A multi-dimensional sensing panoramic monitoring method and system for power transmission lines

Abstract

This application discloses a multi-dimensional sensing panoramic monitoring method and system for transmission lines. The method includes: acquiring and preprocessing multi-dimensional sensing data for each acquisition point of the transmission line; performing feature engineering on the preprocessed multi-dimensional sensing data of each acquisition point, and fusing them to obtain feature vectors for each acquisition point; constructing a random forest regression model; defining the health status of the transmission line and generating health status labels; selecting feature vectors of acquisition points labeled with health status in historical transmission lines as training data and inputting them into the constructed random forest regression model for model training; inputting the currently acquired feature vectors of each acquisition point into the trained random forest regression model, obtaining and analyzing the health status assessment results of each acquisition point, and generating a health status warning. This application can accurately assess the health status of transmission lines and provide timely warnings, improving the monitoring efficiency and security of power systems.

Description

Technical Field

[0001] This application relates to the technical field of intelligent monitoring of transmission lines, specifically to a multi-dimensional perception panoramic monitoring method and system for transmission lines. Background Technology

[0002] With the continuous development of power systems, the importance of intelligent monitoring in the power sector is becoming increasingly prominent. In application scenarios such as remote monitoring of power transmission lines and preventive maintenance of power equipment, effective monitoring systems can ensure the stable operation of power systems, reduce the probability of faults, and minimize economic losses and social impacts caused by power outages and other faults.

[0003] In the traditional field of power monitoring, various methods are mainly used to monitor transmission lines and related power equipment. On the one hand, simple fixed monitoring equipment is installed on the towers, such as devices capable of basic image capture and basic temperature monitoring sensors, to obtain limited line status information. On the other hand, regular manual inspections are also arranged, with maintenance personnel conducting on-site checks of the condition of transmission lines and equipment.

[0004] However, existing power monitoring systems have significant shortcomings, with severe information silos and a lack of effective data fusion mechanisms between different types of monitoring equipment. This makes it impossible to achieve multi-dimensional data fusion analysis, resulting in a significant reduction in accuracy when issuing health warnings due to a lack of comprehensive data support, making it difficult to quickly and accurately determine when to issue health warnings. Summary of the Invention

[0005] In order to promptly detect substandard health status of transmission lines and generate early warnings, solve the problem of information silos in existing power monitoring systems, realize multi-dimensional data fusion analysis, and improve early warning capabilities, this application provides a multi-dimensional perception panoramic monitoring method and system for transmission lines.

[0006] Firstly, this application provides a multi-dimensional sensing panoramic monitoring method for transmission lines, including:

[0007] For each acquisition point of the transmission line, multi-dimensional sensing data of the acquisition point is acquired, including: image data of the acquisition point, time series data of various sensors, lidar point cloud data of the distance between the acquisition point and obstacles around the acquisition point, and environmental data, and preprocessing is performed for each acquisition point;

[0008] Feature engineering is performed on the preprocessed multidimensional sensing data from each acquisition point, and feature vectors from each acquisition point are obtained by fusion. The feature engineering process includes: obtaining image features, temporal features of various sensors, and point cloud features of LiDAR using a single-modal feature extraction method; obtaining cross-modal interaction features based on physical rules and cross-modal interaction features based on correlation analysis rules using a cross-modal interaction feature construction method; and completing feature selection using a Gini importance selection method.

[0009] Construct a random forest regression model; define the health status of transmission lines and generate health status labels for transmission lines; select the feature vectors of each collection point marked with health status in historical transmission lines as training data, input them into the constructed random forest regression model, and obtain the trained random forest regression model;

[0010] Input the feature vectors of each currently acquired collection point into the trained random forest regression model to obtain the health assessment results of each collection point; compare the health assessment results of each collection node with the preset health threshold, and generate a health warning for collection nodes that are less than the preset health threshold.

[0011] By adopting the above scheme, feature engineering and fusion of multi-dimensional sensing data from various collection points can be performed, which can extract effective features from multi-source heterogeneous data, improve data quality and usability; by using the constructed random forest regression model to obtain health assessment results and comparing them with preset health thresholds, the system can promptly detect situations where the health of transmission lines does not meet the standards and generate early warnings, solve the problem of information silos in the existing power monitoring system, realize multi-dimensional data fusion analysis, and improve early warning capabilities.

[0012] Preferred options also include:

[0013] For each time-series feature after feature engineering, a pre-built LSTM prediction model is used to predict the features of the corresponding future time period.

[0014] The features predicted for future time periods based on each time-series feature after feature engineering are fused together with the features after feature engineering to obtain a brand-new feature vector. The brand-new feature vectors of each currently acquired collection point are then input into the retrained random forest regression model to obtain the health assessment results of each collection point, replacing the original health assessment results of each collection point. Among these, the brand-new feature vectors of each collection point marked with health in the historical transmission line are selected as incremental training data to retrain the random forest regression model.

[0015] Construct an isolated forest anomaly detection model; using historical data of features predicted by an LSTM prediction model in various dimensions as input, train the isolated forest anomaly detection model to identify normal and abnormal feature data in each dimension, and output labeled abnormal feature data in each dimension; use the trained isolated forest anomaly detection model to determine whether there are abnormal feature data labels in the feature data of each dimension predicted by the LSTM prediction model for the corresponding future time period, and combine deep learning algorithms to trigger a preset type of anomaly warning for feature data of each dimension with abnormal data labels.

[0016] By adopting the above scheme, a pre-constructed LSTM prediction model is used to predict future time-series features after feature engineering. These features are then fused with the original features to form a new feature vector. A retrained random forest regression model is used to evaluate the health of the collection points, improving the accuracy and timeliness of the health assessment and reflecting the future health status of the transmission lines. An isolated forest anomaly detection model is constructed to identify normal and abnormal data in the feature dimensions predicted by the LSTM prediction model. Combined with deep learning algorithms, anomaly data triggers early warnings, enabling timely detection of potential anomalies in the transmission lines and ensuring their safe and stable operation.

[0017] Preferably, it also includes: collecting terrain data and equipment data for each collection point of the transmission line;

[0018] For each data collection point in the transmission line, including environmental data, terrain data, and equipment data, the scenario type corresponding to the current data collection point is determined, including: normal scenario and abnormal scenario; the abnormal scenario includes: extreme weather scenario, complex terrain scenario, external force damage scenario, and equipment aging scenario.

[0019] If the current acquisition point is determined to be an anomalous scene, feature engineering optimization is performed on the preprocessed multidimensional perception data of the current acquisition point to replace feature engineering processing, and feature vectors of each acquisition point are obtained by fusion. The feature engineering optimization process includes: obtaining image features, time-series features of various sensors, and LiDAR point cloud features using a single-modal feature extraction method; obtaining cross-modal interaction features constructed based on physical rules that conform to the current anomalous scene and cross-modal interaction features constructed based on correlation analysis rules using a cross-modal interaction feature construction method; and completing feature selection using a Gini importance selection method.

[0020] Multiple random forest regression models are constructed, and each random forest regression model is trained using the feature vectors of historical collection points under a certain scenario. Based on the currently determined abnormal scenario, the corresponding random forest regression model is obtained, and the feature vectors of each collection point obtained at the moment are input into the trained random forest regression model to obtain the health assessment results of each collection point.

[0021] By adopting the above scheme, terrain data and equipment data are collected at each collection point of the transmission line. Combined with environmental data, the scene type is determined and abnormal scenes are identified. In abnormal scenes, feature engineering optimization is performed on the multi-dimensional sensing data to obtain feature vectors that conform to the physical rules of abnormal scenes. Multiple random forest regression models for different scenes are constructed, and the corresponding model is selected according to the abnormal scene to evaluate the health, thereby improving the accuracy of transmission line health assessment in abnormal scenes.

[0022] Preferred options also include:

[0023] The health of transmission lines is defined by a weighted assessment of multiple dimensions of indicators, including: structural health, electrical health, and environmental health.

[0024] For each acquisition point of the transmission line, the status data of the corresponding multi-dimensional sensing data acquisition device is obtained, and the weights of different dimension indicators are dynamically adjusted according to the status data of the multi-dimensional sensing data acquisition device. This includes: counting the number of devices whose status data of the corresponding acquisition devices belonging to the dimensions of structural health, electrical health and environmental health are at risk of failure, and dynamically adjusting the weights of different dimension indicators according to the proportion of the number of devices at risk of failure counted in different dimensions.

[0025] By adopting the above scheme, the health status of transmission lines is comprehensively measured by defining a multi-dimensional weighted evaluation of indicators. The weights of different indicators are dynamically adjusted according to the status data of the multi-dimensional sensing data acquisition device, so that the health status assessment is closer to the actual situation, avoiding the impact of some acquisition device failures on the accuracy of the assessment, and improving the reliability and effectiveness of the assessment results.

[0026] Preferably, the construction of the random forest regression model includes:

[0027] Initialize the random forest parameters, including: number of decision trees, maximum depth, and minimum number of samples for split nodes;

[0028] Bootstrap sampling is used to extract a subset of samples with replacement from the feature vector of the original data, construct multiple decision trees, and randomly select some features when splitting each decision tree to obtain several final decision trees as training data.

[0029] The training data is used to fit a random forest regression model. The input of the model is the feature vector obtained from the collection points. The prediction output is made by individual decision trees. The mean of the prediction output of all individual decision trees is calculated or the cluster mean is obtained by K-means clustering to obtain the overall prediction output of the forest and obtain the health assessment result.

[0030] By adopting the above scheme, a random forest regression model is fitted using training data. The feature vectors of the collection points are used as input, and the overall prediction output of the forest is obtained by calculating the mean of the prediction output of a single decision tree or the mean of K-means clustering. This allows for a more accurate assessment of the health of each collection point on the transmission line.

[0031] Preferred options also include:

[0032] The edge-cloud collaborative computing architecture is used to deploy and complete the health and anomaly warning of transmission lines. This includes: using the edge computing center to complete the feature engineering processing of multi-dimensional sensing data, feature prediction for future time periods, and anomaly warning using the isolated forest detection model; and using the cloud computing center to complete the health assessment results using the random forest regression model.

[0033] By adopting the above solution, the health and anomaly early warning systems for transmission lines are deployed using an edge-cloud collaborative computing architecture, reducing data transmission pressure, improving processing efficiency and the real-time nature of anomaly early warning, and ensuring the accuracy and stability of health assessment.

[0034] Preferred options also include:

[0035] After feature selection using the Gini importance screening method, VIF collinearity detection is performed on the selected features to obtain the variance inflation coefficient of each selected feature, and features with variance inflation coefficients not greater than the preset variance inflation coefficient are retained.

[0036] By adopting the above scheme, VIF collinearity detection is performed on the screened features, retaining features whose variance inflation coefficient is not greater than the preset value, avoiding the interference of collinearity between features on the model, and improving the accuracy of the random forest regression model in assessing the health of transmission lines.

[0037] Secondly, this application provides a multi-dimensional sensing panoramic monitoring system for power transmission lines, comprising:

[0038] The power transmission line data acquisition module is used to acquire multi-dimensional sensing data for each acquisition point of the power transmission line, including: image data of the acquisition point, time series data of various sensors, lidar point cloud data of the distance between the acquisition point and obstacles around the acquisition point, and environmental data, and to perform preprocessing for each acquisition point;

[0039] The transmission line data processing module is used to perform feature engineering processing on the pre-processed multi-dimensional sensing data of each acquisition point, and fuse them to obtain the feature vectors of each acquisition point. The feature engineering processing includes: obtaining image features, time-series features of various sensors, and lidar point cloud features using single-modal feature extraction methods; obtaining cross-modal interaction features constructed based on physical rules and cross-modal interaction features constructed based on correlation analysis rules using cross-modal interaction feature construction methods; and completing feature selection using Gini importance filtering methods.

[0040] The power transmission line health assessment module is used to construct a random forest regression model; define the health status of power transmission lines and generate health status labels; select feature vectors of each collection point marked with health status in historical power transmission lines as training data, input them into the constructed random forest regression model, and obtain the trained random forest regression model; input the feature vectors of each currently acquired collection point into the trained random forest regression model respectively, and obtain the corresponding health status assessment results of each collection point; compare the health status assessment results of each collection node with the preset health status threshold, and generate a health status warning for collection nodes that are lower than the preset health status threshold.

[0041] By adopting the above scheme, a transmission line data acquisition module is designed to acquire and preprocess multi-dimensional sensing data from each acquisition point of the transmission line, providing a comprehensive and accurate data foundation for subsequent analysis; a transmission line data processing module is designed to perform feature engineering and fusion on the preprocessed data to effectively extract data features; and a transmission line health assessment module is designed to accurately assess the health of each acquisition point by constructing and training a random forest regression model, and to issue early warnings for acquisition nodes with health values ​​lower than preset values, thereby realizing real-time monitoring and effective management of the transmission line.

[0042] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the method described above.

[0043] Fourthly, this application provides a computer device, the computer device including a memory, a processor and a program stored in the memory and executable thereon, the program being executed by the processor to implement the steps of the method described above.

[0044] In summary, this application has the following beneficial effects:

[0045] 1. Integrating multiple types of sensors to collect multi-dimensional sensing data and performing preprocessing, feature engineering, and fusion, the data is input into a constructed and trained random forest regression model to evaluate the health of transmission lines, promptly detect potential faults and anomalies, solve the information silo problem of existing power monitoring systems, realize the fusion analysis of multi-dimensional data, improve early warning capabilities, and ensure the stable operation of the power system;

[0046] 2. Utilize the LSTM prediction model to predict the features of future time periods and fuse them with the original features. Retrain the random forest regression model to obtain more accurate health assessment results. At the same time, construct an isolated forest anomaly detection model to identify abnormal data. Combine deep learning algorithms to trigger early warnings for abnormal data, improve the prediction and early warning capabilities for potential anomalies in transmission lines, and ensure the stable operation of transmission lines.

[0047] 3. Considering abnormal scenarios, feature engineering optimization is performed on abnormal scenarios to construct random forest regression models adapted to different scenarios, thereby improving the accuracy of transmission line health assessment under abnormal scenarios. Attached Figure Description

[0048] Figure 1 This is a flowchart of the multi-dimensional perception panoramic monitoring method for transmission lines described in a specific embodiment;

[0049] Figure 2 This is a schematic diagram of the structure of the multi-dimensional perception panoramic monitoring system for power transmission lines described in a specific embodiment. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] This application mainly adopts a scheme of collecting multi-dimensional data of transmission lines to assess health and issue early warnings, which achieves the effect of breaking down information silos and improving health early warning capabilities. The following is a further detailed description of this application.

[0052] like Figure 1 As shown in the figure, this application discloses a multi-dimensional perception panoramic monitoring method for transmission lines, including: data acquisition and preprocessing, feature engineering processing, model building and training, health assessment and early warning, etc., which are described step by step below.

[0053] S1. For each acquisition point of the transmission line, acquire multi-dimensional sensing data of the acquisition point.

[0054] Specifically, several key points (such as towers and conductors) are selected as data acquisition points along the transmission line, and data acquisition devices are installed at these points to collect multi-dimensional sensing data. This includes image data (visible light images, infrared images, etc.), time-series data from various sensors (such as conductor temperature, ice thickness, insulator leakage current, partial discharge, tower tilt angle, etc.), lidar point cloud data showing the distances between the data acquisition points and surrounding obstacles (such as the minimum distance from the conductor to the ground, the distance from the conductor to the top of a tree), and environmental data.

[0055] Image data can be collected using visual monitoring equipment on power poles, such as video surveillance equipment and image capture devices. These devices are typically installed at appropriate locations on the poles to comprehensively and clearly capture the transmission lines and their surrounding conditions. The acquisition of time-series data from various sensors relies on multiple sensors, such as conductor temperature monitoring sensors to collect conductor temperature data and conductor galloping status monitoring sensors to monitor conductor galloping status data. These sensors must be securely installed on the poles or conductors to ensure the accuracy of data acquisition. LiDAR point cloud data is collected by LiDAR systems mounted on drones or other mobile devices, providing distance information between the collection point and surrounding obstacles (such as trees and buildings). Environmental data is collected through environmental monitoring sensors, such as small integrated weather stations and icing sensors.

[0056] Preprocessing is performed for each collection point; the preprocessing process includes data cleaning to remove noise and erroneous data from the collected data, and format unification to convert data of different formats into a comparable unified format; data cleaning can use filtering algorithms, such as Gaussian filtering to remove noise, and format unification can be based on the corresponding conversion rules formulated according to the data type.

[0057] S2. Perform feature engineering processing on the preprocessed multidimensional sensing data of each acquisition point, and fuse them to obtain the feature vector of each acquisition point.

[0058] To avoid information silos and accurately complete the fusion analysis of multi-dimensional data, the key lies in processing data features. This application sets up feature engineering processing technology that includes multi-layer data processing to achieve a good fusion effect and assist in subsequent health assessment.

[0059] The feature engineering process specifically includes three layers. The first layer is single-modal feature extraction, which extracts features of the current dimension from the currently collected data. The second layer considers the physical coupling and interrelationships between multiple dimensions of data, thereby constructing cross-modal interactive features. The third layer, based on the feature data obtained in the first two layers, ensures accurate subsequent health assessment by calculating feature importance and selecting features with high importance. The details of each layer will be described below.

[0060] First, for obtaining image features, temporal features of various sensors, and point cloud features of LiDAR using a single-modal feature extraction method, the entire feature extraction can be set to include: dynamic feature extraction, static feature extraction, and spatiotemporal feature extraction.

[0061] For image data acquisition, image feature extraction can employ algorithms such as edge detection and feature point matching, or model algorithms such as YOLOv5 / ResNet and GLCM, to extract static features such as morphological features and texture features, including: the proportion of insulator damage area, pollution level, conductor strand breakage, and tower corrosion area segmentation, etc. For example, in conductor strand breakage identification, the image is first preprocessed, using a 5×5 Gaussian filter for noise reduction, then the gradient is calculated using the Sobel operator, edge localization is performed using the Canny algorithm, and finally the defect size is calculated; and considering time-series image analysis, dynamic features such as conductor galloping trajectory and icing growth rate are extracted.

[0062] For various sensor time-series characteristics, feature extraction can be performed by using a sliding window to extract time-domain features, such as the mean / variance of temperature and vibration intensity, the kurtosis coefficient of vibration, and the leakage current pulse count of insulators; by using FFT transform and wavelet packet energy entropy to extract frequency-domain features, such as the fundamental frequency of conductor galloping and the non-stationary characteristics of vibration signals; by using correlation analysis techniques to extract spatiotemporal correlation features; and by extracting time-series data trend features through indicators such as hourly / daily rate of change and abrupt change amplitude, such as the hourly rate of change of tower tilt / ice thickness (hourly rate of change = (current value - value 1 hour ago) / 1 hour), daily rate of change (daily rate of change = (current value - value 24 hours ago) / 24 hours), and abrupt change amplitude (abrupt change amplitude = current value - sliding window mean).

[0063] To address the characteristics of lidar point cloud data, it is necessary to first preprocess the lidar point cloud data, use a random forest classifier to separate the point cloud types, then perform Poisson reconstruction to generate a continuous three-dimensional model, and then extract features such as conductor sag, distance to the ground, and tree height, specifically including: minimum Euclidean distance between conductor and tree, minimum distance from conductor to ground, and surface fitting of conductor clearance height to ground (catenary equation fitting), etc.

[0064] Secondly, cross-modal interaction features constructed based on physical rules and cross-modal interaction features constructed based on correlation analysis rules are obtained by using cross-modal interaction feature construction methods.

[0065] Cross-modal interaction characteristics constructed based on physical rules can be built according to the physical characteristics and operating laws of transmission lines. For example, by combining materials science formulas, the coupling characteristics between icing thickness and ambient temperature can be calculated to characterize the icing stress caused by sudden temperature changes, and a cross-modal interaction characteristic can be constructed equal to the icing thickness. Ambient temperature gradient; For example, based on electromagnetic principles, combined with leakage current pulse counting and relative humidity, a cross-modal interaction characteristic is constructed to characterize the accelerated insulation degradation caused by wet pollution synergy: insulator leakage pulse. Ambient humidity; similarly, there exists a cross-modal interaction characteristic that combines conductor galloping amplitude and wind speed to characterize the wind-induced galloping enhancement effect = conductor galloping amplitude. Ambient wind speed, etc.

[0066] Cross-modal interaction features constructed based on correlation analysis rules are built by analyzing the correlation between different modal data, such as: calculating the cross-correlation function of tower tilt and wind speed, calculating the nonlinear correlation strength with tree distance and conductor vibration amplitude through mutual information; and Granger causality test results of icing thickness and ambient temperature and humidity.

[0067] Then, feature selection is performed using the Gini importance ranking method. This includes: ranking the features acquired at each collection point according to their Gini importance, setting an importance threshold (e.g., 0.01), retaining features with importance greater than the threshold, and removing other unimportant features. Furthermore, further optimization of the important feature selection strategy can be planned, including: after feature selection using the Gini importance ranking method, performing VIF collinearity detection on the selected features, obtaining the variance inflation coefficient of each selected feature, retaining features with variance inflation coefficients no greater than a preset variance inflation coefficient, and finally using SHAP values ​​to explain feature contributions and avoid Gini importance bias.

[0068] Finally, feature fusion is performed on the above features from each collection point, such as feature concatenation, to obtain the fused multidimensional feature vector corresponding to each collection point.

[0069] S3. Construct a random forest regression model; define the health status of transmission lines and generate health status labels for transmission lines; train the random forest regression model and use the trained random forest regression model to complete the health status assessment results for each collection point.

[0070] First, we construct a random forest regression model.

[0071] Initialize the random forest parameters, including: number of decision trees, maximum depth, minimum number of samples per split node; obtain training data.

[0072] Bootstrap sampling is employed, extracting a subset of samples with replacement from the feature vectors of the original data to construct multiple decision trees. Each decision tree is then split by randomly selecting a subset of features, resulting in several final decision trees used as training data. The feature vectors used as the original data refer to the standardized feature vectors from each historical data point. The decision tree splitting criteria adhere to the following principles: minimizing the mean squared error and randomly selecting features during splitting. Features (reducing correlation between trees); the splitting threshold t satisfies: In the formula, N is the number of samples in the current node. , y represents the mean health index of the left and right subtree samples, respectively; y is the sample label (health index labeled by experts).

[0073] A random forest regression model is fitted using training data. The model's input is the feature vector obtained from the collection points, and the output is the health assessment result of the collection points. To further optimize the random forest regression model, this embodiment adopts collaborative decision-making between individual trees and the forest to obtain the final health assessment result of the collection points. This includes: pre-predicting the output using individual decision trees, calculating the mean of the predicted outputs of all individual decision trees, or using K-means clustering to obtain the overall predicted output of the forest, and obtaining the health assessment result.

[0074] Suppose the i-th decision tree in the forest is , ( (This is a standardized feature vector). Suppose the forest contains m decision trees, and the final health index H is the mean of the predictions from all trees: Alternatively, K-means clustering can be used to calculate the cluster mean, discarding predictions for individual trees that are far from the cluster centers, and retaining the remaining ones. ,get This allows you to map the final health score to a percentage system.

[0075] Secondly, the health status of transmission lines was defined and health status labels were generated.

[0076] Specifically, the health status labels for each data collection point in the transmission line can be set to a percentage system. The specific health status label (score) settings are shown in Table 1 below:

[0077] Table 1 Health Labels

[0078]

[0079] Furthermore, to better refine the health label settings, a multi-dimensional weighted evaluation of transmission line health can be defined, including: structural health, electrical health, and environmental health; the formula is: .

[0080] Accordingly, a random forest regression model is fitted using training data. The input of the model is the feature vector obtained from the collection points, and the output is the health assessment result of the collection points. The process includes first obtaining the structural health, electrical health, and environmental health, and then calculating them by weight.

[0081] The weighting of each health indicator can be dynamically adjusted based on the status data of the multi-dimensional sensing data acquisition devices. Specifically, this includes: acquiring status data of the corresponding multi-dimensional sensing data acquisition devices for each acquisition point on the transmission line, including whether there is a risk of failure; separately counting the number of devices with status data indicating a risk of failure in the corresponding dimensions of structural health, electrical health, and environmental health; and dynamically adjusting the weights of different indicators according to the proportion of devices with risk of failure in each dimension. Furthermore, the multi-dimensional sensing data is pre-defined to belong to specific health levels, such as: tower tilt angle belonging to structural health, peak leakage current belonging to electrical health, and distance between conductors and trees belonging to environmental health.

[0082] Then, the constructed random forest regression model is trained using the training data.

[0083] Specifically, historical feature vectors of each collection point on the transmission line and the health status of each collection point based on expert experience annotation are obtained as training data. The data is divided into training and test sets. The training set is pre-input into the constructed random forest regression model. The parameters of the model are optimized through multiple iterations to obtain a well-trained random forest regression model. The model trained on the test set is then evaluated, and the predictive performance index of the model is calculated.

[0084] Furthermore, once multi-dimensional health indicators such as structural health, electrical health, and environmental health are adopted, the historical feature vectors of each collection point of the transmission line and the multi-dimensional health of each collection point based on expert experience annotation are obtained as training data during the training process to complete the training.

[0085] Finally, the trained random forest regression model was used to complete the health assessment results for each data collection point.

[0086] Specifically, the feature vectors of each currently acquired collection point are input into the trained random forest regression model to obtain the health assessment results of each collection point, including the health value and status level.

[0087] S4. Generate a health alert based on the health assessment results of each collection point.

[0088] Specifically, the health assessment results of each data collection node are compared with a preset health threshold (e.g., 60 points). A health warning is generated for data collection nodes whose health status is below the preset threshold. Correspondingly, data collection points with a poor health status can be identified as fault points, and a severe health warning is generated to promptly inform the data collection point of its location for maintenance.

[0089] The implementation principle of this embodiment is as follows: by comprehensively collecting multi-dimensional sensing data of transmission lines, preprocessing and feature engineering the data, and fully mining the effective information in the data; using a random forest regression model to evaluate the health of transmission lines; and generating early warnings by comparing with preset health levels, potential problems of transmission lines can be detected in a timely manner, thereby improving the health warning capability and ensuring the stable operation of the power system.

[0090] In one specific embodiment, to further improve the accuracy and timeliness of health assessment and enhance the monitoring and early warning capabilities for abnormal conditions in transmission lines, the method further includes:

[0091] For each time-series feature after feature engineering, a pre-built LSTM prediction model is used to predict the features of the corresponding future time period.

[0092] Taking the prediction of temperature in the next X hours using an LSTM prediction model as an example, the structure of the LSTM prediction model is shown in Table 2 below. The loss function is set to mean squared error (MSE), the optimizer to be Adam, and the preset learning rate and decay rate are used. Other time-series features can use the same LSTM prediction model structure, and adjustments are made for features involving environmental data (such as ice thickness). The cross-modal interaction features of ambient temperature gradient involve the LSTM prediction model structure layer. An attention mechanism can be added after the LSTM convolutional layer 2 to assign higher weights to environmental data (such as ambient temperature) to improve prediction accuracy.

[0093] Table 2 LSTM Prediction Model Structure

[0094]

[0095] The specific steps include initializing the input sequence, predicting the temperature for the first hour, updating the input window, predicting the temperature for the second hour, iterating to N hours, and obtaining the final predicted temperature for the next N hours.

[0096] Starting with optimizing the random forest regression model, this paper utilizes predictive feature enhancement to make the assessment more forward-looking. Specifically, the features predicted for future time periods for each time-series feature after feature engineering are fused with the feature-engineered features to obtain a new feature vector. The new feature vectors of each data collection point are then input into the retrained random forest regression model to obtain the health assessment results for each data collection point, replacing the original health assessment results. In addition to feature enhancement, feature enhancement training is also required. Specifically, the new feature vectors of each data collection point marked with health status in the historical transmission line are selected as incremental training data to retrain the random forest regression model, so that the retrained random forest regression model can be used for subsequent health assessments.

[0097] To ensure the safe and stable operation of power transmission lines and promptly detect potential anomalies, an isolated forest anomaly detection model was chosen. This model includes initializing isolated trees, recursively segmenting, and constructing the forest. Using historical data predicted by an LSTM prediction model for each dimension as input, the isolated forest anomaly detection model is trained to identify normal and abnormal features in each dimension, outputting labeled abnormal features. The trained isolated forest anomaly detection model is then used to determine whether there are anomalous feature labels for each dimension in the future time period predicted by the LSTM prediction model. Combined with deep learning algorithms, a pre-defined type of anomaly warning is triggered for feature data with anomalous labels, assisting in subsequent power transmission line maintenance.

[0098] Specifically, a transmission line anomaly early warning module can be constructed using deep learning algorithms. The model's input consists of features of various dimensions for future time periods marked with abnormal characteristics, and the output triggers anomaly early warnings of preset types. For example, for heavy-load lines with temperatures >80℃ or light-load lines with temperatures >60℃, the isolated forest model identifies data that deviates from the normal line range and triggers a conductor temperature early warning; for tower tilt angles >3°, the isolated forest model identifies data that deviates from the normal tower tilt range and generates an icing thickness early warning.

[0099] Furthermore, considering the specific allocation of computing resources, in order to more timely complete health status early warning and transmission line anomaly early warning, an edge-cloud collaborative computing architecture is adopted to deploy health status early warning and anomaly early warning for transmission lines. This includes: using edge computing centers to complete feature engineering processing of multi-dimensional sensing data, feature prediction for future time periods, and anomaly early warning using an isolated forest detection model, that is, uploading the collected data to edge computing centers around the collection point for corresponding processing and early warning; and using cloud computing centers to complete health status assessment results using a random forest regression model, that is, using edge computing centers to perform feature engineering processing and predictive features to upload to the cloud computing center to complete the health status assessment.

[0100] The implementation principle of this embodiment is as follows: By introducing an LSTM prediction model to predict features for future time periods, it is possible to grasp the changing trends of transmission lines in advance. The health status is reassessed after fusing the predicted features with the original features, making the assessment results more accurate; an isolated forest anomaly detection model is used to identify abnormal data and trigger early warnings, so as to discover potential risks to transmission lines more promptly and ensure the safe and stable operation of the power system.

[0101] In one specific embodiment, to better adapt to changes in transmission lines under different scenarios and improve the accuracy and relevance of health assessments, thereby more effectively ensuring the safe operation of transmission lines in various complex scenarios, the method further includes:

[0102] For each data collection point of the transmission line, collect the corresponding terrain data, such as geographical location and equipment data, such as tower data and conductor data.

[0103] For each data collection point in the transmission line, including environmental data, terrain data, and equipment data, determine the scene type corresponding to the current data collection point, including: normal scene and abnormal scene, including: extreme weather scene, complex terrain scene, external force damage scene, equipment aging scene, etc.

[0104] When facing normal-state scenarios for transmission line health monitoring, the health assessment methods described in the above embodiments can be directly adapted; when facing abnormal-state scenarios for transmission line health monitoring, further optimizations are made in dimensions such as feature engineering, model structure, and hyperparameter optimization, specifically including:

[0105] If the current acquisition point is determined to be an anomalous scenario, feature engineering optimization is performed on the preprocessed multidimensional sensing data of the current acquisition point to replace the previous feature engineering process, and feature vectors of each acquisition point are obtained by fusion. The feature engineering optimization process includes: obtaining image features, time-series features of various sensors, and lidar point cloud features using single-modal feature extraction methods; obtaining cross-modal interaction features based on physical rules conforming to the current anomalous scenario, cross-modal interaction features based on correlation analysis rules, and cross-modal interaction features constructed specifically for the current scenario, i.e., dedicated constructed features; for example, in extreme weather scenarios, features are constructed by incorporating atmospheric conditions. The intensity-response characteristics of image factors and equipment status are constructed, such as: cross-modal interaction characteristics characterizing the dynamic relationship between icing load and tower stability = (snow thickness × wind speed) / tower tilt; in complex terrain scenarios, terrain and equipment coupling characteristics are constructed, such as: cross-modal interaction quantifying the amplification effect of terrain on tower stability = tower tilt / terrain slope; in external force damage scenarios, spatiotemporal abrupt change characteristics are constructed, such as: cross-modal interaction characteristics characterizing the approach speed of construction machinery = moving speed of objects detected by lidar; in equipment aging scenarios, interaction terms incorporating equipment operating time and key parameters are constructed, such as: cross-modal interaction characteristics quantifying aging rate = insulator operating years × leakage current growth rate.

[0106] If the current data collection point is determined to be an anomalous scenario, a dedicated model can be trained for each anomalous scenario to complete model optimization and hyperparameter optimization. Specifically, multiple random forest regression models are constructed, and each random forest regression model is trained using the feature vectors of historical data collection points under a specific scenario. Based on the currently determined anomalous scenario, the corresponding random forest regression model is obtained, and the feature vectors of each data collection point are input into the trained random forest regression model to obtain the health assessment results for each data collection point.

[0107] Furthermore, each random forest regression model utilizes feature vectors from historical data collection points within a specific scenario for training. Different types of anomalous scenarios allow for further model structural improvements. For example, for extreme weather scenarios (blizzards, high temperatures, thunderstorms, etc.), a weighted random forest model is used, assigning higher weights to samples from extreme weather periods to enhance the model's learning of features during high-risk periods. For complex terrain scenarios (mountains, jungles, canyons, etc.), a hierarchical random forest model is used, dividing the model into sub-forests based on terrain type. Each sub-forest learns only the data corresponding to its terrain, and the final result is fused through weighted calculation (sub-forests with higher terrain complexity have higher weights). For external damage scenarios (approaching construction machinery, theft, impact, etc.), an incremental random forest model is used, employing an incremental learning structure with a fixed initial tree structure. Only a small number of trees are added for new damage samples to capture new features. For equipment aging scenarios (insulator aging, conductor fatigue, tower corrosion, etc.), a time-window random forest model is used, dividing the data into time windows. Each window trains a sub-forest to capture features from different aging stages, and the corresponding window's sub-forest is used for prediction.

[0108] The implementation principle of this embodiment is as follows: considering the characteristics of different scenarios of transmission lines, the scenario type is determined by collecting terrain data and equipment data; feature engineering optimization is performed for abnormal scenarios and a special random forest regression model is used, which can better adapt to the changes of transmission lines under different scenarios, improve the accuracy and pertinence of health assessment, and thus more effectively ensure the safe operation of transmission lines under various abnormal scenarios.

[0109] like Figure 2 As shown in the figure, this application discloses a multi-dimensional sensing panoramic monitoring system for transmission lines, specifically including:

[0110] The transmission line data acquisition module 101 is used to acquire multi-dimensional sensing data of each acquisition point of the transmission line, including: image data of the acquisition point, time series data of various sensors, lidar point cloud data of the distance between the acquisition point and obstacles around the acquisition point, and environmental data, and to perform preprocessing for each acquisition point.

[0111] The transmission line data processing module 102 is used to perform feature engineering processing on the preprocessed multidimensional sensing data of each acquisition point, and fuse them to obtain the feature vectors of each acquisition point. The feature engineering processing includes: obtaining image features, time-series features of various sensors, and lidar point cloud features using a single-modal feature extraction method; obtaining cross-modal interaction features constructed based on physical rules and cross-modal interaction features constructed based on correlation analysis rules using a cross-modal interaction feature construction method; and completing feature selection using a Gini importance selection method.

[0112] The transmission line health assessment module 103 is used to construct a random forest regression model; define the health status of the transmission line and generate health status labels; select the feature vectors of each collection point marked with health status in the historical transmission line as training data, input them into the constructed random forest regression model, and obtain the trained random forest regression model; input the feature vectors of each collection point currently acquired into the trained random forest regression model respectively, and obtain the corresponding health status assessment results of each collection point; compare the health status assessment results of each collection node with the preset health status threshold, and generate a health status warning for collection nodes that are less than the preset health status threshold.

[0113] In one specific embodiment, the system further includes:

[0114] The transmission line data prediction module 104 is used to predict the features of the corresponding future time period for each time-series feature after feature engineering processing using a pre-built LSTM prediction model.

[0115] The transmission line health assessment optimization module 105 is used to fuse the features of the future time period predicted for each time-series feature after feature engineering with the features after feature engineering to obtain a brand-new feature vector. The brand-new feature vectors of each collection point are input into the retrained random forest regression model to obtain the health assessment results of each collection point and replace the original health assessment results of each collection point. Among them, the brand-new feature vectors of each collection point marked with health in the historical transmission line are selected as incremental training data to retrain the random forest regression model.

[0116] The transmission line anomaly early warning module 106 constructs an isolated forest anomaly detection model. Taking historical data of features predicted by the LSTM prediction model in various dimensions as input, the isolated forest anomaly detection model is trained to identify normal and abnormal feature data in each dimension, and outputs labeled abnormal feature data in each dimension. The trained isolated forest anomaly detection model is used to determine whether there are abnormal feature data labels in the features of each dimension predicted by the LSTM prediction model for the corresponding future time period. Combined with deep learning algorithms, a preset type of anomaly early warning is triggered for feature data in each dimension that has abnormal data labels.

[0117] In one specific embodiment, the power transmission line data acquisition module 101 in the system is further configured to acquire terrain data and equipment data for each acquisition point of the power transmission line; and to determine the scene type corresponding to the current acquisition point for the environmental data, terrain data and equipment data of each acquisition point in the power transmission line, including: normal state scene and abnormal state scene;

[0118] The transmission line data adaptive processing module 107, upon determining that the current acquisition point is in an abnormal scenario, performs feature engineering optimization processing on the pre-processed multi-dimensional sensing data of the current acquisition point to replace feature engineering processing, and fuses to obtain feature vectors of each acquisition point; the feature engineering optimization processing includes: obtaining image features, time-series features of various sensors, and lidar point cloud features using a single-modal feature extraction method; obtaining cross-modal interaction features constructed based on physical rules conforming to the current abnormal scenario and cross-modal interaction features constructed based on correlation analysis rules using a cross-modal interaction feature construction method; and completing feature selection using a Gini importance selection method;

[0119] The transmission line adaptive health assessment module 108 constructs multiple random forest regression models. Each random forest regression model is trained using the feature vectors of historical collection points under a certain scenario. Based on the currently determined abnormal scenario, the corresponding random forest regression model is obtained, and the feature vectors of each collection point obtained at the moment are input into the trained random forest regression model to obtain the health assessment results of each collection point.

[0120] This application also discloses a computer-readable storage medium.

[0121] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the multi-dimensional sensing panoramic monitoring method for power transmission lines described above. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] This application also discloses a computer device.

[0123] Specifically, the computer device includes a memory and a processor, and the memory stores a computer program that can be loaded by the processor and executed for the aforementioned multi-dimensional perception panoramic monitoring method for power transmission lines.

[0124] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A multi-dimensional sensing panoramic monitoring method for transmission lines, characterized in that, include: For each acquisition point of the transmission line, multi-dimensional sensing data of the acquisition point is acquired, including: image data of the acquisition point, time series data of various sensors, lidar point cloud data of the distance between the acquisition point and obstacles around the acquisition point, and environmental data, and preprocessing is performed for each acquisition point; Feature engineering is performed on the preprocessed multidimensional sensing data from each acquisition point, and feature vectors from each acquisition point are obtained by fusion. The feature engineering process includes: obtaining image features, temporal features of various sensors, and point cloud features of LiDAR using a single-modal feature extraction method; obtaining cross-modal interaction features based on physical rules and cross-modal interaction features based on correlation analysis rules using a cross-modal interaction feature construction method; and completing feature selection using a Gini importance selection method. Construct a random forest regression model; define the health status of transmission lines and generate health status labels for transmission lines; select the feature vectors of each collection point marked with health status in historical transmission lines as training data, input them into the constructed random forest regression model, and obtain the trained random forest regression model; Input the feature vectors of each currently acquired collection point into the trained random forest regression model to obtain the health assessment results of each collection point; compare the health assessment results of each collection node with the preset health threshold, and generate a health warning for collection nodes that are less than the preset health threshold. Also includes: For each time-series feature after feature engineering, a pre-built LSTM prediction model is used to predict the features of the corresponding future time period. The features predicted for future time periods based on each time-series feature after feature engineering are fused together with the features after feature engineering to obtain a brand-new feature vector. The brand-new feature vectors of each currently acquired collection point are then input into the retrained random forest regression model to obtain the health assessment results of each collection point, replacing the original health assessment results of each collection point. Among these, the brand-new feature vectors of each collection point marked with health in the historical transmission line are selected as incremental training data to retrain the random forest regression model. Construct an isolated forest anomaly detection model; using historical data of features predicted by the LSTM prediction model in various dimensions as input, train the isolated forest anomaly detection model to identify normal and abnormal feature data in each dimension, and output labeled abnormal feature data in each dimension; use the trained isolated forest anomaly detection model to determine whether there are abnormal feature data labels in the features of each dimension predicted by the LSTM prediction model for the corresponding future time period, and combine deep learning algorithms to trigger a preset type of anomaly warning for feature data in each dimension with abnormal data labels; Also includes: For each data collection point along the power transmission line, terrain data and equipment data are collected and acquired. For each data collection point in the transmission line, including environmental data, terrain data, and equipment data, the scenario type corresponding to the current data collection point is determined, including: normal scenario and abnormal scenario; the abnormal scenario includes: extreme weather scenario, complex terrain scenario, external force damage scenario, and equipment aging scenario. If the current acquisition point is determined to be an anomalous scene, feature engineering optimization is performed on the preprocessed multidimensional perception data of the current acquisition point to replace feature engineering processing, and feature vectors of each acquisition point are obtained by fusion. The feature engineering optimization process includes: obtaining image features, time-series features of various sensors, and LiDAR point cloud features using a single-modal feature extraction method; obtaining cross-modal interaction features constructed based on physical rules that conform to the current anomalous scene and cross-modal interaction features constructed based on correlation analysis rules using a cross-modal interaction feature construction method; and completing feature selection using a Gini importance selection method. Multiple random forest regression models are constructed, and each random forest regression model is trained using the feature vectors of historical collection points under a certain scenario. Based on the currently determined abnormal scenario, the corresponding random forest regression model is obtained, and the feature vectors of each collection point obtained at the moment are input into the trained random forest regression model to obtain the health assessment results of each collection point.

2. The multi-dimensional sensing panoramic monitoring method for transmission lines according to claim 1, characterized in that, Also includes: The health of transmission lines is defined by a weighted assessment of multiple dimensions of indicators, including: structural health, electrical health, and environmental health. For each acquisition point of the transmission line, the status data of the corresponding multi-dimensional sensing data acquisition device is obtained, and the weights of different dimension indicators are dynamically adjusted according to the status data of the multi-dimensional sensing data acquisition device. This includes: counting the number of devices whose status data of the corresponding acquisition devices belonging to the dimensions of structural health, electrical health and environmental health are at risk of failure, and dynamically adjusting the weights of different dimension indicators according to the proportion of the number of devices at risk of failure counted in different dimensions.

3. The multi-dimensional sensing panoramic monitoring method for transmission lines according to claim 1, characterized in that, The construction of the random forest regression model includes: Initialize the random forest parameters, including: number of decision trees, maximum depth, and minimum number of samples for split nodes; Bootstrap sampling is used to extract a subset of samples with replacement from the feature vector of the original data, construct multiple decision trees, and randomly select some features when splitting each decision tree to obtain several final decision trees as training data. The training data is used to fit a random forest regression model. The input of the model is the feature vector obtained from the collection points. The prediction output is made by individual decision trees. The mean of the prediction output of all individual decision trees is calculated or the cluster mean is obtained by K-means clustering to obtain the overall prediction output of the forest and obtain the health assessment result.

4. The multi-dimensional sensing panoramic monitoring method for transmission lines according to claim 1, characterized in that, Also includes: The edge-cloud collaborative computing architecture is used to deploy and complete the health and anomaly warning of transmission lines; including: using the edge computing center to complete the feature engineering processing of multi-dimensional sensing data, feature prediction for future time periods, and anomaly warning using the isolated forest detection model. The health assessment results were obtained using a cloud computing center and a random forest regression model.

5. The multi-dimensional sensing panoramic monitoring method for transmission lines according to claim 1, characterized in that, Also includes: After feature selection using the Gini importance screening method, VIF collinearity detection is performed on the selected features to obtain the variance inflation coefficient of each selected feature, and features with variance inflation coefficients not greater than the preset variance inflation coefficient are retained.

6. A multi-dimensional sensing panoramic monitoring system for power transmission lines, characterized in that, include: The power transmission line data acquisition module is used to acquire multi-dimensional sensing data for each acquisition point of the power transmission line, including: image data of the acquisition point, time series data of various sensors, lidar point cloud data of the distance between the acquisition point and obstacles around the acquisition point, and environmental data, and to perform preprocessing for each acquisition point; The transmission line data processing module is used to perform feature engineering processing on the pre-processed multi-dimensional sensing data of each acquisition point, and fuse them to obtain the feature vectors of each acquisition point. The feature engineering processing includes: obtaining image features, time-series features of various sensors, and lidar point cloud features using single-modal feature extraction methods; obtaining cross-modal interaction features constructed based on physical rules and cross-modal interaction features constructed based on correlation analysis rules using cross-modal interaction feature construction methods; and completing feature selection using Gini importance filtering methods. The power transmission line health assessment module is used to construct a random forest regression model; define the health status of the power transmission line and generate health status labels; select feature vectors of each collection point marked with health status in the historical power transmission line as training data, input them into the constructed random forest regression model, and obtain the trained random forest regression model; input the feature vectors of each currently acquired collection point into the trained random forest regression model respectively, and obtain the corresponding health status assessment results of each collection point; compare the health status assessment results of each collection node with the preset health status threshold, and generate a health status warning for collection nodes that are lower than the preset health status threshold; The transmission line data prediction module is used to predict the features of the corresponding future time period for each time-series feature after feature engineering processing using a pre-built LSTM prediction model. The transmission line health assessment optimization module is used to fuse the features predicted for future time periods for each time-series feature after feature engineering with the feature-engineered features to obtain a new feature vector. The new feature vectors of each currently acquired collection point are then input into the retrained random forest regression model to obtain the corresponding health assessment results for each collection point, replacing the original health assessment results. In particular, the new feature vectors of each collection point marked with health status in the historical transmission line are selected as incremental training data to retrain the random forest regression model. The transmission line anomaly early warning module is used to construct an isolated forest anomaly detection model. It takes historical data of features predicted by the LSTM prediction model as input, trains the isolated forest anomaly detection model to identify normal and abnormal feature data in each dimension, and outputs labeled abnormal feature data in each dimension. It uses the trained isolated forest anomaly detection model to determine whether there are abnormal feature data labels in the features of each dimension predicted by the LSTM prediction model for the corresponding future time period, and combines deep learning algorithms to trigger a preset type of anomaly early warning for feature data in each dimension with abnormal data labels. The transmission line data acquisition module is also used to acquire terrain data and equipment data for each acquisition point of the transmission line; and to determine the scene type corresponding to the current acquisition point for the environmental data, terrain data and equipment data of each acquisition point in the transmission line, including: normal state scene and abnormal state scene; The transmission line data adaptive processing module is used to determine that the current acquisition point is an anomalous scenario. It performs feature engineering optimization on the pre-processed multi-dimensional sensing data of the current acquisition point to replace the original feature engineering process, and fuses the feature vectors of each acquisition point. The feature engineering optimization process includes: acquiring image features, time-series features from various sensors, and lidar point cloud features using single-modal feature extraction; acquiring cross-modal interactive features based on physical rules conforming to the current anomalous scenario and cross-modal interactive features based on correlation analysis rules using cross-modal interactive feature construction methods; and completing feature selection using Gini importance filtering. The adaptive health assessment module for transmission lines is used to construct multiple random forest regression models. Each random forest regression model is trained using the feature vectors of historical collection points under a certain scenario. Based on the currently determined abnormal scenario, the corresponding random forest regression model is obtained, and the feature vectors of each collection point obtained at the moment are input into the trained random forest regression model to obtain the health assessment results of each collection point.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method as described in any one of claims 1 to 5.

8. A computer device, characterized in that, The computer device includes a memory, a processor, and a program stored in and executable on the memory, the program being executed by the processor to implement the steps of the method as described in any one of claims 1 to 5.

Images