A multispectral remote sensing inspection and fault diagnosis method for power facilities

Abstract

The present invention relates to the technical field of fault detection of electric power facilities, and discloses a multispectral remote sensing inspection and fault diagnosis method for electric power facilities. The method obtains multispectral image data of electric power facilities through a multispectral remote sensing device carried on an unmanned aerial vehicle platform, and performs preprocessing to extract spectral features. A spectral feature analysis model is constructed to evaluate the health status of electric power facilities, and a fault diagnosis model is constructed in combination with historical fault data to determine the type and location of the fault. In order to optimize the fault diagnosis process, the present invention models fault diagnosis as a Markov decision process, defines the state space, action space, reward function and output strategy, and uses a reinforcement learning algorithm to continuously optimize decision-making actions, thereby improving the accuracy and efficiency of fault diagnosis. The present invention realizes the intelligent and automated inspection and fault diagnosis of electric power facilities, provides strong technical support for the safe and stable operation of electric power facilities, and has broad application prospects and important practical value.

Description

Technical Field

[0001] The present invention relates to the field of power facility fault detection, and in particular to a multi-spectral remote sensing inspection and fault diagnosis method for power facilities. Background Art

[0002] With the rapid development of the power industry and the continuous expansion of the power grid, the safe and stable operation of power facilities has become particularly important. With the rapid advancement of remote sensing and drone technologies, the use of drones equipped with remote sensing equipment for power facility inspections has become a research hotspot. Traditional remote sensing inspection methods, mostly based on visible light imagery, struggle to effectively detect subtle defects or potential faults in power facilities. Multispectral remote sensing, as an advanced remote sensing technology, can acquire spectral information of targets across different wavelengths, providing richer and more detailed ground features. Therefore, it holds broad application prospects in power facility inspections.

[0003] Current multispectral remote sensing inspection technology still faces some challenges. The processing and analysis methods for multispectral image data are still imperfect. Effectively extracting the spectral characteristics of power facilities from massive amounts of data and accurately assessing their health status are pressing issues. While multispectral remote sensing technology can provide rich feature information, integrating these features with power facility fault diagnosis to achieve intelligent and automated fault diagnosis remains a challenge.

[0004] The fault diagnosis process for power facilities often involves multiple factors, including the facility's operating status, historical fault data, and maintenance resources. The interaction between these factors complicates fault diagnosis. Traditional fault diagnosis methods, often based on experience or simple rules, struggle to adapt to the diverse and complex nature of power facility faults. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-spectral remote sensing inspection and fault diagnosis method for electric power facilities to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multispectral remote sensing inspection and fault diagnosis method for power facilities, the method comprising:

[0007] Step 1: Use the multispectral remote sensing equipment on the UAV platform to inspect power facilities and obtain multispectral image data of the target area;

[0008] Step 2: Preprocess the acquired multispectral image data, including correction, denoising and segmentation, to extract the spectral characteristic data of the power facilities;

[0009] Step 3: Construct a spectral feature analysis model, the input of which is the preprocessed spectral feature data, and the output is the health status assessment value of the power facility;

[0010] Step 4: Construct a fault diagnosis model. This model uses the output of the spectral feature analysis model and combines it with the historical fault data of the power facilities to diagnose the faults of the power facilities and output the fault type and location.

[0011] Step 5: Construct a fault diagnosis decision model, modeling the fault diagnosis process of power facilities as a Markov decision process. Specific definitions include: ① State space: defined as the combination of the real-time spectral feature data and health status assessment values ​​of the power facility; ② Action space: defined as the set of executable inspection and maintenance operations; ③ Reward function: designed based on the stable operation and fault reduction goals of the power facility, including a reward mechanism for state stability, fault prevention, and maintenance efficiency; and combined with the health status assessment value output by the spectral feature analysis model, additional rewards are given for operations that improve health status, and additional penalties are given for operations that degrade health status; ④ Output strategy: defined as the method for selecting the optimal action under the current state, and the decision actions generated are continuously optimized through the reinforcement learning algorithm;

[0012] Step 6: Perform real-time fault diagnosis based on the trained model, collect spectral feature data of power facilities in real time as model input, and select the optimal decision action based on the current status and strategy.

[0013] Preferably, the executable inspection and maintenance operation set includes: adjusting inspection frequency, changing spectrum acquisition parameters, arranging on-site inspections, performing preventive maintenance, replacing aging components, and adjusting equipment operating parameters.

[0014] Preferably, the spectral feature analysis model is constructed using a convolutional neural network (CNN) algorithm, and its network structure specifically includes:

[0015] Input layer: used to receive preprocessed spectral feature data;

[0016] Hidden layer: contains multiple convolutional layers and pooling layers to extract high-level representations of spectral features;

[0017] Output layer: The high-level features are mapped to the health status assessment value of the power facility through the fully connected layer.

[0018] Preferably, the step of training the spectral feature analysis model includes:

[0019] F101: Initialize the weights and biases of the CNN network;

[0020] F102: Calculates the mean squared error loss as a loss function to measure the difference between the health assessment value predicted by the model and the true value;

[0021] F103: Use the backpropagation algorithm and gradient descent optimization algorithm to calculate the gradient according to the loss function, update the weights and bias terms of the CNN network, and iteratively train the network until the loss function converges;

[0022] F104: Save the trained CNN network model as a spectral feature analysis model in the multispectral remote sensing inspection and fault diagnosis method for power facilities.

[0023] Preferably, the fault diagnosis model is constructed using a support vector machine (SVM) algorithm, and its training steps include:

[0024] F201: Collect spectral data of power facilities and corresponding health status labels to construct a training dataset;

[0025] F202: Use the support vector machine (SVM) algorithm to build a fault diagnosis model. The model formula is:

[0026] in, is the input spectral feature vector, is the support vector, is the corresponding Lagrange multiplier, is the kernel function, is the bias term;

[0027] F203: Use the training data set to train the constructed SVM model and solve the model parameters through the sequential minimum optimization (SMO) algorithm;

[0028] F204: After the training is completed, the model is verified using the validation data set to evaluate the prediction performance of the model. If it does not meet the requirements, return to step F203 to adjust the model parameters until the model performance meets the requirements.

[0029] Preferably, the reward function R is composed of multiple sub-reward items, and the total reward value is obtained by weighted summation. The specific formula is: R = w1 * R_stab + w2 * R_prev + w3 * R_eff + w4 * R_model; wherein, R_stab represents the reward item for state stability, which is calculated according to the degree of fluctuation of the operating state of the power facility; R_prev represents the reward item for fault prevention, which is calculated according to the frequency and severity of faults; R_eff represents the reward item for maintenance efficiency, which is calculated according to the success rate and time consumption of maintenance operations; R_model represents the model evaluation reward item, which rewards or penalties according to the improvement or decrease of the health status assessment value output by the spectral feature analysis model; w1, w2, w3 and w4 are the weight coefficients of each sub-reward item, and w1+w2+w3+w4=1.

[0030] Preferably, a deep deterministic policy gradient (DDPG) algorithm is used to train the fault diagnosis decision model.

[0031] Preferably, the step of training the fault diagnosis decision model includes:

[0032] S1: Initialize the actor network and the critic network. The actor network is used to generate actions, and the critic network is used to evaluate the value of actions.

[0033] S2: Initialize the experience playback buffer D, which is used to store the state transition experience generated during the power facility inspection and fault diagnosis process;

[0034] S3: For each training round:

[0035] S301: Initialize the state s at the beginning of the round;

[0036] S302: Before the round ends, the following steps are executed in a loop: i) the actor network selects action a based on the current state s; ii) action a is executed, and the reward r, the next state s', and the termination flag done are observed; iii) the experience (s, a, r, s', done) is stored in the experience replay buffer D; iv) a batch of experience is randomly sampled from the experience replay buffer D to update the network parameters; v) the sampled experience is used to update the parameters of the actor network and the critic network; vi) the current state s is updated to the next state s';

[0037] S303: When the round ends, reset the state s to the new round start state;

[0038] S4: Repeat step S3 until a satisfactory performance level is achieved.

[0039] Preferably, step v in S302 is specifically: using the sampled experience to update the parameters of the critic network through a temporal difference learning method to minimize the value function prediction error.

[0040] Preferably, the multispectral image data includes spectral information of multiple bands of visible light, infrared light and ultraviolet light.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] Using drone platforms equipped with multispectral remote sensing equipment enables efficient and comprehensive inspections of power facilities, significantly reducing the time and labor required for manual inspections while overcoming limitations imposed by environmental and weather factors. Multispectral remote sensing technology can capture spectral information of power facilities across different wavelengths, providing richer and more detailed feature data. Compared to traditional visible light image inspections, it can more effectively detect subtle defects or potential faults in power facilities.

[0043] The constructed spectral feature analysis model and fault diagnosis model can accurately assess the health status of power facilities based on preprocessed spectral feature data, perform fault diagnosis in combination with historical fault data, and output specific fault types and locations, greatly improving the accuracy and efficiency of fault diagnosis.

[0044] Through Markov decision process modeling, the fault diagnosis process is transformed into a decision-making problem, and the reinforcement learning algorithm is used to continuously optimize the decision-making actions, making fault diagnosis more intelligent and automated, and adapting to the needs of the diversity and complexity of power facility faults.

[0045] The fault diagnosis decision model not only considers the real-time health of power facilities but also incorporates the availability of maintenance resources. Through a reward function and penalty mechanism, it optimizes maintenance action decisions, ensuring the rational allocation and efficient utilization of maintenance resources. It collects spectral signature data from power facilities in real time and selects the optimal action based on the current status and strategy. This enables dynamic monitoring and timely maintenance of power facilities, effectively preventing faults and reducing maintenance costs.

[0046] Through precise health assessment and fault diagnosis, this invention can promptly detect and address potential faults in power facilities, avoiding safety incidents such as power outages and equipment damage caused by these faults, thereby improving the operational safety and reliability of power facilities. Continuous optimization of reinforcement learning algorithms makes fault diagnosis and repair decisions more accurate and efficient, providing a strong guarantee for the long-term stable operation of power facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A diagram showing the processing steps of a multi-spectral remote sensing inspection and fault diagnosis method for electric power facilities according to the present invention; Figure 2 This is the training flow chart for the spectral feature analysis module; Figure 3 Flowchart for training a fault diagnosis decision model using the Deep Deterministic Policy Gradient (DDPG) algorithm. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] See also Figure 1-3 The present invention provides a technical solution: a multi-spectral remote sensing inspection and fault diagnosis method for power facilities, the method comprising:

[0050] Step 1: Data Acquisition: Use a drone platform equipped with multispectral remote sensing equipment to inspect power facilities. The drone flies along a pre-set flight path and altitude, ensuring that the multispectral remote sensing equipment fully covers the power facilities in the target area. During flight, the multispectral remote sensing equipment simultaneously captures spectral image data in multiple bands, including visible light, infrared, and ultraviolet. This data is stored in an onboard storage device and then transmitted and processed after the flight.

[0051] Step 2: Data Preprocessing: The acquired multispectral image data is preprocessed. This includes image correction to eliminate image distortion caused by factors such as lens distortion and uneven lighting; a denoising algorithm to reduce noise interference in the image; and image segmentation techniques to separate the power facilities from the background and extract the spectral signature data of the region of interest (ROI). The preprocessed spectral signature data is stored in a standard format to facilitate subsequent analysis and processing.

[0052] Step 3: Build a spectral feature analysis model: This model is constructed based on machine learning algorithms (such as support vector machines and random forests). The model input is preprocessed spectral feature data, and the output is a health assessment of the power facility. To train the model, a large amount of spectral feature data from power facilities with known health conditions is collected as training samples. Model parameters are optimized through methods such as cross-validation to ensure model accuracy and generalization.

[0053] Step 4: Build a fault diagnosis model: Combined with historical fault data from power facilities, a fault diagnosis model is constructed. This model diagnoses power facility faults based on the output of the spectral signature analysis model, along with information such as the facility's operating status and historical maintenance records. Classification algorithms (such as decision trees and neural networks) are used to identify the fault type, and localization algorithms are used to pinpoint the fault location. The fault diagnosis model also requires training samples for learning and optimization.

[0054] Step 5: Construct a fault diagnosis decision model and model the fault diagnosis process of power facilities as a Markov decision process. The specific definitions are as follows:

[0055] State space: It consists of real-time spectral characteristic data and health status assessment values ​​of power facilities, indicating the current status of power facilities.

[0056] Action space: includes a set of executable operations such as inspection and maintenance, representing possible measures that may be taken on power facilities.

[0057] Reward function: Designed based on the goals of stable operation and fault reduction for power facilities, it includes rewards for state stability, fault prevention, and maintenance efficiency. Furthermore, combined with the health status assessment output by the spectral feature analysis model, additional rewards are given for actions that improve health status, while additional penalties are imposed for actions that degrade health status.

[0058] Output strategy: A reinforcement learning algorithm is used to learn the optimal strategy, i.e., the method of selecting the best action in the current state. Through continuous iterative learning, the generated decision actions are optimized.

[0059] Step 6: Real-time Fault Diagnosis: Real-time fault diagnosis is performed based on the trained model. Spectral signature data from power facilities is collected in real time through drone inspections and serves as model input. The spectral signature analysis model first assesses the health of the power facilities. The fault diagnosis model then performs fault diagnosis based on the health assessment and other relevant information. Finally, the fault diagnosis decision model selects the optimal action based on the current status and strategy, such as scheduling maintenance or adjusting inspection frequency, to ensure the safe and stable operation of the power facilities.

[0060] The present invention will be further described below in conjunction with Examples 1 to 3:

[0061] Example 1:

[0062] This embodiment describes in detail an efficient and intelligent method for multispectral remote sensing inspection and fault diagnosis of power facilities. The specific implementation is as follows: A. Data Acquisition and Processing: Using high-precision multispectral remote sensing equipment aboard an unmanned aerial vehicle (UAV) platform, regular or on-demand inspections of power facilities are performed according to pre-set flight plans and parameter settings. The collected multispectral image data undergoes preprocessing steps such as image correction, denoising, and segmentation to extract clear and accurate spectral signature data of the power facilities.

[0063] B. A set of executable inspection and maintenance operations: Based on the actual operation and maintenance requirements of power facilities, a series of executable inspection and maintenance operations are defined, including:

[0064] Adjust inspection frequency: Dynamically adjust the inspection frequency based on historical fault data, operating environment, and current health assessment of power facilities to ensure more intensive monitoring of critical facilities.

[0065] Changing spectral acquisition parameters: Adjust spectral acquisition parameters (such as band selection and exposure time) to optimize data acquisition quality based on weather conditions, lighting changes, or the detection requirements of specific fault types.

[0066] Arrange on-site inspection: When the spectral signature analysis model or fault diagnosis model indicates a potential fault, arrange for professionals to conduct an on-site inspection to confirm the fault and take appropriate measures.

[0067] Perform preventive maintenance: Based on model predictions and historical data, preventive maintenance or replacement of parts that may fail can be performed to reduce the occurrence of sudden failures.

[0068] Replace aging components: Based on the service life and health status assessment of the components, replace aging key components in a timely manner to ensure the stable operation of power facilities.

[0069] Adjust equipment operating parameters: According to the actual operating conditions and needs of the power facilities, adjust the equipment operating parameters (such as voltage, current, etc.) to optimize equipment performance and extend its service life.

[0070] C. Constructing a spectral feature analysis model. The convolutional neural network (CNN) algorithm is used to construct a spectral feature analysis model. The specific design of its network structure is as follows:

[0071] Input layer: Receives preprocessed spectral feature data, which is represented in the form of a multidimensional array, where each dimension corresponds to a different spectral band and spatial position.

[0072] Hidden layers: These layers consist of multiple convolutional and pooling layers. Convolutional layers extract high-level representations of spectral features, such as edges, textures, and shapes, through convolution operations. Pooling layers reduce the dimensionality of feature data, minimizing computational effort while retaining important features. These layers are arranged alternately to form a deep network structure that fully exploits the information contained in spectral features.

[0073] Output layer: The high-level features extracted by the hidden layer are mapped to a health assessment value of the power facility through a fully connected layer. This assessment value can be a continuous value representing the overall health of the power facility, or a categorical label indicating whether the power facility is in a normal, warning, or fault state.

[0074] During the model training process, a large amount of labeled data (i.e., spectral characteristic data of power facilities with known health conditions) is used to train the CNN model, and the model parameters are optimized through the back-propagation algorithm, so that the model can accurately predict the health status of power facilities.

[0075] D. Constructing a Fault Diagnosis and Decision Model: Based on the spectral signature analysis model, a fault diagnosis model is constructed, combining historical fault data and O&M knowledge of power facilities. This model uses classification or regression algorithms to diagnose power facility faults and outputs the fault type and location. The power facility fault diagnosis process is modeled as a Markov decision process, defining the state space, action space, reward function, and output strategy. Using a reinforcement learning algorithm to continuously optimize the generated decision actions, the model selects the optimal inspection and maintenance actions based on the current power facility status and O&M requirements.

[0076] E. Real-time Fault Diagnosis and Decision-Making: Based on the trained model, spectral signature data from power facilities is collected in real time and input into the spectral signature analysis and fault diagnosis models. Based on the model output and the Markov decision process strategy, the optimal inspection and maintenance actions are selected. By continuously monitoring the status of power facilities and updating the model, real-time, accurate fault diagnosis and intelligent decision-making are achieved, ensuring the safe and stable operation of power facilities.

[0077] Example 2:

[0078] To implement the training of a spectral feature analysis model and a fault diagnosis model in a multispectral remote sensing inspection and fault diagnosis method for electric power facilities, this embodiment provides the following detailed steps:

[0079] The convolutional neural network (CNN) algorithm is used to build a spectral feature analysis model. The specific training steps are as follows:

[0080] F101. Initialize the weights and biases of the CNN network: Use a random initialization method (such as Gaussian distribution or uniform distribution) to assign initial values ​​to the weights and biases of each layer of the CNN network. These initial values ​​will serve as the starting point for network training.

[0081] F102. Calculate the loss function: Select Mean Squared Error (MSE) as the loss function to measure the difference between the health assessment value predicted by the model and the actual value. The loss function formula is: in is the real health status assessment value, is the health status assessment value predicted by the model, and N is the number of samples.

[0082] F103. Network training and optimization: Use backpropagation and gradient descent optimization algorithms (such as stochastic gradient descent (SGD), mini-batch SGD, or Adam) for network training. Calculate the gradient of the loss function, that is, the partial derivative of the loss function with respect to the network weights and biases. Use the gradient descent optimization algorithm to update the network weights and biases so that the loss function gradually decreases. Iterate the network until the loss function converges or the preset number of training rounds is reached.

[0083] F104. Model saving: After network training is completed, save the trained CNN network model, including parameters such as network structure, weights, and bias terms; use the saved model as a spectral feature analysis model in the multispectral remote sensing inspection and fault diagnosis method of power facilities for subsequent spectral feature extraction and health status assessment.

[0084] The fault diagnosis model is trained by using the support vector machine (SVM) algorithm. The specific training steps are as follows:

[0085] F201. Build a training dataset: Collect spectral data from power facilities and their corresponding health status labels (e.g., normal, warning, fault, etc.). Preprocess the spectral data to extract spectral feature vectors. Combine the spectral feature vectors and health status labels to form a training dataset.

[0086] F202, build SVM model: Use SVM algorithm to build fault diagnosis model, the model formula is:

[0087] Where x is the input spectral feature vector, is the support vector, is the corresponding Lagrange multiplier, is the kernel function (such as linear kernel, Gaussian kernel, etc.), and b is the bias term.

[0088] F203, Model training: Use the training data set to train the constructed SVM model. Use the Sequential Minimal Optimization (SMO) algorithm to solve the model parameters, that is, optimize the Lagrange multiplier And the bias term b. Through iterative optimization, the model achieves better classification performance on the training data set.

[0089] F204, Model Validation and Tuning: After training is complete, the model is validated using the validation dataset to evaluate its predictive performance (e.g., precision, recall, F1 score, etc.). If the model performance does not meet the requirements, return to step F203, adjust the model parameters (e.g., kernel function type, penalty parameter C, etc.), and retrain the model. Repeat the validation and tuning process until the model performance meets the requirements.

[0090] Example 3:

[0091] This example describes a method for building a fault diagnosis decision model and uses the Deep Deterministic Policy Gradient (DDPG) algorithm for training. The following are the detailed implementation steps:

[0092] Designing the reward function: The reward function R is a key part of the reinforcement learning algorithm. It is used to evaluate the contribution of the actions taken to the operation and maintenance of the power facilities. The reward function consists of multiple sub-reward items, and the total reward value is obtained by weighted summation. The specific formula is:

[0093] R = w1 * R_stab + w2 * R_prev + w3 * R_eff + w4 * R_model, where R_stab represents the reward for state stability, calculated based on the degree of fluctuation in the power facility's operating state. A positive reward is given for stable operation, while a negative reward is given for significant fluctuations.

[0094] R_prev represents the reward for fault prevention, which is calculated based on the frequency and severity of the fault. If the fault is effectively prevented, a positive reward is given; if the fault occurs frequently or is of high severity, a negative reward is given.

[0095] R_eff represents the reward for repair efficiency, which is calculated based on the success rate and time of the repair operation. If the repair operation is successful and takes a short time, a positive reward is given; if the repair operation fails or takes a long time, a negative reward is given.

[0096] R_model represents the model evaluation reward item, which rewards or penalizes based on the improvement or decrease in the health assessment value output by the spectral feature analysis model. An improvement in the health assessment value results in a positive reward, while a decrease in the health assessment value results in a negative reward. w1, w2, w3, and w4 are the weight coefficients for each sub-reward item, satisfying the following: w1 + w2 + w3 + w4 = 1. These weight coefficients can be adjusted based on actual conditions to reflect the importance of different sub-reward items in the total reward.

[0097] Training the fault diagnosis decision model: The deep deterministic policy gradient (DDPG) algorithm is used to train the fault diagnosis decision model. The specific steps are as follows:

[0098] S1. Network Initialization: Initialize the Actor Network and Critic Network. The Actor Network is used to generate actions, that is, to select the optimal inspection and maintenance operations based on the current state; the Critic Network is used to evaluate the value of actions, that is, to assess the contribution of the actions taken to the operation and maintenance of power facilities.

[0099] S2. Initialize the experience replay buffer: Initialize the experience replay buffer D, which is used to store the state transition experience generated during the power facility inspection and fault diagnosis process, including information such as state, action, reward, next state, and termination flag.

[0100] S3, training round loop, for each training round, perform the following steps:

[0101] S301 State initialization: Initialize the state s at the beginning of the round, which may include information such as spectrum characteristics, historical operating status, and maintenance records of the power facility.

[0102] S302 In-round step loop: Before the round ends, the following steps are executed in a loop:

[0103] i) Action selection: The actor network selects action a according to the current state s, i.e., it selects the optimal inspection and maintenance operation.

[0104] ii) Action Execution and Observation: Execute action a and observe the reward r, next state s′, and the termination flag done. The reward r is calculated based on the reward function, the next state s′ is updated based on the actual operating status of the power facility, and the termination flag done is used to determine whether the round has ended.

[0105] iii) Experience storage: The experience (s, a, r, s′, done) is stored in the experience replay buffer D for subsequent use in updating network parameters.

[0106] iv) Experience Sampling and Network Update: A batch of experiences is randomly sampled from the experience replay buffer D and used to update the parameters of the actor network and the critic network. An optimization algorithm such as gradient descent is used to calculate the loss function based on the sampled experiences and update the network parameters.

[0107] v) State update: Update the current state s to the next state s′ to select and execute the next action.

[0108] S303 Round termination processing: When the round ends (that is, done is true), reset the state s to the new round start state and prepare to start the next round of training.

[0109] S4. Training termination condition: Repeat step S3 until a satisfactory performance level is achieved. The performance level can be determined by evaluating the model's performance on the validation dataset, such as when metrics such as accuracy, recall, and F1 score reach preset thresholds, or when the loss function during training converges to a small value and no longer changes significantly.

[0110] Step v in S302 is specifically: using the sampled experience to update the parameters of the critic network through a temporal difference learning method to minimize the value function prediction error.

[0111] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0112] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-spectral remote sensing inspection and fault diagnosis method for power facilities, characterized in that: The method comprises: Step 1: Use the multispectral remote sensing equipment on the UAV platform to inspect power facilities and obtain multispectral image data of the target area; Step 2: Preprocess the acquired multispectral image data, including correction, denoising and segmentation, to extract the spectral characteristic data of the power facilities; Step 3: Construct a spectral feature analysis model, the input of which is the preprocessed spectral feature data, and the output is the health status assessment value of the power facility; Step 4: Construct a fault diagnosis model. This model uses the output of the spectral feature analysis model and combines it with the historical fault data of the power facilities to diagnose the faults of the power facilities and output the fault type and location. Step 5: Construct a fault diagnosis decision model, modeling the fault diagnosis process of power facilities as a Markov decision process. Specific definitions include: ① State space: defined as the combination of the real-time spectral feature data and health status assessment values ​​of the power facility; ② Action space: defined as the set of executable inspection and maintenance operations; ③ Reward function: designed based on the stable operation and fault reduction goals of the power facility, including a reward mechanism for state stability, fault prevention, and maintenance efficiency; and combined with the health status assessment value output by the spectral feature analysis model, additional rewards are given for operations that improve health status, and additional penalties are given for operations that degrade health status; ④ Output strategy: defined as the method for selecting the optimal action under the current state, and the decision actions generated are continuously optimized through the reinforcement learning algorithm; Step 6: Perform real-time fault diagnosis based on the trained model, collect spectral feature data of power facilities in real time as model input, and select the optimal decision action based on the current status and strategy.

2. A multi-spectral remote sensing inspection and fault diagnosis method for electric power facilities according to claim 1, characterized in that: The executable inspection and maintenance operation set includes: adjusting inspection frequency, changing spectrum acquisition parameters, arranging on-site inspections, performing preventive maintenance, replacing aging components, and adjusting equipment operating parameters.

3. The multi-spectral remote sensing inspection and fault diagnosis method for electric power facilities according to claim 1, characterized in that: The spectral feature analysis model is constructed using a convolutional neural network (CNN) algorithm, and its network structure specifically includes: Input layer: used to receive preprocessed spectral feature data; Hidden layer: contains multiple convolutional layers and pooling layers to extract high-level representations of spectral features; Output layer: The high-level features are mapped to the health status assessment value of the power facility through the fully connected layer.

4. A multi-spectral remote sensing inspection and fault diagnosis method for electric power facilities according to claim 3, characterized in that: The steps for training a spectral feature analysis model include: F101: Initialize the weights and biases of the CNN network; F102: Calculates the mean squared error loss as a loss function to measure the difference between the health assessment value predicted by the model and the true value; F103: Use the backpropagation algorithm and gradient descent optimization algorithm to calculate the gradient according to the loss function, update the weights and bias terms of the CNN network, and iteratively train the network until the loss function converges; F104: Save the trained CNN network model as a spectral feature analysis model in the multispectral remote sensing inspection and fault diagnosis method for power facilities.

5. The method for multi-spectral remote sensing inspection and fault diagnosis of electric power facilities according to claim 1, characterized in that: The fault diagnosis model is constructed using the support vector machine (SVM) algorithm, and its training steps include: F201: Collect spectral data of power facilities and corresponding health status labels to construct a training dataset; F202: Use the support vector machine (SVM) algorithm to build a fault diagnosis model. The model formula is: in, is the input spectral feature vector, is the support vector, is the corresponding Lagrange multiplier, is the kernel function, is the bias term; F203: Use the training data set to train the constructed SVM model and solve the model parameters through the sequential minimum optimization (SMO) algorithm; F204: After the training is completed, the model is verified using the validation data set to evaluate the prediction performance of the model. If it does not meet the requirements, return to step F203 to adjust the model parameters until the model performance meets the requirements.

6. The method for multi-spectral remote sensing inspection and fault diagnosis of electric power facilities according to claim 1, characterized in that: The reward function R consists of multiple sub-reward items, and the total reward value is obtained by weighted summation. The specific formula is: R = w1 * R_stab + w2 * R_prev + w3 * R_eff + w4 * R_model; where R_stab represents the reward item for state stability, which is calculated based on the degree of fluctuation in the operating state of the power facility; R_prev represents the reward item for fault prevention, which is calculated based on the frequency and severity of faults; R_eff represents the reward item for maintenance efficiency, which is calculated based on the success rate and time consumption of maintenance operations; R_model represents the model evaluation reward item, which rewards or punishes according to the improvement or decrease in the health status assessment value output by the spectral feature analysis model; w1, w2, w3 and w4 are the weight coefficients of each sub-reward item, and w1+w2+w3+w4=1.

7. A multi-spectral remote sensing inspection and fault diagnosis method for electric power facilities according to claim 6, characterized in that: The deep deterministic policy gradient (DDPG) algorithm is used to train the fault diagnosis decision model.

8. A multi-spectral remote sensing inspection and fault diagnosis method for electric power facilities according to claim 7, characterized in that: The steps for training a fault diagnosis decision model include: S1: Initialize the actor network and the critic network. The actor network is used to generate actions, and the critic network is used to evaluate the value of actions. S2: Initialize the experience playback buffer D, which is used to store the state transition experience generated during the power facility inspection and fault diagnosis process; S3: For each training round: S301: Initialize the state s at the beginning of the round; S302: Before the round ends, the following steps are executed in a loop: i) the actor network selects action a based on the current state s; ii) action a is executed, and the reward r, the next state s', and the termination flag done are observed; iii) the experience (s, a, r, s', done) is stored in the experience replay buffer D; iv) a batch of experience is randomly sampled from the experience replay buffer D to update the network parameters; v) the sampled experience is used to update the parameters of the actor network and the critic network; vi) the current state s is updated to the next state s'; S303: When the round ends, reset the state s to the new round start state; S4: Repeat step S3 until a satisfactory performance level is achieved.

9. A multi-spectral remote sensing inspection and fault diagnosis method for electric power facilities according to claim 8, characterized in that: Step v in S302 is specifically: using the sampled experience to update the parameters of the critic network through a temporal difference learning method to minimize the value function prediction error.

10. The method for multi-spectral remote sensing inspection and fault diagnosis of electric power facilities according to claim 1, characterized in that: The multispectral image data includes spectral information of multiple bands of visible light, infrared light and ultraviolet light.

Images