Power distribution tower pole deformation monitoring system based on visual inspection

Abstract

The invention discloses a power distribution tower pole deformation monitoring system based on visual inspection, relates to the technical field of power transmission and transformation monitoring, and specifically comprises debris flow susceptibility monitoring, meteorological disaster monitoring, landslide disaster monitoring, monitoring equipment access, edge algorithm model establishment and Beidou data processing. The tower and the surrounding environment monitoring analysis early warning system are deployed to complete access and management of the monitoring equipment and the intelligent terminal, the intelligent terminal, the camera and other sensor equipment are deployed on the tower to realize acquisition of health data and image data of the tower, the Beidou module is added to the intelligent terminal to realize Beidou communication capability, and the tower health data and image data are acquired. A monitoring algorithm model and a data and image compression algorithm model are deployed on the intelligent terminal, analysis and compression of data and images are achieved, the data and the images are uploaded to a monitoring center through a Beidou module, and monitoring and early warning of the remote tower and the surrounding environment of the remote tower are achieved.

Description

Technical Field

[0001] This invention relates to the field of power transmission and transformation monitoring, specifically to a visual inspection-based power distribution tower deformation monitoring system. Background Technology

[0002] Due to the large number and wide distribution of power poles, the monitoring of power poles and their surrounding environment presents significant challenges in data collection, transmission, and analysis. Furthermore, the added value of power poles as infrastructure cannot be fully utilized. This is particularly true in mountainous regions where natural disasters such as earthquakes, rainfall, and snowmelt, or human engineering activities like excavation of slope toes, loading of upper slopes, and blasting, can lead to flash floods, landslides, and mudslides. Short-duration heavy rainfall can cause flash floods, resulting in multiple deaths and damage to power poles. Early detection and timely response can significantly reduce losses.

[0003] However, the existing monitoring methods for power distribution towers only involve deploying smart terminals, cameras, and other sensor devices on the towers to collect tower health data and image data. However, there is no monitoring for disasters such as mudslides, landslides, and floods. This single monitoring method reduces the early warning capability for the towers. Summary of the Invention

[0004] This application proposes a visual inspection-based power distribution tower deformation monitoring system, which has the following advantages: it provides early warning of tower deformation by monitoring the surrounding environment; it deploys monitoring algorithm models and data and image compression algorithm models on intelligent terminals to achieve data and image analysis and compression; and it uploads data to the monitoring center via a Beidou module to achieve monitoring and early warning of towers in remote areas and their surrounding environment, thus solving the technical problems mentioned in the background.

[0005] To achieve the above objectives, this application adopts the following technical solution: a visual detection-based power distribution tower deformation monitoring system, including monitoring of debris flow susceptibility, meteorological disasters, landslide disasters, access to monitoring equipment, establishment of edge algorithm models, and processing of BeiDou data;

[0006] Specifically, the monitoring of debris flow susceptibility refers to the evaluation of debris flow susceptibility using machine learning algorithms. The monitoring of debris flow susceptibility consists of four processes: data preprocessing, construction of a susceptibility assessment model, partitioning debris flow susceptibility based on model results, and evaluation of model performance.

[0007] The monitoring of meteorological disasters specifically refers to the study of interannual variations in precipitation data;

[0008] The monitoring of landslide disasters specifically refers to using a YOLO V5 deep learning network to train a model to identify landslides;

[0009] The access to the monitoring equipment specifically refers to building a million-level monitoring equipment access capability, monitoring data visualization analysis and early warning capability, and remote operation and maintenance management capability based on a cloud-edge-device integrated architecture.

[0010] The establishment of the edge algorithm model specifically refers to the management personnel remotely updating the algorithm model according to the monitoring needs of different regions and periods, and applying the research and training results of the monitoring center's algorithm model to the edge side;

[0011] The processing of BeiDou data specifically refers to the use of data compression algorithms to comprehensively improve the efficiency of data transmission via satellite communication for monitoring data.

[0012] Preferably, the monitoring of debris flow susceptibility includes the selection and classification of debris flow influencing factors, the analysis of the importance of influencing factors, and the establishment of an influencing factor model.

[0013] Preferably, the selection and classification of debris flow influencing factors adopts the natural discontinuity method to determine the optimal classification interval, and the specific operation is as follows: 11 influencing factors are selected, including elevation, slope, aspect, topography, land use type, population density, stratigraphic lithology, geological structure, average annual rainfall, distance to river, and vegetation cover. For a set of data X={x1, x2, ..., x...} n} (x1<=x2<=…<=x n Let its variance be SDAM. Assume the data set is divided into M classes, and the classification interval Y for any class is known. i ={x1, y1, y2,…, y M-1 , x n} (x1) <y1<y2<…<y M-1 <x n ), calculate the variance for each class, and the range of each class is (x1, y1), (y1, y2), (y2, y3), ..., (y M-2 ,y M-1 ), (y M-1 ,x n The variances of all classes are summed, and this sum is denoted as SDCM. This is applied to different classification intervals Y. i Each has a corresponding SDCMi;

[0014] The optimal classification interval is the one that corresponds to the smallest SDCMi.

[0015] The calculation formulas for SDAM and SDCM are as follows:

[0016]

[0017] In the formula, N j This represents the number of samples in the j-th class.

[0018] The smallest SDCM value corresponds to the largest variance-to-fit goodness-of-fit (GVF) value. The GVF calculation formula is as follows:

[0019]

[0020] In the formula, the value of GVF ranges from [0,1].

[0021] Preferably, the analysis of the importance of the influencing factors specifically refers to: (using information on geological background and debris flow development characteristics to qualitatively select debris flow influencing factors);

[0022] The importance of the selected impact factors was assessed using quantitative methods.

[0023] The most suitable influencing factor is selected by combining qualitative and quantitative methods.

[0024] The specific steps for analyzing the importance of the influencing factors are as follows: Information entropy, conditional entropy, and information gain are calculated and used to calculate the information gain ratio.

[0025] Entropy is used to measure the uncertainty of a random variable. The greater the uncertainty, the higher the entropy. For a set of data Y={y1,y2,…,y...}, ... n The formula for calculating the entropy of Y is as follows:

[0026]

[0027] In the formula, p(y i ) represents y i The probability of occurrence;

[0028] Conditional entropy is the entropy of a random variable under specified conditions. By definition, conditional entropy is less than or equal to total entropy. This is because adding completely irrelevant information does not change the uncertainty of a random variable, while adding relevant information reduces its uncertainty. The formula for calculating conditional entropy is as follows:

[0029]

[0030] In the formula, H(Y|X) represents the entropy of Y given that X occurs, p(x,y) represents the probability that x and y occur simultaneously, and p(y|x) represents the probability that y occurs given that x occurs.

[0031] The formula for calculating information gain is as follows:

[0032] ,

[0033] The formula for calculating information gain ratio is as follows:

[0034] ;

[0035] The information gain ratio is obtained by calculating the numerator and denominator separately and then dividing the result.

[0036] Preferably, the establishment of the influence factor model specifically refers to: performing weighted summation on multiple inputs received by the neuron connection, comparing the result of the weighted summation with a threshold, and finally obtaining the model output through an activation function;

[0037] This can be expressed using a mathematical formula as follows:

[0038]

[0039] In the formula, y k The output is represented by f(x), where f(x) represents the activation function and x is the value of x. i The input factor, w ki b represents the weight. k Indicates bias.

[0040] Preferably, the monitoring of meteorological disasters specifically refers to: using ensemble empirical mode decomposition based on wavelet analysis to divide time series precipitation data into a limited number of hidden oscillation modes and a trend component, and then using ensemble empirical mode decomposition-Markov model to predict precipitation using sparse samples.

[0041] Preferably, the monitoring of the landslide disaster specifically refers to: using the YOLOv5 algorithm to transform the target detection task into a regression problem for real-time target detection and localization.

[0042] Preferably, the access of the monitoring device includes cloud-edge-device integration, device templates, device examples, and multi-protocol support.

[0043] Preferably, the establishment of the edge algorithm model includes model management, model publishing, and model updating.

[0044] Preferably, the processing of the BeiDou data includes monitoring data acquisition and preprocessing, adaptive selection of compression algorithms, and BeiDou short message transmission.

[0045] The present invention has the following beneficial effects:

[0046] 1. This invention deploys a monitoring, analysis, and early warning system for power poles and their surrounding environment. This system enables the access and management of monitoring equipment and smart terminals. Smart terminals, cameras, and other sensor devices are deployed on the power poles to collect health data and image data. The smart terminals incorporate a BeiDou module for communication capabilities. Monitoring algorithm models and data / image compression algorithm models are deployed on the smart terminals to analyze and compress data and images. This data is then uploaded to the monitoring center via the BeiDou module, enabling monitoring and early warning of remote power poles and their surrounding environment.

[0047] 2. This invention collects monitoring data on the physical and environmental indicators of power poles through edge devices, including pole displacement data, tilt data, acceleration data, temperature, wind speed, wind direction data, and surrounding environmental image data. It uses algorithm models to reason and analyze the data on changes in the power pole, geographical environment, and climate environment, and conducts real-time early warning of power pole tilt and natural disasters such as landslides, mudslides, and fires in the surrounding area, realizing a deep perception of the health status of the power pole and its surrounding environment. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the neuron structure in this invention;

[0049] Figure 2 This is a schematic diagram illustrating a specific prediction example for a certain site within the structure of this invention;

[0050] Figure 3 This is a schematic diagram illustrating the detailed iterative process of the EEMD algorithm in this invention. Detailed Implementation

[0051] The technical solution of the present invention will be clearly and completely described below with reference to preferred embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] like Figures 1 to 3 A visual inspection-based power distribution tower deformation monitoring system includes monitoring debris flow susceptibility, meteorological disasters, landslide disasters, access to monitoring equipment, establishment of edge algorithm models, and processing of BeiDou data.

[0053] Specifically, debris flow susceptibility monitoring refers to the evaluation of debris flow susceptibility using machine learning algorithms. Debris flow susceptibility monitoring consists of four processes: data preprocessing, construction of susceptibility assessment models, partitioning debris flow susceptibility based on model results, and evaluation of model performance.

[0054] Meteorological disaster monitoring specifically refers to the study of interannual variations in precipitation data;

[0055] Landslide disaster monitoring specifically refers to using a YOLO V5 deep learning network to train a model to identify landslides;

[0056] The access of monitoring equipment specifically refers to building a cloud-edge-device integrated architecture to enable access to millions of monitoring devices, visualized analysis and early warning of monitoring data, and remote operation and maintenance management capabilities.

[0057] The establishment of edge algorithm models specifically refers to the management personnel remotely updating the algorithm models according to the monitoring needs of different regions and periods, and applying the research and training results of the monitoring center's algorithm models to the edge side;

[0058] The processing of BeiDou data specifically refers to the use of data compression algorithms to comprehensively improve the efficiency of data transmission via satellite communication for monitoring data.

[0059] Monitoring debris flow susceptibility includes selecting and classifying debris flow influencing factors, analyzing the importance of influencing factors, and establishing influencing factor models.

[0060] The selection and classification of debris flow influencing factors adopted the natural discontinuity method to determine the optimal classification interval. Specifically, 11 influencing factors were selected, including elevation, slope, aspect, topography, land use type, population density, stratigraphic lithology, geological structure, average annual rainfall, distance to river, and vegetation cover. For a set of data X={x1, x2,…,x…} n} (x1<=x2<=…<=x n Let its variance be SDAM. Assume the data set is divided into M classes, and the classification interval Y for any class is known. i ={x1, y1, y2,…, y M-1 , x n} (x1) <y1<y2<…<y M-1 <x n ), calculate the variance for each class, and the range of each class is (x1, y1), (y1, y2), (y2, y3), ..., (y M-2 ,y M-1 ), (y M-1 ,x n The sum of these variances is the sum of the variances of all classes, denoted as SDCM. For different classification intervals Yi, there is a corresponding SDCMi.

[0061] The optimal classification interval is the one that corresponds to the smallest SDCMi.

[0062] The calculation formulas for SDAM and SDCM are as follows:

[0063]

[0064] In the formula, N j This represents the number of samples in the j-th class.

[0065] The smallest SDCM value corresponds to the largest variance-to-fit goodness-of-fit (GVF) value. The GVF calculation formula is as follows:

[0066]

[0067] In the formula, the value of GVF ranges from [0,1].

[0068] The analysis of the importance of influencing factors specifically refers to: calculating the values ​​of information entropy, conditional entropy, and information gain, which are then used to calculate the information gain ratio.

[0069] For a set of data Y={y1,y2,…,y n The formula for calculating the entropy of Y is as follows:

[0070]

[0071] In the formula, p(y i ) represents y i The probability of occurrence;

[0072] The formula for calculating conditional entropy is as follows:

[0073]

[0074] In the formula, H(Y|X) represents the entropy of Y given that X occurs, p(x,y) represents the probability that x and y occur simultaneously, and p(y|x) represents the probability that y occurs given that x occurs.

[0075] The formula for calculating information gain is as follows:

[0076] ,

[0077] The formula for calculating information gain ratio is as follows:

[0078] ;

[0079] The information gain ratio is obtained by calculating the numerator and denominator separately and then dividing the results. The analysis of the importance of influencing factors involves qualitatively selecting these factors by analyzing geological background and debris flow development characteristics; evaluating the importance of the selected factors using quantitative methods; and finally, combining qualitative and quantitative methods to select the most suitable influencing factor.

[0080] The establishment of the influence factor model specifically refers to: the multiple inputs received by the neuron are weighted and summed, the result of the weighted sum is compared with the threshold, and finally the output of the model is obtained through the activation function;

[0081] This can be expressed using a mathematical formula as follows:

[0082]

[0083] In the formula, y k The output is represented by f(x), where f(x) represents the activation function and x is the value of x. i The input factor, w ki b represents the weight. k Indicates bias;

[0084] like Figure 1 As shown, a multilayer feedforward neural network receives input from debris flow influencing factors, calculates the model's output through hidden layers, and compares the output with the actual results to obtain information such as model accuracy. Based on the model's output, a loss function can be constructed. The loss function is used to evaluate the model error. After calculating the model's loss, algorithms such as gradient descent are used to back-calculate the model's weight parameters for the next iteration. The model error is then recalculated based on the updated model parameters. This iterative process is repeated multiple times until the result meets the requirements. This is the basic solution approach for neural networks.

[0085] Meteorological disaster monitoring specifically refers to: using wavelet analysis-based ensemble empirical mode decomposition (EMD) to divide time-series precipitation data into a finite number of hidden oscillation patterns and a trend component; then, using an EMD-Markov model (DMarkov model for short) to predict precipitation with sparse samples. The wavelet analysis-based EMD algorithm is abbreviated as W-EEMD. Abnormal precipitation is detected as a potential oscillation pattern, rather than noise data. Since precipitation, as a product of atmospheric circulation, is usually a periodic phenomenon, wavelet analysis can obtain the number of potential oscillation patterns. This is because wavelet analysis can more easily detect the periodicity of data.

[0086] like Figure 2 As shown, suppose we are predicting precipitation for a region, which includes several stations. For a given station (e.g., station i), the n-year precipitation data from the previous year can be provided as follows: The goal is to predict next year's rainfall at all stations. For each site, the W-EEMD method is used to transform the nonlinear time series data X. iThe data is decomposed into r IMFs and one residual, so that a hidden oscillation is treated as an IMF rather than noise data. Here, the number of IMFs r is obtained by performing wavelet analysis on historical data Xi. Then, a Markov model is used to predict each pattern one by one. Finally, the site... The precipitation forecast is composed of all r+1 predictions from the Markov model. The predictions from all stations constitute the precipitation forecast for the region next year. However, there are some differences in the specific parameter settings for the Markov model calculation method. State levels are defined. For each IMF, this paper classifies them into 5 state levels, setting the state level classification criteria α to 0.8 and β to 0.4. Five state transition matrices are calculated, with k set to 1~5, and transition probability matrices m1~m5, corresponding to state transition time spans of 1~5 years. This is because historical states closer to the prediction time have a greater impact on the prediction results; therefore, this paper selects the 5 years closest to the prediction year for probability calculation. Furthermore, to balance the influence of the other 4 predicted states (except for the state with the largest pi), σ is set to 2.

[0087] like Figure 3 As shown, the EEMD algorithm is a recursive iterative process. The specific algorithm flow is as follows:

[0088] S1. Add random white noise with a certain signal-to-noise ratio to the rainy season precipitation time series Xi to obtain the rainy season precipitation time series X' with added noise;

[0089] S2. Find the maximum point of X' and fit the upper envelope X'. ma Find the minimum point of X' and fit the lower envelope X'. min ;

[0090] S3, Find the upper envelope line X' ma And the lower envelope X' min The mean value of the upper and lower envelopes is m1;

[0091] S4. Subtract the noise-added rainy season precipitation time series X' from the mean of the upper and lower envelopes m1 to obtain the remaining signal d1; replace X' with the remaining signal d1 to obtain a new noise-added rainy season precipitation time series.

[0092] S5. Repeat steps S3 to S4 until the set number of iterations M1 is met, and d1' is obtained.

[0093] S6. Repeat steps S1 to S5 until the set number of iterations M2 is met, and obtain M2 d1's. Take the average of the M2 d1's to obtain the first feature term IMF1.

[0094] S7. To obtain more IMFs, subtract the time series Xi from IMF1 to obtain the first-order residual r1. Use r1 as the new rainy season precipitation time series and repeat steps S1 to S6 until the above rainy season precipitation time decomposition series is obtained: IMF1, IMF2, ..., IMFn, and the nth-order residual rn is the trend term.

[0095] Overall, for the first iteration, the first potential oscillation mode is extracted as IMF1, and the rest are the initial residuals, which will become the input for the next iteration. Therefore, the potential oscillation mode extracted in the first iteration is IMFi, and the remaining part is regarded as the residual.

[0096] Landslide disaster monitoring specifically refers to the use of the YOLOv5 algorithm. By transforming the target detection task into a regression problem, it enables real-time target detection and localization. YOLOv5 achieves faster detection speeds through optimized network structure and inference processes, making it suitable for processing real-time landslide data. This is significant for the automatic identification and early warning of landslides, helping to reduce the need for manual intervention and improve response speed and processing efficiency. It provides a simple and intuitive API interface, facilitating model training and tuning for researchers. Furthermore, YOLOv5 supports various data augmentation and model ensemble techniques, such as MixUp and Ensemble, which can further improve model performance. YOLOv5 is a version of the YOLO target detection network model that employs a new network structure called Linear Convolutional Layer (LinConV) to replace traditional convolutional layers. This new network structure improves network performance and efficiency. YOLO V5 mainly consists of three parts: feature extraction, feature fusion, and prediction. Feature extraction: YOLO V5 uses multiple convolutional layers of different sizes to extract image features, which can better detect targets of different sizes. Feature fusion: It uses a feature fusion method of multiple feature maps of different sizes to obtain the final feature map, which can improve the accuracy of target detection. Prediction: It uses a method similar to a multi-scale sliding window to predict the position and category of the target, which can better detect targets of different sizes.

[0097] The access to monitoring devices includes cloud-edge-device integration, device templates, device examples, and multi-protocol support.

[0098] 1) Cloud-edge-device integration

[0099] Cloud: The cloud can be deployed in a single machine or in a cluster. As the core of management, the cloud is responsible for the management of edge nodes, AI models, applications and devices, as well as user permissions, and is also responsible for interaction and integration with third parties.

[0100] Cloud-edge collaboration component: The cloud-edge collaboration component supports single-node mode to cope with different edge environments. In single-machine mode, it can run on a general Linux kernel. The cloud-edge collaboration component is responsible for the synchronization, reporting, monitoring, and remote operation and maintenance of edge models, data, and status.

[0101] Edge side: Responsible for accessing and controlling terminal devices, collecting data and formulating data rules, executing model inference, and reporting inference results. It also uses containers to run various business applications.

[0102] End devices: These can be cameras, microphones, environmental sensors, Bluetooth devices, or other common IoT devices, etc.

[0103] Peripheral systems: By introducing systems such as logging, monitoring, and security, the system assists in the stable, reliable, and secure operation of the hardware and software.

[0104] 2) Equipment template

[0105] In actual monitoring, the number of monitoring devices can be enormous. At the monitoring center, a unified template can be defined for each type of device; for example, similar devices like cameras, stress sensors, and wind speed sensors can all be defined under one template. Creating device instances using these templates reduces manual operation costs and prevents errors. By virtualizing physical terminal devices at the monitoring center, unified visual management of the devices is achieved, reducing the complexity of maintaining and managing a massive number of terminal devices.

[0106] 3) Equipment Examples

[0107] Based on the equipment templates, managers can selectively and specifically configure the monitoring data to be collected for different towers and their surrounding environment, avoiding [unclear - possibly related to monitoring needs]. Simultaneously, managers can monitor the equipment's health status and collected monitoring data in real time through a visual interface, and can issue alarms for abnormal situations.

[0108] 4) Multi-protocol support

[0109] It should support access protocols for cameras and various types of sensor devices, such as 4G / 5G, BeiDou, Modbus, OPC-UA, RS485, etc.

[0110] The establishment of edge computing models includes model management, model deployment, and model updates.

[0111] 1) Model Management

[0112] Model management includes storage management and operation management of models. Model storage management mainly relies on model repositories, through which algorithm models are classified and stored. Administrators can select appropriate algorithms from the model repository according to the application scenario and version of the algorithm and distribute them. Model operation management provides visual monitoring capabilities for algorithm models, monitoring their running status, running logs, and computation status.

[0113] 2) Model Release

[0114] Model deployment enables algorithms to be deployed to designated edge sides according to the administrator's usage needs, realizing targeted or batch deployment of visual algorithm models.

[0115] 3) Model Update

[0116] As the algorithm model accumulates daily data and is continuously optimized and upgraded through training, it is inevitable that the edge-side algorithm will need to be updated and iterated. The model update capability can be achieved by completing the update and upgrade operation of the edge-side algorithm model at the monitoring center, and the model version can also be selected according to actual needs.

[0117] Functional value

[0118] Edge algorithm model management capabilities enable managers to remotely update algorithm models according to monitoring needs in different regions and at different times, thereby achieving targeted monitoring effects. They can also easily apply the research and training results of the monitoring center's algorithm models to the edge side.

[0119] The processing of BeiDou data includes monitoring data acquisition and preprocessing, adaptive selection of compression algorithms, and BeiDou short message transmission.

[0120] 1) Monitoring data collection and preprocessing

[0121] Design a data acquisition probe to collect data from power data sensor devices in remote areas, including power poles and their surrounding environment. Clean and aggregate the data to filter out key data from the sensor devices.

[0122] 2) Adaptive selection of compression algorithm

[0123] For each compressed message type and characteristic, a rule-based approach is used to select a combination of fine-grained lossless compression algorithms. For different message types, the appropriate compression encoding method is matched according to the message data structure characteristics.

[0124] 3) BeiDou short message transmission

[0125] The compressed data is synchronized from the edge to the monitoring center via the Beidou user terminal. At the same time, the data is stored at the edge in the form of a cache and logs are generated to facilitate later debugging.

[0126] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention. The invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A visual inspection-based power distribution tower deformation monitoring system, characterized in that: This includes monitoring debris flow susceptibility, meteorological disasters, landslides, access to monitoring equipment, establishment of edge computing models, and processing of BeiDou data; Specifically, the monitoring of debris flow susceptibility refers to the evaluation of debris flow susceptibility using machine learning algorithms. The monitoring of debris flow susceptibility consists of four processes: data preprocessing, construction of a susceptibility assessment model, partitioning debris flow susceptibility based on model results, and evaluation of model performance. The monitoring of meteorological disasters specifically refers to the study of interannual variations in precipitation data; The monitoring of landslide disasters specifically refers to using a YOLO V5 deep learning network to train a model to identify landslides; The access to the monitoring equipment specifically refers to building a million-level monitoring equipment access capability, monitoring data visualization analysis and early warning capability, and remote operation and maintenance management capability based on a cloud-edge-device integrated architecture. The establishment of the edge algorithm model specifically refers to the management personnel remotely updating the algorithm model according to the monitoring needs of different regions and periods, and applying the research and training results of the monitoring center's algorithm model to the edge side; The processing of BeiDou data specifically refers to using data compression algorithms to comprehensively improve the efficiency of data transmission via satellite communication for monitoring data.

2. The power distribution tower deformation monitoring system based on vision detection according to claim 1, characterized in that: The monitoring of debris flow susceptibility includes the selection and classification of debris flow influencing factors, the analysis of the importance of influencing factors, and the establishment of an influencing factor model.

3. The distribution tower deformation monitoring system based on vision detection according to claim 2, characterized in that: The selection and classification of debris flow influencing factors adopts the natural discontinuity method to determine the optimal classification interval, and the specific operation is as follows: for a set of data X={x1, x2, ..., x... n } (x1<=x2<=…<=x n Let the variance be SDAM. Assume the data set is divided into M classes, and the classification interval for any given class is known as Yi = {x1, y1, y2, ..., y...}. M-1 , x n } (x1) <y1<y2<…<y M-1 <x n ), calculate the variance for each class, and the range of each class is (x1, y1), (y1, y2), (y2, y3), ..., (y M-2 ,y M-1 ), (y M- 1,x n The sum of these variances is the sum of the variances of all classes, denoted as SDCM. For different classification intervals Yi, there is a corresponding SDCMi. The optimal classification interval is the one that corresponds to the smallest SDCMi. The calculation formulas for SDAM and SDCM are as follows: In the formula, Nj represents the number of samples in the j-th class. The smallest SDCM value corresponds to the largest variance-to-fit goodness-of-fit (GVF) value. The GVF calculation formula is as follows: In the formula, the value of GVF ranges from [0,1].

4. The distribution tower deformation monitoring system based on vision detection according to claim 2, characterized in that: The analysis of the importance of the influencing factors specifically refers to: (qualitatively selecting debris flow influencing factors by analyzing information on geological background and debris flow development characteristics); The importance of the selected impact factors was assessed using quantitative methods. The most suitable influencing factor is selected by combining qualitative and quantitative methods. The specific steps for analyzing the importance of the influencing factors are as follows: Information entropy, conditional entropy, and information gain are calculated and used to calculate the information gain ratio. For a set of data Y={y1,y2,…,y n The formula for calculating the entropy of Y is as follows: In the formula, p(y i ) represents y i The probability of occurrence; The formula for calculating conditional entropy is as follows: In the formula, H(Y|X) represents the entropy of Y given that X occurs, p(x,y) represents the probability that x and y occur simultaneously, and p(y|x) represents the probability that y occurs given that x occurs. The formula for calculating information gain is as follows: ， The formula for calculating information gain ratio is as follows: ； The information gain ratio is obtained by calculating the numerator and denominator separately and then dividing the result.

5. The distribution tower deformation monitoring system based on vision detection according to claim 2, characterized in that: The establishment of the influence factor model specifically refers to: weighting and summing multiple inputs received by the neuron connection, comparing the weighted sum with a threshold, and finally obtaining the model output through an activation function; This can be expressed using a mathematical formula as follows: In the formula, y k The output is represented by f(x), where f(x) represents the activation function and x is the value of x. i The input factor, w ki b represents the weight. k Indicates bias.

6. The power distribution tower deformation monitoring system based on vision detection according to claim 1, characterized in that: The monitoring of meteorological disasters specifically refers to: using ensemble empirical mode decomposition based on wavelet analysis to divide time series precipitation data into a limited number of hidden oscillation modes and a trend component, and then using ensemble empirical mode decomposition-Markov model to predict precipitation using sparse samples.

7. The distribution tower deformation monitoring system based on vision detection according to claim 1, characterized in that: The monitoring of landslide disasters specifically refers to the use of the YOLOv5 algorithm to transform the target detection task into a regression problem for real-time target detection and localization.

8. The distribution tower deformation monitoring system based on vision detection according to claim 1, characterized in that: The access to the monitoring equipment includes cloud-edge-device integration, device templates, device examples, and multi-protocol support.

9. The distribution tower deformation monitoring system based on vision detection according to claim 1, characterized in that: The establishment of the edge algorithm model includes model management, model publishing, and model updating.

10. The distribution tower deformation monitoring system based on vision detection according to claim 1, characterized in that: The processing of BeiDou data includes monitoring data acquisition and preprocessing, adaptive selection of compression algorithms, and BeiDou short message transmission.

Images