A substation route inspection planning method based on deep learning

Abstract

The present invention discloses a substation route inspection planning method based on deep learning, which includes the following steps: obtaining substation equipment status data and environmental monitoring data; performing multi-dimensional feature analysis on the equipment status data based on a preset deep feature extraction model to generate an equipment health assessment vector; generating multiple candidate paths through a reinforcement learning algorithm based on the environmental monitoring data and the equipment health assessment vector, screening the multiple candidate paths to obtain an initial inspection path, and dynamically optimizing the initial inspection path by integrating geographic information in real time; based on the drone positioning data and equipment status feedback, if it is detected that the health is lower than the threshold or the environmental risk exceeds the limit, triggering local path re-planning, and giving priority to navigating to high-risk equipment. The present invention has the following advantages and effects: it can improve the intelligence, accuracy and safety of inspection route planning, and ensure that the drone can efficiently and reliably perform substation equipment inspection tasks.

Description

Technical Field

[0001] The present invention relates to drone inspection route planning, and in particular to a substation route inspection planning method based on deep learning. Background Art

[0002] As power systems continue to expand, substations, as a crucial component of power transmission and distribution networks, have a significant impact on the safety and reliability of these systems. Therefore, substation inspections have become a crucial component of ensuring power system security. Traditional substation inspections rely primarily on manual inspections and regular testing. However, these two methods suffer from low efficiency, limited coverage, and slow response times, making them inadequate for modern substation inspections.

[0003] With the development of drone and deep learning technologies, drone-based inspection solutions have become a hot topic in the power industry. Drones offer advantages such as efficiency, flexibility, and automation. They can quickly obtain the operating status of substation equipment, monitor equipment health in real time, and provide more comprehensive and accurate inspection data. However, in practice, due to the complexity of substation inspection areas and environmental uncertainty, traditional route planning methods struggle to cope with complex working environments, such as dynamic changes in equipment status and real-time changes in obstacles.

[0004] Existing drone-based inspection route planning methods have the following main deficiencies: First, they do not fully consider the impact of equipment status and environmental risks on the inspection route, resulting in low safety and efficiency of the inspection route; second, most path planning lacks real-time response to changes in equipment health and cannot adjust the inspection route in time to deal with equipment failures or environmental changes; third, most existing methods ignore the need for multi-objective optimization and cannot simultaneously take into account multiple factors such as efficiency, safety and emergency tasks.

[0005] Therefore, in response to the above problems, the present invention proposes a substation route inspection planning method based on deep learning, aiming to improve the intelligence, accuracy and safety of inspection path planning, and ensure that drones can perform substation equipment inspection tasks efficiently and reliably. Summary of the Invention

[0006] The purpose of the present invention is to provide a substation route inspection planning method based on deep learning to solve the problems raised in the background technology.

[0007] The above technical objectives of the present invention are achieved through the following technical solutions:

[0008] A substation route inspection planning method based on deep learning includes the following steps:

[0009] S100, obtaining substation equipment status data and environmental monitoring data; wherein the equipment status data includes equipment temperature, vibration frequency, and insulation resistance value, and the environmental monitoring data includes meteorological parameters, light intensity, and obstacle distribution;

[0010] S200, performing multi-dimensional feature analysis on the device status data based on a preset deep feature extraction model to generate a device health assessment vector, where the generated device health assessment vector can reflect different states of the device; wherein the deep feature extraction model includes a dual-modal structure of a convolutional neural network and a long short-term memory network, which is used to extract spatial features and time series of the device status data;

[0011] S300, based on environmental monitoring data and equipment health assessment vectors, generates multiple candidate paths through a reinforcement learning algorithm, selects an initial inspection path from the multiple candidate paths, and dynamically optimizes the initial inspection path by integrating geographic information in real time;

[0012] S400: Based on the drone positioning data and device status feedback, if it is detected that the health level is below the threshold or the environmental risk exceeds the limit, local path replanning is triggered, and navigation is prioritized to high-risk devices.

[0013] By adopting the above-mentioned technical solution, a variety of equipment status data and environmental monitoring data are integrated to provide an efficient and intelligent solution for substation inspection planning; substation inspection tasks usually include complex equipment monitoring and environmental assessment requirements. The present invention can accurately judge the health status of the equipment to formulate the optimal inspection path; the deep feature extraction model is used to evaluate the health of the equipment, which can extract effective features from the various operating states of the equipment to form a comprehensive health assessment vector, which can not only quantify the operating status of the equipment, but also reflect the potential failure risks of the equipment, thereby providing important information for the optimization of the inspection path; at the same time, the present invention also enhances the dynamic adaptability of path planning. The drone can dynamically adjust the inspection path according to real-time environmental data and equipment health assessment results. This mechanism significantly improves the flexibility of the inspection task and the ability to cope with complex environmental changes, thereby realizing efficient and safe drone inspections in complex industrial environments such as substations.

[0014] Further configuration is that, in S100, the substation equipment status data and environmental monitoring data are obtained specifically as follows:

[0015] The distributed sensor network collects equipment temperature, vibration frequency, and insulation resistance values, and normalizes the equipment status data. The distributed sensor network is deployed at key monitoring nodes of substation equipment and includes infrared sensors, acceleration sensors, and insulation monitoring units.

[0016] Meteorological parameters, light intensity and obstacle distribution are obtained through the multispectral camera and lidar carried by the drone.

[0017] By adopting the above technical solutions, comprehensive monitoring of equipment and environment is achieved through distributed sensor networks and sensor systems on drones. Equipment status data is an important basis for judging equipment health, while environmental monitoring data provides detailed information about the external environment for path planning. The collection of data from multiple sources helps the model better identify and evaluate the health status of equipment. The deployment of the sensor network enables each key node in the substation to provide real-time feedback on the working status of equipment. Distributed monitoring improves the efficiency of inspections and avoids blind spots and errors that may be caused by manual inspections. At the same time, the real-time nature of data collection ensures a rapid response to emergencies. Sensors such as multispectral cameras and lidars on human-machine aircraft provide all-round perception capabilities for environmental monitoring, especially in the complex environment inside the substation. They can accurately obtain information such as obstacle distribution, light intensity and weather changes, thereby providing real-time and accurate environmental data support for path planning.

[0018] Further configuration is that the S200 is specifically:

[0019] S201. Input the device temperature, vibration frequency, and insulation resistance value into the parallel channels of the convolutional neural network, respectively, and extract local abnormal features using a multi-scale convolution kernel;

[0020] S202, inputting the time series of local abnormality features and historical device status data into a long short-term memory network;

[0021] S203: Concatenate the spatial features and the time series features into tensors and map them into a device health assessment vector through a fully connected layer.

[0022] By adopting the above technical solution, convolutional neural networks and long-short-term memory networks can fully exploit the spatial and temporal features in device status data. The convolutional neural network extracts local anomaly features using multi-scale convolution kernels, which is crucial for monitoring device abnormalities and effectively identifying minor faults or early signs of failure. The long-short-term memory network can conduct in-depth analysis of historical device status data and capture the dynamic changes in time series. This is particularly critical during device operation, as the device's health status is not only affected by the current environment but also closely related to its historical operation. The long-short-term memory network can effectively process data spanning long periods of time and accurately predict the device's future health status. The spatial and temporal features are concatenated into tensors to generate a health assessment vector, which not only comprehensively assesses the device's operating status but also reflects the device's risk level. This comprehensive assessment method provides a scientific basis for subsequent path planning and can issue timely warnings when device status anomalies occur, ensuring the efficient execution of inspection tasks.

[0023] Further configuration is that the S201 is specifically as follows:

[0024] Standardize equipment temperature data, vibration frequency signals, and insulation resistance values ​​to eliminate dimensional differences.

[0025] The standardized device temperature data, vibration frequency signal, and insulation resistance value are input into three independent input channels of the convolutional neural network, each corresponding to a data type, to achieve multimodal data parallel processing to generate local anomaly feature tensors.

[0026] The S202 is specifically as follows: arranging the local abnormal features into a time series in chronological order, and aligning the timestamps with the historical device status data;

[0027] Add time coding parameters to historical data;

[0028] Construct a long short-term memory network, whose input layer receives the concatenated local abnormal features and time encoding parameters, and dynamically adjusts the information flow through the gating mechanism of input gate, forget gate and output gate;

[0029] The S203 is specifically as follows: tensor splicing the spatial features output by the convolutional neural network and the temporal features output by the long short-term memory network, constructing a fully connected network, and constraining the output value to be within the range of 0.0-1.0 to generate a device health assessment vector:

[0030] If the health level is greater than the set first health level threshold, it indicates a normal state;

[0031] If the health level is between the first health level threshold and the second health level threshold, it indicates that there is a potential risk;

[0032] If the health level is lower than the second health level threshold, it indicates a high-risk state and triggers an alarm.

[0033] By adopting the above technical solution, the device data is processed through standardization, eliminating the dimensional differences between different sensor data and ensuring that the data can be effectively integrated into the same model. Standardization is a key step in data preprocessing. It ensures the balance of various types of data when input into the deep learning model and avoids the interference of data of different scales on the model learning process. The parallel processing of multimodal data is achieved through three independent input channels of the convolutional neural network, each channel corresponding to a data type. This parallel processing method can efficiently extract the characteristics of device data, especially when facing complex device states. It can fully explore the potential information in various types of sensor data and help improve the accuracy of fault prediction. The spatial features output by the convolutional neural network and the temporal features output by the long short-term memory network are tensor-spliced ​​to generate a fully connected network, which not only improves the accuracy of the evaluation but also ensures the stability of the equipment health assessment. By constraining the output value to the range of 0.0-1.0, the different states of equipment health can be clearly defined, providing clear guidance for subsequent inspection decisions.

[0034] Further configuration is that the S300 is specifically:

[0035] S301. Define the three-dimensional perception space of the substation area, including equipment coordinates, obstacle distribution, and environmental risks; divide the substation area into grid cells, and mark each grid cell as one of the following states:

[0036] A passable area means there are no obstacles within it;

[0037] Low-risk obstacles, which represent stationary obstacles;

[0038] High-risk obstacles, indicating dynamic obstacles or high-risk areas in the environment;

[0039] S302, designing a multi-objective reward function;

[0040] S303. Optimize the initial path through a path smoothing algorithm to ensure that the UAV flight trajectory is continuous and complies with kinematic constraints.

[0041] By adopting the above technical solution, by defining the three-dimensional perception space of the substation area and dividing it into grid units, a comprehensive perception of the substation environment can be achieved. This area division can clearly mark the passable area, low-risk obstacle area and high-risk obstacle area, which helps the drone avoid obstacles during the inspection process and ensure flight safety; at the same time, the management after zoning can improve the flexibility of path planning, allowing the drone to plan the inspection path more efficiently; the design of the multi-objective reward function takes into account multiple factors such as efficiency, safety and emergency tasks, so that the path planning process not only focuses on the shortest distance of the path, but also reasonably balances task priority and flight safety. This multi-faceted target optimization ensures that the drone can complete the inspection task efficiently and safely; by optimizing the initial path through the path smoothing algorithm, the discontinuity in the path can be eliminated, the drone can avoid sharp turns and vibrations during flight, and improve the stability of flight and the accuracy of task execution.

[0042] It is further configured that the multi-objective reward function in S302 includes an efficiency reward function, a safety reward function and an emergency task reward function;

[0043] For the efficiency reward function, the value R of the efficiency reward function is calculated based on the path time and the preset threshold. time ;

[0044] For the safety reward function, each time a high-risk obstacle is successfully avoided, the value of the safety reward function R is calculated based on the actual obstacle avoidance distance and the safety distance. safe ;

[0045] For the emergency task reward function, each time a device with a health value less than the second health threshold in a high-risk state is inspected, the value R of the emergency task reward function is calculated based on the health value of the device in the high-risk state. emergency ;

[0046] Assign weights to the efficiency reward function, safety reward function, and emergency task reward function respectively, and calculate the value R of the multi-objective reward function total .

[0047] By adopting the above technical solution, a multi-objective reward function can simultaneously optimize multiple aspects of path planning, including path efficiency, safety, and emergency response priority. Through this multi-objective optimization, the drone can not only quickly complete routine inspection tasks, but also prioritize high-risk equipment when equipment health is poor, thereby improving the emergency response capability of the inspection. The safety reward function can effectively guide the drone to avoid high-risk obstacles, ensuring safety during flight. The efficiency reward function ensures the optimization of the path by calculating the path time and preset thresholds, so that the inspection task can be completed in the shortest time. The emergency task reward function ensures timely response in the event of an emergency by prioritizing high-risk equipment. By assigning weights to each reward function, the priority of different objectives can be flexibly adjusted according to the actual situation, allowing the drone to make the most appropriate choice in different mission scenarios.

[0048] Further configuration is that the S303 is specifically as follows:

[0049] S3031. Generate multiple candidate paths using the Monte Carlo tree search algorithm and select paths that meet the following conditions:

[0050] The path length is less than 1.2 times the theoretical shortest path;

[0051] Covering more than 90% of high-risk equipment;

[0052] S3032. Calculate the value R of the multi-objective reward function for each selected candidate path. total ,select the candidate path with the highest score as the initial inspection path;

[0053] S3033. Perform cubic Bezier curve interpolation on the waypoint sequence of the initial inspection path to generate a continuous trajectory and constrain the path curvature to prevent the drone from making sharp turns.

[0054] S3034, based on the value R of the path multi-objective reward function total Dynamically adjust the drone's flight speed:

[0055] If the value of the multi-objective reward function R total If the value is greater than the set first function threshold, the flight speed of the drone is increased to the first speed;

[0056] If the value of the multi-objective reward function R total When the drone is between the first function threshold and the second function threshold, the drone maintains the default flight speed;

[0057] If the value of the multi-objective reward function R total When it is less than the second function threshold, the UAV flight speed is reduced to the second speed;

[0058] S3035. If there is a device in the substation whose health is less than a third health threshold, a hovering waypoint is added to the path and an audible and visual alarm is triggered.

[0059] By adopting the above technical solution, multiple candidate paths are generated through the Monte Carlo tree search algorithm, and the optimal path is screened through a multi-objective reward function. This not only avoids the uncertainty of path selection, but also enables dynamic optimization based on multiple factors such as equipment health status and environmental risks, ensuring the comprehensiveness and effectiveness of the inspection path; Bezier curve interpolation can generate a smooth path, avoiding the problem of sharp turns caused by excessive path curvature, which is crucial for the flight stability of the drone and the accuracy of mission execution, especially in complex environments, to ensure that the drone can successfully complete the mission; the strategy of dynamically adjusting the flight speed enables the drone to automatically adjust the flight speed according to the value of the multi-objective reward function, thereby ensuring the execution efficiency and safety of different inspection tasks. By flexibly adjusting the flight speed, the drone can maintain the optimal flight state under different environments and mission requirements.

[0060] Further configuration is that the S400 is specifically:

[0061] S401: If the health of a device in the substation decreases and is less than the second health threshold, increase the weight of its emergency task reward function in the multi-objective reward function; if the health of the device further decreases and is less than the third health threshold, force the drone to prioritize inspection and update the path score;

[0062] S402: If a sudden increase in wind speed is detected, the weight of the safety reward function is reduced in real time, and the path is recalculated.

[0063] By adopting the above technical solution and dynamically adjusting the weight of the emergency task reward function, it is possible to ensure that drones prioritize inspecting high-risk equipment when the equipment health deteriorates. This mechanism can effectively respond to sudden equipment failures or potential risks and prevent the impact of faulty equipment on the entire system from expanding. By increasing the weight of emergency tasks, drones can adjust their inspection paths in a timely manner, so that when the equipment health further deteriorates, they can quickly adjust the priority and optimize the path to ensure that critical equipment is inspected in a timely manner.

[0064] A further configuration is that after the above-mentioned S400, the following steps are further included:

[0065] S500: Collect path tracking errors, device false detection rates, and obstacle avoidance failure events during actual inspections, and generate a report.

[0066] By adopting the above technical solution, the generated report provides inspectors with detailed inspection records, helping to identify potential problems and provide improvement measures. This function not only improves inspection efficiency but also ensures that the health status of equipment is fully assessed, thereby improving the management level and safety of substation equipment.

[0067] In summary, the present invention has the following beneficial effects: the innovations of the present invention in equipment health assessment, path planning optimization, and real-time emergency response help to improve the efficiency and safety of inspection tasks, reduce the cost of manual intervention, and provide strong technical support for the efficient operation and maintenance of substations. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 Schematic diagram of the process of the embodiment. DETAILED DESCRIPTION

[0069] The present invention will be further described in detail below with reference to the accompanying drawings.

[0070] As attached Figure 1 As shown;

[0071] This embodiment discloses a substation route inspection planning method based on deep learning, which includes the following steps:

[0072] S100, obtaining substation equipment status data and environmental monitoring data; wherein the equipment status data includes equipment temperature, vibration frequency, and insulation resistance value, and the environmental monitoring data includes meteorological parameters, light intensity, and obstacle distribution;

[0073] S200, performing multi-dimensional feature analysis on the device status data based on a preset deep feature extraction model to generate a device health assessment vector, where the generated device health assessment vector can reflect different states of the device; wherein the deep feature extraction model includes a dual-modal structure of a convolutional neural network and a long short-term memory network, which is used to extract spatial features and time series of the device status data;

[0074] S300, based on environmental monitoring data and equipment health assessment vectors, generates multiple candidate paths through a reinforcement learning algorithm, selects an initial inspection path from the multiple candidate paths, and dynamically optimizes the initial inspection path by integrating geographic information in real time;

[0075] S400: Based on the drone positioning data and device status feedback, if it is detected that the health level is below the threshold or the environmental risk exceeds the limit, local path replanning is triggered, and navigation is prioritized to high-risk devices.

[0076] In S100, the substation equipment status data and environmental monitoring data are obtained as follows:

[0077] The distributed sensor network collects equipment temperature, vibration frequency, and insulation resistance values, and normalizes the equipment status data. The distributed sensor network is deployed at key monitoring nodes of substation equipment and includes infrared sensors, acceleration sensors, and insulation monitoring units.

[0078] Meteorological parameters, light intensity, and obstacle distribution are acquired through the multispectral camera and lidar carried by the drone, and the environmental monitoring data is spatially gridded. The granularity of the spatial gridding is adaptively adjusted according to the area of ​​the substation and the obstacle density.

[0079] Among them, S200 is specifically:

[0080] S201. Input the device temperature, vibration frequency, and insulation resistance value into the parallel channels of the convolutional neural network, respectively, and extract local abnormal features using a multi-scale convolution kernel;

[0081] S202, inputting the time series of local abnormality features and historical device status data into a long short-term memory network;

[0082] S203. Perform tensor splicing on the spatial features and the time series features, and map them into an equipment health assessment vector through a fully connected layer; wherein the dimension of the equipment health assessment vector corresponds one-to-one to the number of substation equipment types.

[0083] Among them, S201 is specifically:

[0084] The device temperature data, vibration frequency signal, and insulation resistance values ​​were standardized to eliminate dimensional differences. Specifically, the device temperature data was standardized and converted into a 128×128 pixel grayscale heat map, where the grayscale value of each pixel corresponds to the high and low device surface temperature. The vibration frequency signal was short-time Fourier transformed in a 0.1 second time window to generate a time-frequency map to characterize the change in vibration intensity of different frequencies over time. The insulation resistance value was generated using a sliding window to generate a two-dimensional time series matrix with a window length of 10 seconds and a step size of 5 seconds.

[0085] The standardized device temperature data, vibration frequency signal, and insulation resistance value are input into three independent input channels of the convolutional neural network. Each channel corresponds to a data type, realizing multimodal data parallel processing to generate local anomaly feature tensors. Specifically, the grayscale heat map, time-frequency map, and time series matrix are input into three independent channels of the convolutional neural network. Each channel uses 3×3, 5×5, and 7×7 multi-scale convolution kernels for feature extraction, where:

[0086] The 3×3 convolution kernel is used to detect localized minor anomalies on the device surface, including areas of sudden temperature rise or transient spikes in the vibration spectrum.

[0087] A 5×5 convolution kernel is used to identify correlation anomalies between equipment components, including abnormal temperature gradients or spatial diffusion patterns of vibration fluctuations;

[0088] The 7×7 convolution kernel is used to analyze overall trends, including the uniformity of temperature distribution across the substation or the periodicity of vibration;

[0089] Finally, the multi-scale convolution results are fused through residual connections, and the channel attention mechanism is used to weight important features to generate a local anomaly feature tensor.

[0090] S202 specifically includes: arranging the local abnormal features into a time series in chronological order and aligning the timestamps with the historical device status data; the historical device status data includes the device temperature, vibration frequency, and insulation resistance value records for the past 24 hours;

[0091] Add time coding parameters to historical data, including seasonal factors, day and night signs, and load cycle signs;

[0092] Construct a long short-term memory network, whose input layer receives the concatenated local abnormal features and time encoding parameters, and dynamically adjusts the information flow through the gating mechanism of input gate, forget gate and output gate;

[0093] The input gate assigns weights based on the current feature importance. If the insulation resistance value is detected to drop by more than 5% within the same load cycle, the weight of the feature at that time step is increased.

[0094] The forget gate decides to retain or discard information based on the relevance of historical data. If a temperature fluctuation is determined to be sporadic noise, its long-term impact is reduced.

[0095] The hidden state is output to characterize the degradation trend of the device and the remaining life is predicted based on the Weibull distribution model. The calculation formula is:

[0096] ;

[0097] Where η is the characteristic life parameter; β is the shape parameter, which is dynamically adjusted through the latent variable output by the long short-term memory network; t is time, and R(t) represents the reliability of the device after time t, that is, the probability that the device has not failed from the initial state to time t.

[0098] S203 specifically involves: compressing the spatial features output by the convolutional neural network into a 512-dimensional vector through global maximum pooling, compressing the temporal features output by the long short-term memory network into a 256-dimensional vector, performing tensor splicing on the spatial features output by the convolutional neural network and the temporal features output by the long short-term memory network to generate a 768-dimensional mixed feature vector, and reducing the dimensionality to 256 dimensions through principal component analysis to eliminate redundancy; constructing a fully connected network with three hidden layers, using Leaky ReLU as the activation function, and strictly matching the output layer dimension with the number of substation equipment types. The output value is constrained to the range of 0.0-1.0 through the Sigmoid function to generate an equipment health assessment vector:

[0099] If the health level is greater than the set first health level threshold, it indicates a normal state;

[0100] If the health level is between the first health level threshold and the second health level threshold, it indicates that there is a potential risk;

[0101] If the health level is lower than the second health level threshold, it indicates a high-risk state and triggers an alarm.

[0102] More specifically, in this embodiment, the first health threshold is set to 0.8, and the second health threshold is set to 0.5;

[0103] If the health level is greater than 0.8, it indicates a normal state and is marked green;

[0104] If 0.5≤Health≤0.8, it indicates potential risk and is marked in yellow;

[0105] If the health level is less than 0.5, it indicates a high-risk state and triggers an alarm, and is marked in red.

[0106] Dynamic thresholds are set based on the device type. The temperature threshold for outdoor devices is 5% higher than that for indoor devices, and the vibration tolerance for high-frequency use devices is reduced by 10%.

[0107] Among them, S300 is specifically:

[0108] S301. Define the three-dimensional perception space of the substation area, including equipment coordinates, obstacle distribution, and environmental risks. Divide the substation area into 0.5 m x 0.5 m grid cells, and mark each grid cell as one of the following states:

[0109] A passable area means there are no obstacles within it;

[0110] Low-risk obstacles represent static obstacles;

[0111] High-risk obstacles refer to dynamic obstacles or high-risk environmental areas. Dynamic obstacles include flying birds, and high-risk environmental areas include wind speeds greater than 10m / s.

[0112] S302, designing a multi-objective reward function;

[0113] S303. Optimize the initial path through a path smoothing algorithm to ensure that the UAV flight trajectory is continuous and complies with kinematic constraints.

[0114] The multi-objective reward function in S302 includes an efficiency reward function, a safety reward function, and an emergency task reward function;

[0115] For the efficiency reward function, based on the path time and the preset threshold T max To calculate the value of the efficiency reward function R time ; If the path time T is less than the preset threshold T max , the value of the efficiency reward function R time =1−T / T max ;

[0116] For the safety reward function, each time a high-risk obstacle is successfully avoided, the value of the safety reward function R is calculated based on the actual obstacle avoidance distance and the safety distance. safe ; Each time a high-risk obstacle is successfully bypassed, the value of the safety reward function R safe =(D safe −d) / D safe , where d is the actual obstacle avoidance distance, D safe For safe distance, it is set at 2 meters;

[0117] For the emergency task reward function, each time a device with a health value less than the second health threshold in a high-risk state is inspected, the value R of the emergency task reward function is calculated based on the health value of the device in the high-risk state. emergency ; Each time a device with a health level lower than the second health threshold is inspected and in a high-risk state, the value of the emergency task reward function R emergency =2×(1−health);

[0118] Assign weights to the efficiency reward function, safety reward function, and emergency task reward function respectively, and calculate the value R of the multi-objective reward function total .

[0119] The value R of the multi-objective reward function total =a1R time +a2R safe +a3R emergency , where a1, a2, and a3 represent the weights of the efficiency reward function, the safety reward function, and the emergency task reward function in the multi-objective reward function respectively; the initial value of the multi-objective reward function R total =0.5R time +0.3R safe+0.2R emergency .

[0120] Among them, S303 is specifically:

[0121] S3031. Generate multiple candidate paths using the Monte Carlo tree search algorithm and select paths that meet the following conditions:

[0122] The path length is less than 1.2 times the theoretical shortest path;

[0123] Covering more than 90% of high-risk equipment;

[0124] S3032. Calculate the value R of the multi-objective reward function for each selected candidate path. total ,select the candidate path with the highest score as the initial inspection path;

[0125] S3033. Perform cubic Bezier curve interpolation on the waypoint sequence of the initial inspection path to generate a continuous trajectory, and constrain the path curvature to a maximum of 0.1 per meter to prevent the drone from making sharp turns.

[0126] S3034, based on the value R of the path multi-objective reward function total Dynamically adjust the drone's flight speed:

[0127] If the value of the multi-objective reward function R total If the value is greater than the set first function threshold, the flight speed of the drone is increased to the first speed;

[0128] If the value of the multi-objective reward function R total When the drone is between the first function threshold and the second function threshold, the drone maintains the default flight speed;

[0129] If the value of the multi-objective reward function R total When it is less than the second function threshold, the UAV flight speed is reduced to the second speed;

[0130] More specifically, in this embodiment, the first function threshold is set to 0.8, the second function threshold is set to 0.5, the first speed is set to 3 m / s, the default flight speed is set to 2 m / s, and the second speed is set to 1 m / s;

[0131] If the value of the multi-objective reward function R total >0.8, the drone's flight speed is increased to 3m / s;

[0132] If 0.5≤the value of the multi-objective reward function R total ≤0.8, the drone maintains the default flight speed of 2m / s;

[0133] If the value of the multi-objective reward function R totalIf the value is less than 0.5, the flight speed of the drone will be reduced to 1m / s.

[0134] S3035. If there is equipment in the substation whose health is less than the third health threshold, a hovering waypoint is added to the path and an audible and visual alarm is triggered. The drone will stay for 10 seconds and take photos for evidence. The third health threshold is set to 0.3.

[0135] Among them, S400 is specifically:

[0136] S401: If the health of a device in the substation decreases and falls below the second health threshold, increase the weight of its emergency task reward function in the multi-objective reward function to 0.3; if the health of the device further decreases and falls below the third health threshold, force the drone to prioritize inspection and update the path score;

[0137] S402: If a sudden increase in wind speed is detected, the safety reward function weight is reduced to 0.2 in real time, and the path is recalculated. At the same time, if the light intensity is detected to be lower than the set light intensity threshold, a searchlight command is automatically added to the drone and the flight altitude is increased to 3 meters.

[0138] Among them, after S400, the following steps are also included:

[0139] S500: Collect path tracking errors, device false detection rates, and obstacle avoidance failure events during actual inspections, and generate a report.

[0140] Example 1

[0141] The goal was to verify efficient inspections under routine conditions. A routine inspection was conducted at a 500kV substation. Equipment health was generally good, and the environmental conditions were stable, with a wind speed of 3m / s and a light intensity of 800lux. The goal was to complete the inspection of 120 devices within 30 minutes.

[0142] In S100, the device status data is as follows:

[0143] Equipment temperature: The maximum surface temperature of transformer A is 52°C, which is below the threshold of 55°C; the temperature of insulator B is 48°C;

[0144] Vibration frequency: The vibration amplitude of circuit breaker C is 0.2mm, which is within the normal range of 0.1-0.3mm;

[0145] Insulation resistance value: The resistance value of arrester D is 500MΩ, which is within the normal range;

[0146] Environmental monitoring data are as follows:

[0147] Utility poles and fences are marked as low-risk obstacles;

[0148] Meteorological parameters: wind speed 3m / s, no precipitation.

[0149] In S200, the convolutional neural network detected a local temperature hotspot in transformer A, but the overall distribution was uniform. The long-short-term memory network analyzed historical equipment status data and predicted that its remaining life would be greater than 5 years.

[0150] The device health evaluation vector is as follows:

[0151] The health of transformer A is 0.85, the health of insulator B is 0.78, the health of circuit breaker C is 0.82, and the health of the lightning arrester is 0.93, all indicating normal status and marked green.

[0152] In S300, a multi-objective reward function R is selected by calculation. total The candidate path with the highest value is taken as the initial inspection path; the multi-objective reward function R of the initial inspection path total The value of is calculated as follows:

[0153] Theoretical shortest path L min =1.2 km, path length L=1.3 km;

[0154] The value of the efficiency reward function R time =1−28 / 30≈0.07;

[0155] Without obstacle avoidance events, the value of the safety reward function is R safe =0;

[0156] Without high-risk equipment, the value of the emergency task reward function is R emergency =0;

[0157] The value R of the multi-objective reward function total =0.5×0.07+0.3×0+0.2×0=0.035;

[0158] The unmanned flight speed is set to 3m / s.

[0159] In S400, the drone flew along the planned path and completed the inspection in 29 minutes with a path deviation of less than 0.2 meters.

[0160] The results are analyzed as follows:

[0161] The actual time consumption is 51.7% shorter than the traditional average 60-minute manual inspection;

[0162] The consistency rate between equipment health assessment and manual review reached 98%.

[0163] Example 2

[0164] Verify dynamic response in complex environments. During one inspection, a sudden strong wind of 15m / s and interference from a flock of birds necessitated urgent route adjustments and priority inspection of high-risk equipment.

[0165] In S100, the device status data is as follows:

[0166] The insulation resistance of transformer E drops to 150MΩ;

[0167] Environmental monitoring data are as follows:

[0168] Obstacle update: 3 new flying birds, and wind speed increased to 15m / s;

[0169] Light intensity: reduced to 200 lux due to cloud cover.

[0170] In S200,

[0171] The convolutional neural network detected an abnormal local insulation resistance of transformer E, and the long short-term memory network predicted that its remaining life was less than 6 months.

[0172] The health of transformer E is 0.25, marked in red;

[0173] In S400, the weight of the emergency task reward function is increased to 0.3, and the weight of the safety reward function is reduced to 0.2;

[0174] The path is replanned; a Monte Carlo tree search is used to generate a new path around the bird, covering transformer E, and the path length is increased to 1.5 km; at the same time, the flight speed is reduced to 1 m / s, and a hovering instruction is added.

[0175] The results are analyzed as follows:

[0176] Successfully avoided all flying birds, with a minimum obstacle avoidance distance of 1.8 meters;

[0177] The fault of transformer E was confirmed in a timely manner and was later analyzed to be caused by damage to the internal insulation layer, avoiding a potential power outage.

[0178] It can be seen from the above-mentioned Examples 1 and 2 that the present invention exhibits significant advantages in both conventional and complex scenarios. This method provides reliable technical support for intelligent inspection of substations and has broad application prospects.

[0179] This specific embodiment is merely an explanation of the present invention and is not intended to limit the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed. However, as long as such modifications are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A substation route inspection planning method based on deep learning, characterized in that: The following steps are included: S100, obtaining substation equipment status data and environmental monitoring data; wherein the equipment status data includes equipment temperature, vibration frequency, and insulation resistance value, and the environmental monitoring data includes meteorological parameters, light intensity, and obstacle distribution; S200, performing multi-dimensional feature analysis on the device status data based on a preset deep feature extraction model to generate a device health assessment vector, where the generated device health assessment vector can reflect different states of the device; wherein the deep feature extraction model includes a dual-modal structure of a convolutional neural network and a long short-term memory network, which is used to extract spatial features and time series features of the device status data; This includes: tensor splicing of the spatial features output by the convolutional neural network and the temporal features output by the long short-term memory network, and reducing the dimension through principal component analysis to eliminate redundancy; S300, based on environmental monitoring data and equipment health assessment vectors, generates multiple candidate paths through a reinforcement learning algorithm, screens the candidate paths to obtain an initial inspection path, and dynamically optimizes the initial inspection path by integrating geographic information in real time. S300 includes a multi-objective reward function and a path smoothing algorithm. The multi-objective reward function includes an efficiency reward function, a safety reward function and an emergency task reward function; The path smoothing algorithm includes: Generate multiple candidate paths through the Monte Carlo tree search algorithm; Based on the value of the multi-objective reward function, the initial inspection path is selected; Perform cubic Bezier curve interpolation on the waypoint sequence of the initial inspection path and constrain the path curvature; Dynamically adjust the UAV flight speed according to the value of the path multi-objective reward function; S400: Based on the drone positioning data and device status feedback, if the health level is detected to be below the threshold or the environmental risk exceeds the limit, local path replanning is triggered, and navigation is prioritized to high-risk devices, including: S401: If the health of a device in the substation decreases and is less than the second health threshold, increase the weight of its emergency task reward function in the multi-objective reward function; if the health of the device further decreases and is less than the third health threshold, force the drone to prioritize inspection and update the path score; S402: If a sudden increase in wind speed is detected, the weight of the safety reward function is reduced in real time, and the path is recalculated.

2. A substation route inspection planning method based on deep learning according to claim 1, characterized in that: In S100, the substation equipment status data and environmental monitoring data are obtained as follows: The distributed sensor network collects equipment temperature, vibration frequency, and insulation resistance values, and normalizes the equipment status data. The distributed sensor network is deployed at key monitoring nodes of substation equipment and includes infrared sensors, acceleration sensors, and insulation monitoring units. Meteorological parameters, light intensity and obstacle distribution are obtained through the multispectral camera and lidar carried by the drone.

3. The substation route inspection planning method based on deep learning according to claim 2 is characterized by: The S200 is specifically: S201. Input the device temperature, vibration frequency, and insulation resistance value into the parallel channels of the convolutional neural network, respectively, and extract local abnormal features using a multi-scale convolution kernel; S202, inputting the time series of local abnormality features and historical device status data into a long short-term memory network; S203: Concatenate the spatial features and the time series features into tensors and map them into a device health assessment vector through a fully connected layer.

4. A substation route inspection planning method based on deep learning according to claim 3, characterized in that: The S201 is specifically as follows: Standardize equipment temperature data, vibration frequency signals, and insulation resistance values ​​to eliminate dimensional differences. The standardized device temperature data, vibration frequency signal, and insulation resistance value are input into three independent input channels of the convolutional neural network, each corresponding to a data type, to achieve multimodal data parallel processing to generate local anomaly feature tensors. The S202 is specifically as follows: arranging the local abnormal features into a time series in chronological order, and aligning the timestamps with the historical device status data; Add time coding parameters to historical data; Construct a long short-term memory network, whose input layer receives the concatenated local abnormal features and time encoding parameters, and dynamically adjusts the information flow through the gating mechanism of input gate, forget gate and output gate; The S203 is specifically as follows: tensor splicing the spatial features output by the convolutional neural network and the temporal features output by the long short-term memory network, constructing a fully connected network, and constraining the output value to be within the range of 0.0-1.0 to generate a device health assessment vector: If the health level is greater than the set first health level threshold, it indicates a normal state; If the health level is between the first health level threshold and the second health level threshold, it indicates that there is a potential risk; If the health level is lower than the second health level threshold, it indicates a high-risk state and triggers an alarm.

5. A substation route inspection planning method based on deep learning according to claim 4, characterized in that: The S300 is specifically: S301. Define the three-dimensional perception space of the substation area, including equipment coordinates, obstacle distribution, and environmental risks; divide the substation area into grid cells, and mark each grid cell as one of the following states: A passable area means there are no obstacles within it; Low-risk obstacles, which represent stationary obstacles; High-risk obstacles, indicating dynamic obstacles or high-risk areas in the environment; S302, designing a multi-objective reward function; S303. Optimize the initial path through a path smoothing algorithm to ensure that the UAV flight trajectory is continuous and complies with kinematic constraints.

6. The substation route inspection planning method based on deep learning according to claim 5 is characterized by: For the efficiency reward function, the value R of the efficiency reward function is calculated based on the path time and the preset threshold. time ; For the safety reward function, each time a high-risk obstacle is successfully avoided, the value of the safety reward function R is calculated based on the actual obstacle avoidance distance and the safety distance. safe ; For the emergency task reward function, each time a device with a health value less than the second health threshold in a high-risk state is inspected, the value R of the emergency task reward function is calculated based on the health value of the device in the high-risk state. emergency ; Assign weights to the efficiency reward function, safety reward function, and emergency task reward function respectively, and calculate the value R of the multi-objective reward function total .

7. A substation route inspection planning method based on deep learning according to claim 6, characterized in that: The S303 is specifically as follows: S3031. Generate multiple candidate paths using the Monte Carlo tree search algorithm and select paths that meet the following conditions: The path length is less than 1.2 times the theoretical shortest path; Covering more than 90% of high-risk equipment; S3032. Calculate the value R of the multi-objective reward function for each selected candidate path. total ,select the candidate path with the highest score as the initial inspection path; S3033. Perform cubic Bezier curve interpolation on the waypoint sequence of the initial inspection path to generate a continuous trajectory and constrain the path curvature to prevent the drone from making sharp turns. S3034, based on the value R of the path multi-objective reward function total Dynamically adjust the drone's flight speed: If the value of the multi-objective reward function R total If the value is greater than the set first function threshold, the flight speed of the drone is increased to the first speed; If the value of the multi-objective reward function R total When the drone is between the first function threshold and the second function threshold, the drone maintains the default flight speed; If the value of the multi-objective reward function R total When it is less than the second function threshold, the UAV flight speed is reduced to the second speed; S3035. If there is a device in the substation whose health is less than a third health threshold, a hovering waypoint is added to the path and an audible and visual alarm is triggered.

8. The substation route inspection planning method based on deep learning according to claim 1 is characterized in that: After the step S400, the following steps are also included: S500: Collect path tracking errors, device false detection rates, and obstacle avoidance failure events during actual inspections, and generate a report.

Images