Method for monitoring and analyzing electrical equipment through linkage of camera and monitoring system

Abstract

The invention provides a method for monitoring and analyzing electrical equipment through linkage of a camera and a monitoring system, and the method comprises the steps: firstly, enabling a patrol system to define a key region needing to be monitored, and enabling the camera to carry out the continuous monitoring in a non-patrol task period; and then acquiring a video stream for analysis, extracting key features by using a computer vision technology, identifying the abnormality of the electrical equipment by means of a trained deep learning model, and immediately alarming and recording related information once the abnormality is found. And when the electrical equipment monitoring system perceives a potential problem, the alarm signal triggers the inspection host, and the camera is controlled to jump to a preset position to collect a field condition. And after the collected data is stored, the authenticity and severity of the abnormity are deeply evaluated by using the intelligent analysis algorithm again. And finally, an analysis result is fed back to an operator to assist in decision making, and the operation process and result are recorded in detail so as to facilitate review and performance optimization. According to the method, all-around, accurate and efficient monitoring analysis of the electrical equipment can be realized, the operation and maintenance level is improved, and safe and stable operation of a power system is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of monitoring electrical equipment, and in particular to a monitoring and analysis method for electrical equipment in linkage with a camera and a monitoring system. Background Art

[0002] During the operation of electrical equipment, it is crucial to ensure its safe and stable operation. Traditional electrical equipment monitoring mainly relies on manual inspections, but this method has many limitations. On the one hand, manual inspections are inefficient and difficult to achieve real-time and continuous monitoring of equipment, which can easily lead to missed inspections; on the other hand, for some complex abnormal situations, manual judgments may not be accurate and timely enough.

[0003] With the development of science and technology, the application of cameras and monitoring systems in electrical equipment monitoring has gradually increased. However, there are still some challenges, such as how to effectively use camera resources during non-patrol tasks to improve the monitoring capabilities of key areas; how to quickly and accurately identify abnormal conditions of electrical equipment from the video stream obtained by the camera, including abnormalities in equipment appearance (oil leakage, deformation, fracture, tilt, etc.), operating environment (fireworks, foreign objects, etc.) and personnel behavior (small animal intrusion, failure to wear safety equipment correctly, abnormal crowd gathering, etc.); when the electrical equipment monitoring system detects potential problems and sends out alarm signals, how to ensure that the cameras in the patrol system can respond quickly and accurately jump to the preset position to take pictures or videos to capture the scene; how to conduct a more in-depth and accurate analysis of the data captured by the linkage to confirm the authenticity and severity of the abnormality; and how to effectively feedback the final analysis results to the operator and provide a basis for subsequent performance optimization.

[0004] In order to solve these problems, there is an urgent need for a more advanced, efficient and accurate method of monitoring and analyzing electrical equipment by linking cameras and monitoring systems to improve the operation and maintenance level of electrical equipment and ensure the safe and stable operation of the power system. Summary of the invention

[0005] The main purpose of the present invention is to provide a monitoring and analysis method for electrical equipment by linking a camera and a monitoring system, so as to solve the problem that manual judgment may not be accurate and timely in complex abnormal situations.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for monitoring and analyzing electrical equipment by linking a camera and a monitoring system, the method comprising: S1. The patrol system determines the key areas that need to be monitored, and the cameras of the patrol system continuously monitor the preset key areas during the non-patrol task execution period; S2, obtain the video stream from the camera for subsequent analysis and processing; S3. Use computer vision technology to process real-time video streams, extract key features from images, call trained deep learning models, identify abnormal conditions of electrical equipment, detect abnormal conditions of electrical equipment, immediately generate alarm information, and record relevant timestamps and location information; S4. When the electrical equipment monitoring system detects a potential problem with the equipment, it sends an alarm signal to the patrol system. The patrol host responds quickly, controls the designated camera to jump to the preset position, and starts taking pictures or recording videos to capture the on-site situation; S5. Store the captured images or recorded video clips, and use the intelligent analysis algorithm to conduct a more in-depth evaluation of the captured data to confirm the authenticity and severity of the anomaly; S6. Feedback the final analysis results to the operator to provide auxiliary decision-making suggestions. All operation processes and analysis results should be recorded in detail for future review and performance optimization.

[0007] In the preferred solution, in step S1, corresponding abnormality detection rules are set for the device appearance, operating environment, and personnel behavior; Among them, the equipment appearance abnormality detection rules include oil leakage, equipment deformation, fracture, and tilt; The rules for detecting abnormalities in the operating environment include fireworks, hanging objects in the air; Abnormal personnel behavior detection rules include small animal intrusion, failure to wear a hard hat correctly, failure to wear long-sleeved work clothes, and personnel gathering / wandering.

[0008] In the preferred embodiment, the specific steps in step S3 are: S3.1. From the real-time video stream, Select a frame of image and extract the feature points in the image through the improved scale-invariant feature transformation algorithm; suppose the image The feature point set extracted from , ; At the same time, calculate the local binary pattern texture features of the local area around each feature point and oriented gradient histogram shape features ; In order to introduce time dimension information, a spatiotemporal feature vector is constructed; for each feature point , and place it in a continuous Frame, where , the position changes within ,in Indicates time frame, LBP feature changes And HOG feature changes Combined into space-time feature vector ; set up for The spatiotemporal feature vector set of the image at each moment, then the spatiotemporal feature matrix of the entire video sequence is ,in is the total number of frames selected; this spatiotemporal feature matrix can simultaneously describe the characteristic changes of electrical equipment in spatial and temporal dimensions, providing richer information for accurately identifying abnormal situations; S3.2. Improvement of deep learning model and anomaly identification: A hybrid deep learning model based on three-dimensional convolutional neural network (3D-CNN) and long short-term memory (LSTM) is used to identify anomalies. 3D-CNN is used to process spatial and local temporal features in the spatiotemporal feature matrix, and LSTM is used to learn the feature change pattern of long time series. 3D - CNN's convolutional layer calculates the spatiotemporal feature matrix Output , the convolution kernel slides simultaneously in the spatial and temporal dimensions, and the convolution kernel weight is , the bias term is , the convolutional layer calculation process is as follows: ; in is the size of the convolution kernel in the spatial dimension, is the size of the convolution kernel in the time dimension, , is the spatial coordinate, is the time coordinate, The function is an activation function. This formula can effectively extract local patterns in spatiotemporal features through convolution operations in three dimensions, and has a good perception of abnormal features of electrical equipment in spatial layout and time changes. The output of 3D-CNN Input into the LSTM network, the hidden state update formula of LSTM is: ; ; ; ; ; ; in for The hidden state of the moment, is the cell state, , , are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the sigmoid function, represents element-wise multiplication, and are the corresponding weight matrices and bias items; LSTM can learn the dependencies in long time series data through these gating mechanisms, model the long-term trend of electrical equipment state changes, and thus more accurately determine whether there are abnormalities; In the last layer of the model, a fully connected layer is used to classify abnormal situations and output the classification results , including normal, oil leakage, and deformation; S3.3, Dynamic threshold adaptive adjustment and abnormal alarm: During the model training process, for each abnormal category , calculate the feature distance distribution of the model output in the training sample ;set up For abnormal category No. The model output feature vector of training samples is is the normal sample feature center vector, then the distance ,in, represents the distance metric, ; Distribution by distance , a method based on kernel density estimation (KDE) is used to calculate the dynamic threshold ; The KDE formula is: ; in As the kernel function, the Gaussian kernel function is used , is the bandwidth parameter; dynamic threshold Can be based on The distribution of ,in, is a pre-set probability threshold, such as $0.95$); this dynamic threshold can be automatically adjusted according to the distribution of training data to meet the detection requirements of different abnormal situations and improve the accuracy of abnormal alarms; In the real-time monitoring process, for each classification result , calculate the distance between the model output feature and the normal sample feature center ,if ,in, If the dynamic threshold value of the corresponding abnormal category is exceeded, it is judged as an abnormal situation, an alarm message is generated immediately, and the current timestamp is recorded. And the location information of the feature points in the image, through the feature point coordinates Sure.

[0009] In the preferred embodiment, the specific steps in step S4 are: S4.1. The receiving module in the patrol system continuously monitors the alarm signal sent by the electrical equipment monitoring system; after receiving the alarm signal, it analyzes the signal content and obtains the type of potential problem of the equipment (type, location of occurrence) and severity information; Based on device location And the camera's installation position and viewing angle parameters, calculate the spatial distance between each camera and the device and perspective bias , Indicates the camera number; spatial distance The calculation formula is as follows: ; in For the The spatial coordinates of the camera; the viewing angle deviation It is obtained by calculating the angle between the device position vector and the camera orientation vector. This distance formula is used to measure the distance between each camera and the device. The viewing angle deviation formula is used to evaluate whether the camera can be directly aimed at the device. The two together serve as the basis for selecting the best camera. S4.2. Define comprehensive evaluation indicators of cameras , taking into account the distance , Perspective Bias And the current working status of the camera , the calculation formula is: ; in, The closer the distance, the higher the priority. The smaller the viewing angle deviation, the higher the priority. Make sure to prioritize available cameras. Indicates that the priority is adjusted according to the severity of the problem. The higher the severity, the greater the priority increase. According to the calculated comprehensive evaluation index , sort all cameras in descending order, select The camera with the largest value is used as the target camera; if there are multiple cameras with the same maximum value value, we further compare their historical response accuracy, which is obtained by recording the ratio of the number of times each camera successfully captured valid information in the past to the total number of responses, and select the camera with higher accuracy; S4.3. For the selected target camera, obtain its current position and attitude, pitch angle , yaw angle , Roll Angle , and the preset target location and target attitude (pitch angle , yaw angle , Roll Angle ); The path planning method based on the improved A* algorithm is used to search for the optimal path from the current position to the target position in the camera's motion space; the path is assumed to consist of a series of intermediate points. ,in, , is the number of path points, the path cost function Defined as: ; in , and is the weight coefficient, which is used to adjust the importance of spatial position change and attitude angle change in the path cost. This formula takes into account the cost of camera attitude adjustment while considering the shortest spatial path to achieve more efficient and smooth path planning. According to the planned path, the camera is controlled to pass through each intermediate point in sequence to quickly jump to the preset position; during the jump process, the position and posture of the camera are monitored in real time, and the path is fine-tuned according to the actual situation; S4.4. When the camera jumps to the preset position, it immediately starts taking pictures or recording videos to obtain images of the equipment site or video sequence ; At the same time, record the shooting start time and end time , and environmental parameters during filming, which together with the image or video data serve as a complete record of the scene; The captured data is preprocessed, including image denoising, using a denoising algorithm based on wavelet transform. The formula is as follows: Assume the original image is , the high frequency coefficients after wavelet decomposition are , the threshold is , then the high-frequency coefficient after denoising is , and then the denoised image is obtained by wavelet reconstruction , video frame rate adjustment operation to improve data quality and transmission efficiency; The pre-processed image or video data as well as time and environmental parameters are transmitted to the monitoring center via a high-speed network to ensure timely and accurate data delivery so that operation and maintenance personnel can conduct further analysis and processing.

[0010] In the preferred embodiment, the specific steps in step S5 are: S5.1. Store the captured images or recorded video clips in the local database and mark the corresponding timestamp for each data file. , Camera number and device location information and other metadata for subsequent query and analysis; For image data, a feature extraction method based on convolutional neural network (CNN) is used; the image is input into the pre-trained CNN model, and the output of a specific layer in the network is selected as the deep feature vector of the image. ; picture The feature map after the CNN model is , whose dimensions are , , is the height and width of the feature map, is the number of feature channels, then the depth feature vector It is obtained by performing spatial pooling operation on the feature map, using global average pooling: ; in is the depth feature vector No. elements, The feature map In Location The The formula can convert the two-dimensional feature map into a one-dimensional feature vector, effectively reducing the feature dimension while retaining the key information in the image for subsequent anomaly assessment. For a video clip, first split it into a series of consecutive video frames ,in is the number of video frames; Then, for each video frame, the same CNN feature extraction method as the image is used to obtain the corresponding deep feature vector ; Then, by calculating the optical flow features between feature vectors of adjacent video frames , the optical flow features are fused with the depth feature vector to obtain the spatiotemporal feature vector of the video clip ; A simple way to merge can be a splicing operation: ;Such spatiotemporal feature vectors can simultaneously capture the spatial and temporal information in the video and better describe the state changes of electrical equipment in the dynamic process; S5.2. Collect a large amount of electrical equipment image and video sample data with labeled abnormality types and severity, and divide them into training sets and test set ; For each sample , marking its exception type and severity ; Based on Bayesian decision theory, calculate the given feature vector In the case of and severity The posterior probability ; According to the Bayesian formula: ; in is the conditional probability, indicating that when the abnormal type is known and severity In the case of probability; is the prior probability, indicating the abnormal type in all samples and severity Probability of occurrence; is the evidence factor, which is used to normalize the posterior probability and is calculated as follows: ; To calculate , assuming that the eigenvector Under each anomaly type and severity, a certain probability distribution is followed for the deep feature vector , let it be in the exception type and severity The mean vector under , the covariance matrix is ,but The calculation formula is: ; in is the dimension of the feature vector; for the spatiotemporal feature vector of the video ,Due to its complex structure, a mixed Gaussian model can be used to describe its probability distribution; Estimating prior probability through training set data And the parameters of the conditional probability distribution; then, for the test set or the data to be evaluated in actual monitoring, according to the calculated posterior probability To determine the authenticity and severity of the anomaly; You can select the anomaly type and severity with the largest posterior probability as the evaluation result, or set a probability threshold , when the maximum a posteriori probability is greater than When a real abnormality is detected, it is considered that the real abnormality is detected, and corresponding measures are taken according to the severity; S5.3. Use the test set The constructed anomaly assessment model is evaluated for performance; indicators such as accuracy, recall, and F1-score are used to measure the classification performance of the model. The calculation formula is as follows: ; ; ; Among them, TP is the true positive example, which indicates the number of samples correctly predicted by the model as abnormal; TN is the true negative example, which indicates the number of samples correctly predicted by the model as normal; FP is the false positive example, which indicates the number of samples incorrectly predicted by the model as abnormal; FN is the false negative example, which indicates the number of samples incorrectly predicted by the model as normal; these indicators can comprehensively evaluate the accuracy of the model in anomaly detection and severity classification; If the model performance does not meet the requirements, optimize the model according to the evaluation results; the optimization methods may include adjusting the prior probability estimation method in the Bayesian decision model, improving the feature extraction method, expanding the number of training samples or improving the sample quality; The model evaluation and optimization process is repeated until the model performance reaches a satisfactory level, ensuring that the model can accurately assess the authenticity and severity of electrical equipment anomalies in practical applications.

[0011] In the preferred embodiment, the specific steps in step S6 are: S6.1. The analysis results include the abnormality type, severity, occurrence time, and location information, and the final analysis results are fed back to the operator in an intuitive and easy-to-understand manner; Generate auxiliary decision-making suggestions based on the analysis results and the preset operation and maintenance strategy rule base; First, define a decision influencing factor matrix , where the elements Indicates The decision factor The degree of influence of a decision option can be quantified through expert experience evaluation or data analysis, and the value range is [0, 1]. Indicates no impact, Indicates strong influence; Set abnormal severity , Equipment Importance , repair cost , then for the decision option "immediately shut down for maintenance" , decision-making influencing factors One way to calculate can be: ; This formula comprehensively considers the impact of abnormality severity, equipment importance and repair cost on the decision of immediate shutdown for maintenance; the more serious the abnormality, the more important the equipment, and the relatively low the repair cost, the more inclined to immediate shutdown for maintenance. The impact factor calculated by this formula can reflect this comprehensive impact relationship; similarly, the impact factors of other decision factors on different decision options can be calculated to construct a complete decision impact factor matrix ; Calculate the combined impact of each decision option , using the weighted summation method: ; in For the The weight of the decision factors can be determined by the analytic hierarchy process (AHP). AHP constructs a hierarchical model of decision factors, compares the relative importance of each factor, constructs a judgment matrix, and then solves the eigenvector of the judgment matrix to obtain the weight. Suppose the judgment matrix , where the elements Indicates The decision factor is relative to The importance ratio of decision factors, then the weight vector By solving the following equation: ; in is the judgment matrix The maximum eigenvalue of ; Determining the weights through AHP can more scientifically reflect the importance of different decision factors in the decision-making process; According to the comprehensive impact value Sort the decision options and select the one with the largest comprehensive impact value as the recommended decision suggestion to the operator; if ( , , If the comprehensive impact values ​​are "immediately stop for maintenance", "arrange emergency inspection", and "continue observation", the operator is recommended to select "immediately stop for maintenance"; S6.2. Record all operation processes in detail, including the camera startup time , jump path information $path$, shooting time , data transmission time , analyze the execution time of the algorithm and the time when the decision proposal was generated At the same time, record the key parameter values ​​in each step, such as the comprehensive evaluation index in the camera selection process , Posterior Probability in Anomaly Assessment Model ; For the analysis results, in addition to the above-mentioned abnormality type, severity and other information, the intermediate results of the analysis process are also recorded, such as the feature vector after feature extraction or , Prior Probability in Bayesian Decision Model and conditional probability distribution parameters; organize these operation processes and analysis result data into structured record files , for subsequent review and performance optimization; During the recording process, some key data are preprocessed to improve data availability and analysis efficiency. For example, for time series data, normalization is performed to make its value range between $[0, 1]$, which is convenient for subsequent data analysis and comparison. The original time value is set to , the normalized time value The calculation formula is: ; in and are the minimum and maximum values ​​of the time parameter in all records respectively; this formula can unify time data of different scales into the same range, which is convenient for comparison and correlation analysis in subsequent analysis; S6.3. Regularly review records Review and analyze the data in order to evaluate the performance of the system and find optimization opportunities; first, analyze the time parameters during the operation and calculate the average time of each operation step and time fluctuations (such as standard deviation ); Analyze the accuracy and stability of abnormality assessment results; calculate the false positive rate (ER) and false negative rate (MR) for different abnormality types and severities, and analyze the stability of the results over different time periods or on different equipment types; If the misjudgment rate or missed judgment rate is high, or the result stability is poor, re-examine the anomaly assessment model; it may be necessary to adjust the prior probability estimation method in the Bayesian decision model, improve the feature extraction method, or expand the number and quality of training samples; at the same time, according to the performance analysis results, the decision-making influencing factor matrix and decision factor weights Make optimization adjustments to improve the accuracy and rationality of decision-making support recommendations.

[0012] The present invention provides a monitoring and analysis method for electrical equipment by linking a camera and a monitoring system, which can monitor key areas uninterruptedly during non-patrol tasks, realize all-weather monitoring of electrical equipment, greatly improve the proportion of time that the equipment is monitored, effectively reduce the risk of abnormal situations being missed, and ensure that the equipment operating status is always within the observable range.

[0013] Multiple cameras acquire data from different perspectives and combine a variety of computer vision technologies to process video streams. They can capture device information from multiple dimensions and levels, and conduct comprehensive monitoring of abnormal conditions in multiple aspects such as device appearance, operating environment, and personnel behavior, without missing any factors that may affect the safe operation of the equipment.

[0014] Advanced deep learning models are used for anomaly identification. After being trained with a large amount of data, they can accurately determine whether electrical equipment has appearance problems such as oil leakage, deformation, breakage, tilt, etc., as well as monitor fireworks, foreign objects and violations in personnel behavior in the operating environment, greatly improving the accuracy of anomaly detection and reducing false alarms and missed alarms.

[0015] Intelligent analysis algorithms are applied again to the data captured in the linkage process to conduct in-depth evaluations from the perspectives of multimodal data fusion and spatiotemporal feature analysis. This can not only confirm the authenticity of the anomaly, but also accurately determine the severity of the anomaly, providing a more valuable basis for subsequent decision-making and helping operation and maintenance personnel to take targeted measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a flow chart of the monitoring and analysis method of the present invention; DETAILED DESCRIPTION Example 1 like Figure 1 As shown, a method for monitoring and analyzing electrical equipment by linkage of a camera and a monitoring system, the method comprising: S1. The patrol system determines the key areas that need to be monitored, and the cameras of the patrol system continuously monitor the preset key areas during the non-patrol task execution period; S2, obtain the video stream from the camera for subsequent analysis and processing; S3. Use computer vision technology to process real-time video streams, extract key features from images, call trained deep learning models, identify abnormal conditions of electrical equipment, detect abnormal conditions of electrical equipment, immediately generate alarm information, and record relevant timestamps and location information; S4. When the electrical equipment monitoring system detects a potential problem with the equipment, it sends an alarm signal to the patrol system. The patrol host responds quickly, controls the designated camera to jump to the preset position, and starts taking pictures or recording videos to capture the on-site situation; S5. Store the captured images or recorded video clips, and use the intelligent analysis algorithm to conduct a more in-depth evaluation of the captured data to confirm the authenticity and severity of the anomaly; S6. Feedback the final analysis results to the operator to provide auxiliary decision-making suggestions. All operation processes and analysis results should be recorded in detail for future review and performance optimization.

[0017] When the monitoring system detects a potential problem, the patrol host quickly controls the camera to jump and capture the on-site situation. The whole process is fast and efficient, ensuring that key information when an abnormality occurs is obtained at the first time, which buys valuable time for timely handling of the problem.

[0018] Generate auxiliary decision-making suggestions based on detailed records of operating processes and analysis results, comprehensively consider various factors, help operators make decisions quickly and scientifically, improve operation and maintenance efficiency, avoid equipment damage or power accidents caused by wrong decisions, and provide data support for long-term operation and maintenance optimization.

[0019] Completely record all operation processes and analysis results for easy review in the future, and be able to clearly trace every link in the system operation, which helps to discover potential problems and deficiencies.

[0020] By regularly reviewing and analyzing recorded data, the system can be optimized for performance in a targeted manner, such as improving algorithms, adjusting parameters, optimizing camera layout, etc., so that the system can continuously adapt to changes in the actual operating environment and needs, and continuously improve monitoring and analysis capabilities and overall performance.

[0021] In the preferred solution, in step S1, corresponding abnormality detection rules are set for the device appearance, operating environment, and personnel behavior; Among them, the equipment appearance abnormality detection rules include oil leakage, equipment deformation, fracture, and tilt; The rules for detecting abnormalities in the operating environment include fireworks, hanging objects in the air; Abnormal personnel behavior detection rules include small animal intrusion, failure to wear a hard hat correctly, failure to wear long-sleeved work clothes, and personnel gathering / wandering.

[0022] Example 2 Further illustrate with reference to Example 1, Figure 1 The structure shown in FIG. 1 , the specific steps in step S3 are: S3.1. From the real-time video stream, Select a frame of image and extract the feature points in the image through the improved scale-invariant feature transformation algorithm; suppose the image The feature point set extracted from , ; At the same time, calculate the local binary pattern texture features of the local area around each feature point and oriented gradient histogram shape features ; In order to introduce time dimension information, a spatiotemporal feature vector is constructed; for each feature point , and place it in a continuous Frame, where , the position changes within ,in Indicates time frame, LBP feature changes And HOG feature changes Combined into space-time feature vector ; set up for The spatiotemporal feature vector set of the image at each moment, then the spatiotemporal feature matrix of the entire video sequence is ,in is the total number of frames selected; this spatiotemporal feature matrix can simultaneously describe the characteristic changes of electrical equipment in spatial and temporal dimensions, providing richer information for accurately identifying abnormal situations; By extracting the key features of the image and constructing a spatiotemporal feature matrix, the state change information of the electrical equipment can be captured more comprehensively, including not only the current appearance state of the equipment, but also considering its dynamic changes in a short period of time, which helps to detect abnormal situations more accurately.

[0023] S3.2. Improvement of deep learning model and anomaly identification: A hybrid deep learning model based on three-dimensional convolutional neural network (3D-CNN) and long short-term memory (LSTM) is used to identify anomalies. 3D-CNN is used to process spatial and local temporal features in the spatiotemporal feature matrix, and LSTM is used to learn the feature change pattern of long time series. 3D - CNN's convolutional layer calculates the spatiotemporal feature matrix Output , the convolution kernel slides simultaneously in the spatial and temporal dimensions, and the convolution kernel weight is , the bias term is , the convolutional layer calculation process is as follows: ; in is the size of the convolution kernel in the spatial dimension, is the size of the convolution kernel in the time dimension, , is the spatial coordinate, is the time coordinate, The function is an activation function. This formula can effectively extract local patterns in spatiotemporal features through convolution operations in three dimensions, and has a good perception of abnormal features of electrical equipment in spatial layout and time changes. The output of 3D-CNN Input into the LSTM network, the hidden state update formula of LSTM is: ; ; ; ; ; ; in for The hidden state of the moment, is the cell state, , , are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the sigmoid function, represents element-wise multiplication, and are the corresponding weight matrices and bias items; LSTM can learn the dependencies in long time series data through these gating mechanisms, model the long-term trend of electrical equipment state changes, and thus more accurately determine whether there are abnormalities; In the last layer of the model, a fully connected layer is used to classify abnormal situations and output the classification results , including normal, oil leakage, and deformation; By leveraging the advantages of 3D-CNN and LSTM, we conduct in-depth analysis of the spatiotemporal characteristics of electrical equipment to improve the model’s ability to identify abnormal situations, especially for some abnormal situations that require comprehensive consideration of spatial layout changes and time trends (such as gradual deformation of equipment, potential risks caused by environmental changes, etc.).

[0024] S3.3, Dynamic threshold adaptive adjustment and abnormal alarm: During the model training process, for each abnormal category , calculate the feature distance distribution of the model output in the training sample ;set up For exception category No. The model output feature vector of training samples is is the normal sample feature center vector, then the distance ,in, represents the distance metric, ; Distribution by distance , a method based on kernel density estimation (KDE) is used to calculate the dynamic threshold ; The KDE formula is: ; in As the kernel function, the Gaussian kernel function is used , is the bandwidth parameter; dynamic threshold Can be based on The distribution of ,in, is a pre-set probability threshold, such as $0.95$); this dynamic threshold can be automatically adjusted according to the distribution of training data to meet the detection requirements of different abnormal situations and improve the accuracy of abnormal alarms; In the real-time monitoring process, for each classification result , calculate the distance between the model output feature and the normal sample feature center ,if ,in, If the dynamic threshold value of the corresponding abnormal category is exceeded, it is judged as an abnormal situation, an alarm message is generated immediately, and the current timestamp is recorded. And the location information of the feature points in the image, through the feature point coordinates Sure.

[0025] The dynamic threshold adaptive adjustment mechanism avoids the misjudgment and missed judgment problems that may occur in fixed thresholds in different scenarios and equipment states. It can more accurately judge abnormal situations based on actual data distribution, improve the reliability and adaptability of anomaly detection, and ensure that accurate alarm information is issued in a timely manner so that operation and maintenance personnel can take corresponding measures.

[0026] Example 3 Further described in conjunction with Example 1, the specific steps in step S4 are: S4.1. The receiving module in the patrol system continuously monitors the alarm signal sent by the electrical equipment monitoring system; after receiving the alarm signal, it analyzes the signal content and obtains the type of potential problem of the equipment (type, location of occurrence) and severity information; Based on device location And the camera's installation position and viewing angle parameters, calculate the spatial distance between each camera and the device and perspective bias , Indicates the camera number; spatial distance The calculation formula is as follows: ; in For the The spatial coordinates of the camera; the viewing angle deviation It is obtained by calculating the angle between the device position vector and the camera orientation vector. This distance formula is used to measure the distance between each camera and the device. The viewing angle deviation formula is used to evaluate whether the camera can be directly aimed at the device. The two together serve as the basis for selecting the best camera. Accurately analyze the alarm signal content, determine the relevant information of the equipment problem, and calculate the key parameters between the camera and the equipment, providing basic data for the subsequent rapid selection of the most suitable camera, ensuring that the on-site situation of the equipment can be captured in a timely and accurate manner.

[0027] S4.2. Define comprehensive evaluation indicators of cameras , taking into account the distance , Perspective Bias And the current working status of the camera , the calculation formula is: ; in, The closer the distance, the higher the priority. The smaller the viewing angle deviation, the higher the priority. Make sure to prioritize available cameras. Indicates that the priority is adjusted according to the severity of the problem. The higher the severity, the greater the priority increase. According to the calculated comprehensive evaluation index , sort all cameras in descending order, select The camera with the largest value is used as the target camera; if there are multiple cameras with the same maximum value value, we further compare their historical response accuracy, which is obtained by recording the ratio of the number of times each camera successfully captured valid information in the past to the total number of responses, and select the camera with higher accuracy; By prioritizing cameras through comprehensive evaluation indicators, it is possible to quickly select the camera that is most suitable for capturing on-site conditions of potential equipment problems from multiple cameras. At the same time, the severity of the equipment problem and the historical performance of the camera are taken into consideration to improve the accuracy and effectiveness of camera selection.

[0028] S4.3. For the selected target camera, obtain its current position and attitude, pitch angle , yaw angle , Roll Angle , and the preset target location and target attitude (pitch angle , yaw angle , Roll Angle ); The path planning method based on the improved A* algorithm is used to search for the optimal path from the current position to the target position in the camera's motion space; the path is assumed to consist of a series of intermediate points. ,in, , is the number of path points, the path cost function Defined as: ; in , and is the weight coefficient, which is used to adjust the importance of spatial position change and attitude angle change in the path cost. This formula takes into account the cost of camera attitude adjustment while considering the shortest spatial path to achieve more efficient and smooth path planning. According to the planned path, the camera is controlled to pass through each intermediate point in sequence to quickly jump to the preset position; during the jump process, the position and posture of the camera are monitored in real time, and the path is fine-tuned according to the actual situation; For example, if an obstacle or other interference factor is encountered, the target location can be reached smoothly by locally replanning the path or adjusting the movement speed.

[0029] Through the optimized path planning algorithm, the camera can jump to the preset position in the fastest and most stable way. At the same time, it can adaptively adjust the path during movement to overcome various obstacles that may be encountered, improve the camera's response speed to potential equipment problems and the accuracy of capturing on-site conditions.

[0030] S4.4. When the camera jumps to the preset position, it immediately starts taking pictures or recording videos to obtain images of the equipment site or video sequence ; At the same time, record the shooting start time and end time , and environmental parameters during filming, which together with the image or video data serve as a complete record of the scene; The captured data is preprocessed, including image denoising, using a denoising algorithm based on wavelet transform. The formula is as follows: Assume the original image is , the high-frequency coefficients after wavelet decomposition are , the threshold is , then the high-frequency coefficient after denoising is , and then the denoised image is obtained by wavelet reconstruction , video frame rate adjustment operation to improve data quality and transmission efficiency; The pre-processed image or video data as well as time and environmental parameters are transmitted to the monitoring center via a high-speed network to ensure timely and accurate data delivery so that operation and maintenance personnel can conduct further analysis and processing.

[0031] Accurately and completely capture the on-site conditions of the equipment, pre-process the data and efficiently transmit it, providing high-quality data support for subsequent abnormal analysis and decision-making, ensuring that operation and maintenance personnel can obtain detailed information on the equipment site in a timely manner and make accurate judgments and handling measures.

[0032] Example 4 Further described in conjunction with Example 1, the specific steps in step S5 are: S5.1. Store the captured images or recorded video clips in the local database and mark the corresponding timestamp for each data file. , Camera number and device location information and other metadata for subsequent query and analysis; For image data, a feature extraction method based on convolutional neural network (CNN) is used; the image is input into the pre-trained CNN model, and the output of a specific layer in the network is selected as the deep feature vector of the image. ; picture The feature map after the CNN model is , whose dimensions are , , is the height and width of the feature map, is the number of feature channels, then the depth feature vector It is obtained by performing spatial pooling operation on the feature map, using global average pooling: ; in is the depth feature vector No. elements, The feature map In Location The The formula can convert the two-dimensional feature map into a one-dimensional feature vector, effectively reducing the feature dimension while retaining the key information in the image for subsequent anomaly assessment. For a video clip, first split it into a series of consecutive video frames ,in is the number of video frames; Then, for each video frame, the same CNN feature extraction method as the image is used to obtain the corresponding deep feature vector ; Then, by calculating the optical flow features between feature vectors of adjacent video frames , the optical flow features are fused with the depth feature vector to obtain the spatiotemporal feature vector of the video clip ; A simple way to merge can be a splicing operation: ;Such spatiotemporal feature vectors can simultaneously capture the spatial and temporal information in the video and better describe the state changes of electrical equipment in the dynamic process; Standardized data storage and extraction of effective feature vectors provide rich and accurate data representation for subsequent anomaly assessment. The depth feature vector of the image can reflect the static appearance characteristics of the device, while the spatiotemporal feature vector of the video can reflect the change of the device status over time. The combination of the two can more comprehensively assess the anomaly.

[0033] S5.2. Collect a large amount of electrical equipment image and video sample data with labeled abnormality types and severity, and divide them into training sets and test set ; For each sample , marking its exception type and severity ; Based on Bayesian decision theory, calculate the given feature vector In the case of and severity The posterior probability ; According to the Bayesian formula: ; in is the conditional probability, indicating that when the abnormal type is known and severity In the case of The probability of is the prior probability, indicating the abnormal type in all samples and severity Probability of occurrence; is the evidence factor, which is used to normalize the posterior probability and is calculated as follows: ; To calculate , assuming that the eigenvector Under each anomaly type and severity, a certain probability distribution is followed for the deep feature vector , let it be in the exception type and severity The mean vector under , the covariance matrix is ,but The calculation formula is: ; in is the dimension of the feature vector; for the spatiotemporal feature vector of the video ,Due to its complex structure, a mixed Gaussian model can be used to describe its probability distribution; Estimating prior probability through training set data And the parameters of the conditional probability distribution; then, for the test set or the data to be evaluated in actual monitoring, according to the calculated posterior probability To determine the authenticity and severity of the anomaly; You can select the anomaly type and severity with the largest posterior probability as the evaluation result, or set a probability threshold , when the maximum a posteriori probability is greater than When a real abnormality is detected, it is considered that the real abnormality is detected, and corresponding measures are taken according to the severity; Using Bayesian decision theory to construct an anomaly assessment model can comprehensively consider prior knowledge (anomaly distribution in training data) and new data (feature vectors monitored in real time) to more accurately judge the authenticity and severity of anomalies, reduce the possibility of misjudgment and missed judgment, and provide a reliable basis for operation and maintenance decisions.

[0034] S5.3. Use the test set The constructed anomaly assessment model is evaluated for performance; indicators such as accuracy, recall, and F1-score are used to measure the classification performance of the model. The calculation formula is as follows: ; ; ; Among them, TP is the true positive example, which indicates the number of samples correctly predicted by the model as abnormal; TN is the true negative example, which indicates the number of samples correctly predicted by the model as normal; FP is the false positive example, which indicates the number of samples incorrectly predicted by the model as abnormal; FN is the false negative example, which indicates the number of samples incorrectly predicted by the model as normal; these indicators can comprehensively evaluate the accuracy of the model in anomaly detection and severity classification; If the model performance does not meet the requirements, optimize the model according to the evaluation results; the optimization methods may include adjusting the prior probability estimation method in the Bayesian decision model, improving the feature extraction method, expanding the number of training samples or improving the sample quality; Repeat the model evaluation and optimization process until the model performance reaches a satisfactory level, ensuring that the model can accurately evaluate the authenticity and severity of electrical equipment anomalies in actual applications. By evaluating and optimizing the model, its performance is continuously improved, so that it can better adapt to various complex situations in actual monitoring scenarios, providing more reliable protection for the safe operation of electrical equipment.

[0035] Example 5 Further described in conjunction with Example 1, the specific steps in step S6 are: S6.1. The analysis results include the abnormality type, severity, occurrence time, and location information, and the final analysis results are fed back to the operator in an intuitive and easy-to-understand manner; Generate auxiliary decision-making suggestions based on the analysis results and the preset operation and maintenance strategy rule base; First, define a decision influencing factor matrix , where the elements Indicates The decision factor The degree of influence of a decision option can be quantified through expert experience evaluation or data analysis, and the value range is [0, 1]. Indicates no impact, Indicates strong influence; Set abnormal severity , Equipment Importance , repair cost , then for the decision option "immediately shut down for maintenance" , decision-making influencing factors One way to calculate can be: ; This formula comprehensively considers the impact of abnormality severity, equipment importance and repair cost on the decision of immediate shutdown for maintenance; the more serious the abnormality, the more important the equipment, and the relatively low the repair cost, the more inclined to immediate shutdown for maintenance. The impact factor calculated by this formula can reflect this comprehensive impact relationship; similarly, the impact factors of other decision factors on different decision options can be calculated to construct a complete decision impact factor matrix ; Calculate the combined impact of each decision option , using the weighted summation method: ; in For the The weight of the decision factors can be determined by the analytic hierarchy process (AHP). AHP constructs a hierarchical model of decision factors, compares the relative importance of each factor, constructs a judgment matrix, and then solves the eigenvector of the judgment matrix to obtain the weight. Suppose the judgment matrix , where the elements Indicates The decision factor is relative to The importance ratio of decision factors, then the weight vector By solving the following equation: ; in is the judgment matrix The maximum eigenvalue of ; Determining the weights through AHP can more scientifically reflect the importance of different decision factors in the decision-making process; According to the comprehensive impact value Sort the decision options and select the one with the largest comprehensive impact value as the recommended decision suggestion to the operator; if ( , , If the comprehensive impact values ​​are "immediately stop for maintenance", "arrange emergency inspection", and "continue observation", the operator is recommended to select "immediately stop for maintenance"; The analysis results are presented to operators in a clear and effective manner, and reasonable auxiliary decision-making suggestions are generated based on multiple factors to help operators make decisions quickly and accurately to deal with abnormal situations of electrical equipment, thereby improving operation and maintenance efficiency and equipment reliability.

[0036] S6.2. Record all operation processes in detail, including the camera startup time , jump path information $path$, shooting time , data transmission time , analyze the execution time of the algorithm and the time when the decision proposal was generated At the same time, record the key parameter values ​​in each step, such as the comprehensive evaluation index in the camera selection process , Posterior Probability in Anomaly Assessment Model ; For the analysis results, in addition to the above-mentioned abnormality type, severity and other information, the intermediate results of the analysis process are also recorded, such as the feature vector after feature extraction or , Prior Probability in Bayesian Decision Model and conditional probability distribution parameters; organize these operation processes and analysis result data into structured record files , for subsequent review and performance optimization; During the recording process, some key data are preprocessed to improve data availability and analysis efficiency; for example, for time series data, normalization is performed to make its value range between $[0, 1]$, which is convenient for subsequent data analysis and comparison; the original time value is set to , the normalized time value The calculation formula is: ; in and are the minimum and maximum values ​​of the time parameter in all records respectively; this formula can unify time data of different scales into the same range, which is convenient for comparison and correlation analysis in subsequent analysis; Completely and accurately record the operation process and analysis results to provide a detailed basis for future review. At the same time, through data preprocessing, a convenient data basis is provided for subsequent performance optimization, which helps to discover problems and improvement directions during system operation and continuously improve system performance and reliability.

[0037] S6.3. Regularly review records Review and analyze the data in order to evaluate the performance of the system and find optimization opportunities; first, analyze the time parameters during the operation and calculate the average time of each operation step and time fluctuations (such as standard deviation ); Analyze the accuracy and stability of abnormality assessment results; calculate the false positive rate (ER) and false negative rate (MR) for different abnormality types and severities, and analyze the stability of the results over different time periods or on different equipment types; If the false positive rate or missed positive rate is high, or the result stability is poor, re-examine the abnormal assessment model. It may be necessary to adjust the prior probability estimation method in the Bayesian decision model (such as adopting a more dynamic prior probability update strategy to continuously revise the prior probability based on real-time monitoring data), improve the feature extraction method (such as increasing the diversity of features or adopting a more advanced deep learning feature extraction model), or expand the number and quality of training samples (collect more abnormal samples under different working conditions).

[0038] At the same time, according to the performance analysis results, the decision-making factor matrix and decision factor weights Make optimization adjustments to improve the accuracy and rationality of decision-making support recommendations.

[0039] Through in-depth analysis of recorded data, we can discover performance bottlenecks and problems during system operation, make targeted optimization adjustments, and continuously improve the system's performance in electrical equipment monitoring and analysis and auxiliary decision-making, ensuring that the system can adapt to changing operation and maintenance needs and provide stronger support for the safe and stable operation of electrical equipment.

[0040] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for monitoring and analyzing electrical equipment by linking a camera and a monitoring system, characterized in that: The method includes: S1. The patrol system determines the key areas that need to be monitored, and the cameras of the patrol system continuously monitor the preset key areas during the non-patrol task execution period; S2, obtain the video stream from the camera for subsequent analysis and processing; S3. Use computer vision technology to process real-time video streams, extract key features from images, call trained deep learning models, identify abnormal conditions of electrical equipment, detect abnormal conditions of electrical equipment, immediately generate alarm information, and record relevant timestamps and location information; S4. When the electrical equipment monitoring system detects a potential problem with the equipment, it sends an alarm signal to the patrol system. The patrol host responds quickly, controls the designated camera to jump to the preset position, and starts taking pictures or recording videos to capture the on-site situation; S5. Store the captured images or recorded video clips, and use the intelligent analysis algorithm to conduct a more in-depth evaluation of the captured data to confirm the authenticity and severity of the anomaly; S6. Feedback the final analysis results to the operator to provide auxiliary decision-making suggestions. All operation processes and analysis results should be recorded in detail for future review and performance optimization.

2. According to claim 1, a method for monitoring and analyzing electrical equipment by linking a camera and a monitoring system, characterized in that: In step S1, corresponding abnormality detection rules are set based on the device appearance, operating environment, and personnel behavior; Among them, the equipment appearance abnormality detection rules include oil leakage, equipment deformation, fracture, and tilt; The rules for detecting abnormalities in the operating environment include fireworks, hanging objects in the air; Abnormal personnel behavior detection rules include small animal intrusion, failure to wear a hard hat correctly, failure to wear long-sleeved work clothes, and personnel gathering / wandering.

3. According to claim 1, a method for monitoring and analyzing electrical equipment by linking a camera and a monitoring system, characterized in that: The specific steps in step S3 are: S3.

1. From the real-time video stream, Select a frame of image and extract the feature points in the image through the improved scale-invariant feature transformation algorithm; suppose the image The feature point set extracted from , ; At the same time, calculate the local binary pattern texture features of the local area around each feature point and oriented gradient histogram shape features ; In order to introduce time dimension information, a spatiotemporal feature vector is constructed; for each feature point , and place it in a continuous Frame, where , the position changes within ,in Indicates time frame, LBP feature changes And HOG feature changes Combined into space-time feature vector ; set up for The spatiotemporal feature vector set of the image at each moment, then the spatiotemporal feature matrix of the entire video sequence is ,in is the total number of frames selected; this spatiotemporal feature matrix can simultaneously describe the characteristic changes of electrical equipment in spatial and temporal dimensions, providing richer information for accurately identifying abnormal situations; S3.

2. Improvement of deep learning model and anomaly identification: A hybrid deep learning model based on three-dimensional convolutional neural network (3D-CNN) and long short-term memory (LSTM) is used to identify anomalies. 3D-CNN is used to process spatial and local temporal features in the spatiotemporal feature matrix, and LSTM is used to learn the feature change pattern of long time series. 3D - CNN's convolutional layer calculates the spatiotemporal feature matrix Output , the convolution kernel slides simultaneously in the spatial and temporal dimensions, and the convolution kernel weight is , the bias term is , the convolutional layer calculation process is as follows: ； in is the size of the convolution kernel in the spatial dimension, is the size of the convolution kernel in the time dimension, , is the spatial coordinate, is the time coordinate, The function is an activation function. This formula can effectively extract local patterns in spatiotemporal features through convolution operations in three dimensions, and has a good perception of abnormal features of electrical equipment in spatial layout and time changes. The output of 3D-CNN Input into the LSTM network, the hidden state update formula of LSTM is: ； ； ； ； ； ； in for The hidden state of the moment, is the cell state, , , are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the sigmoid function, represents element-wise multiplication, and are the corresponding weight matrices and bias items; LSTM can learn the dependencies in long time series data through these gating mechanisms, model the long-term trend of electrical equipment state changes, and thus more accurately determine whether there are abnormalities; In the last layer of the model, a fully connected layer is used to classify abnormal situations and output the classification results , including normal, oil leakage, and deformation; S3.3, Dynamic threshold adaptive adjustment and abnormal alarm: During the model training process, for each abnormal category , calculate the feature distance distribution of the model output in the training sample ;set up For exception category No. The model output feature vector of training samples is is the normal sample feature center vector, then the distance ,in, represents the distance metric, ; Distribution by distance , a method based on kernel density estimation (KDE) is used to calculate the dynamic threshold ; The KDE formula is: ； in As the kernel function, the Gaussian kernel function is used , is the bandwidth parameter; dynamic threshold Can be based on The distribution of ,in, is a pre-set probability threshold, such as $0.95$); this dynamic threshold can be automatically adjusted according to the distribution of training data to meet the detection requirements of different abnormal situations and improve the accuracy of abnormal alarms; In the real-time monitoring process, for each classification result , calculate the distance between the model output feature and the normal sample feature center ,if ,in, If the dynamic threshold value of the corresponding abnormal category is exceeded, it is judged as an abnormal situation, an alarm message is generated immediately, and the current timestamp is recorded. And the location information of the feature points in the image, through the feature point coordinates Sure.

4. According to claim 1, a method for monitoring and analyzing electrical equipment by linking a camera and a monitoring system, characterized in that: The specific steps in step S4 are: S4.

1. The receiving module in the patrol system continuously monitors the alarm signal sent by the electrical equipment monitoring system; after receiving the alarm signal, it analyzes the signal content and obtains the type of potential problem of the equipment (type, location of occurrence) and severity information; Based on device location And the camera's installation position and viewing angle parameters, calculate the spatial distance between each camera and the device and perspective bias , Indicates the camera number; spatial distance The calculation formula is as follows: ； in For the The spatial coordinates of the cameras; Perspective Bias It is obtained by calculating the angle between the device position vector and the camera orientation vector; The distance formula is used to measure the distance between each camera and the device, and the viewing angle deviation formula is used to evaluate whether the camera can be pointed directly at the device. Together, they serve as the basis for selecting the best camera. S4.

2. Define comprehensive evaluation indicators of cameras , taking into account the distance , Perspective Bias And the current working status of the camera , the calculation formula is: ； in, The closer the distance, the higher the priority. The smaller the viewing angle deviation, the higher the priority. Make sure to prioritize available cameras. Indicates that the priority is adjusted according to the severity of the problem. The higher the severity, the greater the priority increase. According to the calculated comprehensive evaluation index , sort all cameras in descending order, select The camera with the largest value is used as the target camera; if there are multiple cameras with the same maximum value value, we further compare their historical response accuracy, which is obtained by recording the ratio of the number of times each camera successfully captured valid information in the past to the total number of responses, and select the camera with higher accuracy; S4.

3. For the selected target camera, obtain its current position and attitude, pitch angle , yaw angle , Roll Angle , and the preset target location and target attitude (pitch angle , yaw angle , Roll Angle ); The path planning method based on the improved A* algorithm is used to search for the optimal path from the current position to the target position in the camera's motion space; the path is assumed to consist of a series of intermediate points. ,in, , is the number of path points, and the path cost function Defined as: ； in , and is the weight coefficient, which is used to adjust the importance of spatial position change and attitude angle change in the path cost. This formula takes into account the cost of camera attitude adjustment while considering the shortest spatial path to achieve more efficient and smooth path planning. According to the planned path, the camera is controlled to pass through each intermediate point in sequence to quickly jump to the preset position; during the jump process, the position and posture of the camera are monitored in real time, and the path is fine-tuned according to the actual situation; S4.

4. When the camera jumps to the preset position, it immediately starts taking pictures or recording videos to obtain images of the equipment site or video sequence ; At the same time, record the shooting start time and end time , and environmental parameters during filming, which together with the image or video data serve as a complete record of the scene; The captured data is preprocessed, including image denoising, using a denoising algorithm based on wavelet transform. The formula is as follows: Assume the original image is , the high-frequency coefficients after wavelet decomposition are , the threshold is , then the high-frequency coefficient after denoising is , and then the denoised image is obtained by wavelet reconstruction , video frame rate adjustment operation to improve data quality and transmission efficiency; The pre-processed image or video data as well as time and environmental parameters are transmitted to the monitoring center via a high-speed network to ensure timely and accurate data delivery so that operation and maintenance personnel can conduct further analysis and processing.

5. According to claim 1, a method for monitoring and analyzing electrical equipment by linking a camera and a monitoring system, characterized in that: The specific steps in step S5 are: S5.

1. Store the captured images or recorded video clips in the local database and mark the corresponding timestamp for each data file. , Camera number and device location information and other metadata for subsequent query and analysis; For image data, a feature extraction method based on convolutional neural network (CNN) is used; Input the image into the pre-trained CNN model and select the output of a specific layer in the network as the deep feature vector of the image ; picture The feature map after the CNN model is , whose dimensions are , , is the height and width of the feature map, is the number of feature channels, then the depth feature vector It is obtained by performing spatial pooling operation on the feature map, using global average pooling: ； in is the depth feature vector No. elements, The feature map In Location The The formula can convert the two-dimensional feature map into a one-dimensional feature vector, effectively reducing the feature dimension while retaining the key information in the image for subsequent anomaly assessment. For a video clip, first split it into a series of consecutive video frames ,in is the number of video frames; Then, for each video frame, the same CNN feature extraction method as the image is used to obtain the corresponding deep feature vector ; Then, by calculating the optical flow features between feature vectors of adjacent video frames , the optical flow features are fused with the depth feature vector to obtain the spatiotemporal feature vector of the video clip ; A simple way to merge can be a splicing operation: ;Such spatiotemporal feature vectors can simultaneously capture the spatial and temporal information in the video and better describe the state changes of electrical equipment in the dynamic process; S5.

2. Collect a large amount of electrical equipment image and video sample data with labeled abnormality types and severity, and divide them into training sets and test set ; For each sample , marking its exception type and severity ; Based on Bayesian decision theory, calculate the given feature vector In the case of and severity The posterior probability ; According to the Bayesian formula: ； in is the conditional probability, indicating that when the abnormal type is known and severity In the case of probability; is the prior probability, indicating the abnormal type in all samples and severity Probability of occurrence; is the evidence factor, which is used to normalize the posterior probability and is calculated as follows: ; To calculate , assuming that the eigenvector Under each anomaly type and severity, a certain probability distribution is followed for the deep feature vector , let it be in the exception type and severity The mean vector under , the covariance matrix is ,but The calculation formula is: ； in is the dimension of the feature vector; for the spatiotemporal feature vector of the video ,Due to its complex structure, a mixed Gaussian model can be used to describe its probability distribution; Estimating prior probability through training set data And the parameters of the conditional probability distribution; then, for the test set or the data to be evaluated in actual monitoring, according to the calculated posterior probability To determine the authenticity and severity of the anomaly; You can select the anomaly type and severity with the largest posterior probability as the evaluation result, or set a probability threshold , when the maximum a posteriori probability is greater than When a real abnormality is detected, it is considered that the real abnormality is detected, and corresponding measures are taken according to the severity; S5.

3. Use the test set The constructed anomaly assessment model is evaluated for performance; indicators such as accuracy, recall, and F1-score are used to measure the classification performance of the model. The calculation formula is as follows: ； ； ； Among them, TP is the true positive example, which indicates the number of samples correctly predicted by the model as abnormal; TN is the true negative example, which indicates the number of samples correctly predicted by the model as normal; FP is the false positive example, which indicates the number of samples incorrectly predicted by the model as abnormal; FN is the false negative example, which indicates the number of samples incorrectly predicted by the model as normal; these indicators can comprehensively evaluate the accuracy of the model in anomaly detection and severity classification; If the model performance does not meet the requirements, optimize the model according to the evaluation results; the optimization methods may include adjusting the prior probability estimation method in the Bayesian decision model, improving the feature extraction method, expanding the number of training samples or improving the sample quality; The model evaluation and optimization process is repeated until the model performance reaches a satisfactory level, ensuring that the model can accurately assess the authenticity and severity of electrical equipment anomalies in practical applications.

6. According to claim 1, a method for monitoring and analyzing electrical equipment by linking a camera and a monitoring system, characterized in that: step S6 The specific steps are: S6.

1. The analysis results include the abnormality type, severity, occurrence time, and location information, and the final analysis results are fed back to the operator in an intuitive and easy-to-understand manner; Generate auxiliary decision-making suggestions based on the analysis results and the preset operation and maintenance strategy rule base; First, define a decision influencing factor matrix , where the elements Indicates The decision factor The degree of influence of a decision option; the quantification of the degree of influence can be obtained through expert experience evaluation or data analysis, and the value range is [0,1]. Indicates no impact, Indicates strong influence; Set abnormal severity , Equipment Importance , repair cost , then for the decision option "immediately shut down for maintenance" , decision-making influencing factors One way to calculate can be: ； This formula comprehensively considers the impact of abnormality severity, equipment importance and repair cost on the decision of immediate shutdown for maintenance; the more serious the abnormality, the more important the equipment, and the relatively low the repair cost, the more inclined to immediate shutdown for maintenance. The impact factor calculated by this formula can reflect this comprehensive impact relationship; similarly, the impact factors of other decision factors on different decision options can be calculated to construct a complete decision impact factor matrix ; Calculate the combined impact of each decision option , using the weighted summation method: ； in For the The weight of each decision factor can be determined by the analytic hierarchy process (AHP). AHP constructs a hierarchical model of decision factors, compares the relative importance of each factor, constructs a judgment matrix, and then solves the eigenvector of the judgment matrix to obtain the weight. Assume the judgment matrix , where the elements Indicates The decision factor is relative to The importance ratio of decision factors, then the weight vector By solving the following equation: ； in is the judgment matrix The maximum eigenvalue of ; Determining the weights through AHP can more scientifically reflect the importance of different decision factors in the decision-making process; According to the comprehensive impact value Sort the decision options and select the one with the largest comprehensive impact value as the recommended decision suggestion to the operator; if ( , , If the comprehensive impact values ​​are "immediately stop for maintenance", "arrange emergency inspection", and "continue observation", the operator is recommended to select "immediately stop for maintenance"; S6.

2. Record all operation processes in detail, including the camera startup time , jump path information $path$, shooting time , data transmission time , analyze the execution time of the algorithm and the time when the decision proposal was generated At the same time, record the key parameter values ​​in each step, such as the comprehensive evaluation index in the camera selection process , Posterior Probability in Anomaly Assessment Model ; For the analysis results, in addition to the above-mentioned abnormality type, severity and other information, the intermediate results of the analysis process are also recorded, such as the feature vector after feature extraction or , Prior Probability in Bayesian Decision Model and conditional probability distribution parameters; organize these operation processes and analysis result data into structured record files , for subsequent review and performance optimization; During the recording process, some key data are preprocessed to improve data availability and analysis efficiency; for example, for time series data, normalization is performed to make its value range between $[0, 1]$, which is convenient for subsequent data analysis and comparison; the original time value is set to , the normalized time value The calculation formula is: ； in and are the minimum and maximum values ​​of the time parameter in all records respectively; This formula can unify time data of different scales into the same range, making it easier to compare and correlate in subsequent analyses; S6.

3. Regularly review records Review and analyze the data in order to evaluate the performance of the system and find optimization opportunities; first, analyze the time parameters during the operation and calculate the average time of each operation step and time fluctuations (such as standard deviation ); Analyze the accuracy and stability of abnormal assessment results; Calculate the false positive rate (ER) and false negative rate (MR) for different abnormality types and severity levels, and analyze the stability of the results over different time periods or different equipment types; If the misjudgment rate or missed judgment rate is high, or the result stability is poor, re-examine the anomaly assessment model; it may be necessary to adjust the prior probability estimation method in the Bayesian decision model, improve the feature extraction method, or expand the number and quality of training samples; at the same time, according to the performance analysis results, the decision-making influencing factor matrix and decision factor weights Make optimization adjustments to improve the accuracy and rationality of decision-making support recommendations.

Images