide auxiliary lighting. Step S103: Perform contrast enhancement, noise reduction, and perspective correction on the acquired original image to obtain a preprocessed image; Step S104: Input the preprocessed image into a lightweight semantic segmentation network to identify and segment the dirty areas on the surface of the cable terminal; Step S105: Calculate the area ratio of the contaminated area on the surface of the cable terminal. ,in To reduce the total number of pixels in contaminated regions output by a lightweight semantic segmentation network. This represents the total number of pixels in the cable termination area. Step S106: Divide the cable terminal surface into For each grid sub-region, the proportion of dirty pixels within each sub-region is counted, and the uniformity of dirt distribution is calculated. ,in Let Variance be the percentage of contamination in each grid sub-region. For the first Line 1 The percentage of dirty pixels in the column grid sub-regions. This represents the average percentage of contaminated pixels across all grid sub-regions. This represents the theoretical maximum variance. Step S107: Determine the area proportion and uniformity By performing weighted combination, the static pollution distribution characterization value is obtained. ,in and Preset static weight coefficients; The calculation of dynamic discharge activity characterization values ​​in step S1 specifically includes: Step S111: Perform 3-level wavelet packet decomposition on the acquired leakage current signal to obtain wavelet packet coefficients for 8 frequency bands. ,in Frequency band number, , The wavelet packet coefficient index, with a value range from 1 to... , For the first The number of wavelet packet coefficients in each frequency band; Step S112: Calculate the energy value of the wavelet packet coefficients for each frequency band. The energy values ​​of each frequency band are normalized to obtain the normalized energy characteristics of each frequency band. ,in Frequency band number, , For the first Energy values ​​of wavelet packet coefficients in each frequency band; Step S113: Assign frequency band weighting coefficients to different frequency bands based on the correlation between each frequency band and historical flashover accidents. ; Step S114: Normalize the energy characteristics of each frequency band. With the corresponding frequency band weighting coefficient By performing weighted summation, the dynamic discharge activity characterization value is obtained. ; Step S1, calculating the environmental aggravation factor, specifically includes: Step S121: Synchronously collect the temperature around the cable terminal. ,humidity Wind speed and particulate matter concentration ; Step S122: Collect the temperature ,humidity Wind speed and particulate matter concentration Inputting a pre-trained multilayer perceptron model yields environmental aggravation factors. .

2. The method for early warning of pollution level of a high-voltage cable terminal on a vehicle roof according to claim 1, characterized in that, Step S2 specifically includes: Step S21: The static pollution distribution characterization value... Multiplying by the first correction factor yields the corrected static pollution distribution characterization value. ,in The static environmental impact coefficient has a first correction factor of 1. ; Step S22: The dynamic discharge activity characterization value... Multiplying by the second correction factor yields the corrected dynamic discharge activity characterization value. ,in The dynamic environmental impact coefficient is [value], and the second correction coefficient is [value]. .

3. The method for early warning of pollution level of a high-voltage cable terminal on a vehicle roof according to claim 2, characterized in that, Step S3 specifically includes: analyzing the corrected static pollution distribution characterization values. and the corrected dynamic discharge activity characterization values By performing a weighted summation, a comprehensive pollution risk index is obtained. ,in For fusion weighting coefficients.

4. The method for early warning of pollution level of a high-voltage cable terminal on a vehicle roof according to claim 3, characterized in that, The reinforcement learning model in step S4 is a deep deterministic policy gradient network, specifically including: Step S41: Define the state space ,in The rate of change of the comprehensive pollution risk index over the past hour. The current operating speed of the train. This represents the cumulative mileage since the last cleaning. This represents the number of alerts triggered in the past 7 days. Step S42: Define the motion space ,in and These are the first warning threshold and the second warning threshold, respectively. Step S43: Define the reward function ; in, To accurately predict reward items, For false alarm penalties, For the penalty items for failure to report, To normalize maintenance costs, , , This is the preset penalty coefficient; Step S44: Construct an experience replay pool using historical running data, based on the current state. Input to maximize cumulative rewards For the goal, among which Using the discount factor, train a deep deterministic policy gradient network to output the optimal action. ,in The optimal static environmental impact coefficient is... The optimal dynamic environmental impact coefficient. To achieve the optimal fusion weight coefficient, The optimal first warning threshold is... The optimal second warning threshold is determined by this threshold.

5. The method for early warning of pollution level of a high-voltage cable terminal on a vehicle roof according to claim 4, characterized in that, Step S7 specifically includes: Step S71: Preset the first warning threshold Second warning threshold The second warning threshold is greater than the first warning threshold. The first and second warning thresholds are preset empirical values ​​when the system is initially running. After the update in step S6, the updated warning thresholds are used. Step S72, when the comprehensive pollution risk index satisfy At that time, the first-level warning is triggered; Step S73, when the comprehensive pollution risk index satisfy At that time, a Level 2 warning was triggered.

6. A vehicle roof high-voltage cable terminal pollution level early warning system, characterized in that, A method for implementing a pollution level warning system for a rooftop high-voltage cable terminal as described in any one of claims 1-5 includes: The data acquisition module is used to collect surface images, leakage current signals, and environmental parameters of the high-voltage cable terminals on the vehicle roof. The static fouling calculation module is used to calculate the static fouling distribution characterization value based on the surface image; The dynamic discharge calculation module is used to calculate the dynamic discharge activity characterization value based on the leakage current signal; The environmental factor calculation module is used to calculate environmental aggravation factors based on environmental parameters. The correction module is used to correct the static pollution distribution characterization value and the dynamic discharge activity characterization value using environmental aggravation factors, respectively, to obtain the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value. The fusion module is used to fuse the corrected static pollution distribution characterization value and the corrected dynamic discharge activity characterization value to obtain a comprehensive pollution risk index. The reinforcement learning optimization module is used to build and train a reinforcement learning model. It uses historical operating data, the current comprehensive pollution risk index, environmental parameters, and historical warnings as the state space, correction coefficients, fusion weights, and warning thresholds as the action space, and warning accuracy and operation and maintenance costs as the reward functions. The module outputs the optimized correction coefficients, fusion weights, and warning thresholds. The reinforcement learning model is a deep deterministic policy gradient network. The parameter update module is used to update the corresponding parameters used in the correction module and the fusion module using the optimized correction coefficients and fusion weights, and to update the preset warning threshold using the optimized warning threshold. The early warning judgment module is used to compare the comprehensive pollution risk index calculated from subsequently collected data with the updated early warning threshold, and trigger the corresponding level of early warning signal based on the comparison result. The communication module is used to provide local alerts via the vehicle-mounted display terminal after a warning signal is triggered, and simultaneously transmit the warning information to the remote operation and maintenance platform via the wireless communication module.

Images