Intelligent monitoring method, device and cloud server for power overhead lines based on self-monitoring

Abstract

The present application provides a method, device and cloud server for intelligent monitoring of power overhead lines based on self-supervision, the method comprising: using a front-end acquisition device to collect real-time multi-dimensional image data of power overhead lines; performing self-supervision pre-training on a seed model transmitted by a cloud server according to historical multi-dimensional image data of power overhead lines to obtain a trained proprietary model; using a proprietary model to identify hidden dangers in real-time multi-dimensional image data, obtaining hidden danger identification results, and transmitting the hidden danger identification results to a visualization monitoring device for visualization display. It is capable of achieving real-time monitoring and intelligent identification of hidden dangers in power overhead lines, and by intelligently adjusting model parameters to adapt to complex and changeable natural environments, reducing false alarms and missed alarms, and ensuring the safe operation of power overhead lines. Through the training and application of proprietary models, while ensuring high-quality and efficient completion of real-time monitoring tasks, the individual accuracy of the proprietary models is continuously optimized.

Description

Technical Field

[0001] The present application belongs to the technical field of power system monitoring, and in particular, relates to an intelligent monitoring method, device and cloud server for power overhead lines based on self-monitoring. Background Art

[0002] Overhead power lines are the core components of the power transmission system, and their safe and stable operation is crucial to the overall reliability of the power system. However, since overhead lines are exposed to the outdoors for a long time, they are greatly affected by natural environmental factors, such as extreme weather such as wildfires, rainstorms, heavy snow, or fallen trees and bird nesting, which can cause different degrees of hidden dangers to overhead lines. If these hidden dangers are not discovered and handled in time, they may lead to serious accidents such as line breakage and short circuits, which in turn affect the continuity and safety of power supply.

[0003] Traditional monitoring of hidden dangers in overhead power lines mainly relies on manual inspections and fixed-point sensor monitoring. Although manual inspections can intuitively detect hidden dangers, they have the following problems: First, they are inefficient. Manual inspections rely on staff to regularly inspect the lines, which is time-consuming and difficult to detect problems in real time; second, they are highly subjective: different personnel have different inspection standards and experience, resulting in inconsistent accuracy in hidden danger detection; third, they are limited in scope: manual inspections are difficult to cover a wide range of lines, especially in complex terrain or bad weather conditions, and there are monitoring blind spots.

[0004] In recent years, with the development of image processing and artificial intelligence technologies, the detection technology of hidden dangers of power overhead lines based on image recognition has begun to attract attention. This type of technology collects image data around the line through cameras installed on drones, inspection robots or fixed positions, and uses computer vision algorithms to automatically identify potential hidden dangers in the image. Although these methods have improved the level of automation of detection to a certain extent, there are still new challenges. For example, under different ambient lighting, weather and seasonal conditions, the clarity and contrast of the image will be affected, which in turn affects the accuracy of recognition. At the same time, the existing image recognition models mostly rely on a large amount of pre-labeled data for supervised learning. Under environmental changes or new scenarios, the recognition effect of the model may drop significantly, and it lacks adaptive capabilities. In addition, with the expansion of the monitoring area and the increase in monitoring frequency, the amount of image data that the system needs to process also increases. How to efficiently process and analyze massive data has become a major technical problem. Summary of the invention

[0005] The main purpose of the embodiments of the present invention is to provide an intelligent monitoring method, device and cloud server for power overhead lines based on self-supervision, which can adapt to the complex and changeable natural environment by intelligently adjusting model parameters, thereby reducing false alarms and missed alarms, and ensuring the safe operation of power overhead lines. Through the training and application of proprietary models, while ensuring high-quality and efficient completion of real-time monitoring tasks, the individual accuracy of the proprietary models is continuously optimized, thereby solving the problem in the prior art that power overhead line detection is easily affected by environmental conditions and has a limited level of automation.

[0006] In the first aspect, a method for intelligent monitoring of power overhead lines based on self-supervision is provided, the method comprising: using a front-end acquisition device to collect real-time multi-dimensional image data of the power overhead lines, the front-end acquisition device comprising a multi-spectral camera and an infrared camera, the multi-spectral camera and the infrared camera supporting timed shooting and event-triggered shooting, and adjusting the shooting frequency according to environmental conditions; performing self-supervision pre-training on a seed model transmitted by a cloud server according to historical multi-dimensional image data of the power overhead lines, and generating a proprietary model adapted to specific environmental characteristics, specifically comprising: receiving a seed model transmitted by a cloud server, parsing the seed model into a weight matrix, a hidden danger feature extraction path description, and a hidden danger feature extraction path description. The algorithm consists of three parts: description, pre-training task parameter configuration, and hierarchical feature consistency check; the environmental and hidden danger characteristics of the power overhead line are used to generate self-supervised training signals; hierarchical feature extraction is performed on the intrinsic structure and characteristics of the historical multi-dimensional image data itself; the extracted features and self-supervised training signals are input into the seed model, the weights and training strategies are adjusted according to the real-time feedback of the seed model, the parameters of the seed model are dynamically optimized, and a proprietary model adapted to the environmental characteristics of different locations is obtained; the proprietary model is used to identify hidden dangers in real-time multi-dimensional image data, and the hidden danger identification results are obtained, and the hidden danger identification results are transmitted to the visualization monitoring equipment for visualization display and early warning.

[0007] In another possible implementation, the seed model transmitted by the cloud server is received, and the seed model is parsed into three parts: weight matrix, hidden danger feature extraction path description, and pre-training task parameter configuration, and a hierarchical feature consistency check is performed, specifically including: comparing the feature response pattern of the weight matrix with the historical feature library of the power overhead line to match the historical high-risk hidden danger areas; calculating the statistical indicators of the output feature distribution of each level of the convolution kernel, including the mean, variance, and kurtosis, and using the chi-square distribution fitting method to compare the convolution output with the hidden danger feature distribution under the specific environment of the power line; in the first layer, verifying whether the main target area of ​​the convolution kernel output covers specific hidden dangers; in the second layer, evaluating whether the features are robust to environmental influencing factors, and verifying the stability of hidden danger feature extraction by superimposing feature comparisons under different conditions; in the third layer, verifying whether global features and local features appear in the detection results at the same time to ensure the comprehensive adaptability of the model.

[0008] In a possible implementation, the environmental and hidden danger characteristics of the power overhead lines are used to generate self-supervised training signals, specifically including: dividing the power overhead lines into mountainous areas, plains, and cities, obtaining the hidden danger characteristics of each area and performing significance scoring to obtain regional characteristic enhancement signals; generating hidden danger migration mode signals based on historical hidden danger records of the power overhead lines combined with environmental characteristics; dividing the historical hidden danger records of the power overhead lines into different hidden danger areas according to location distribution, determining the frequency of occurrence of hidden danger characteristics in the corresponding area based on the characteristics of each hidden danger area, using environmental characteristics as an influencing factor for hidden danger migration, generating a migration matrix, and obtaining a hidden danger migration mode signal;

[0009] The time-frequency characteristics of the periodic signal are analyzed for the hidden danger characteristics, and the frequency domain signal is obtained according to the following relationship:

[0010]

[0011] in is the short-time Fourier transform, is the frequency enhancement factor, is the frequency characteristic weight function, which is used to increase the significance of specific hidden danger frequency. is the enhancement coefficient, which indicates the enhancement intensity of the high-significance frequency component. is the time decay factor, is a time disturbance function, which indicates the intensity change of hidden danger characteristics in a specific time period. is the time attenuation coefficient, which is used to control the overall impact of time disturbance on the frequency domain signal; the amplitude and phase information of the frequency domain signal are mapped into a two-dimensional spectrum diagram, in which each pixel represents the intensity of the frequency component at the hidden danger position.

[0012] In another possible implementation, the power overhead lines are divided into mountainous areas, plains, and cities, and the hidden danger characteristics of each area are obtained and scored for significance, so as to obtain regional characteristic enhancement signals, including: dividing the transmission line environment into mountainous areas, plains, cities and other areas, and generating a hidden danger feature set for the area based on actual data, wherein, for mountainous areas, the sag and tension distribution characteristics of the conductors caused by snow accumulation and freezing, and the bottom-up diffusion characteristics of the channel smoke caused by wildfires are extracted; for plain areas, the threat characteristics of vegetation interference areas to the conductors are extracted; for urban areas, the construction machinery characteristics of high-frequency interference sources and industrial emission smoke characteristics are extracted; for the hidden danger characteristics of each area, the significance score is calculated according to the following relationship :

[0013]

[0014] in is the correction term for environmental impact factors, , through the coupled calculation of environmental impact and distance attenuation, the close-range hidden danger characteristics in the critical environment are given priority, is the environmental sensitivity coefficient of the characteristic, which indicates the degree to which a specific hidden danger characteristic is dynamically affected by the environment. is the distance decay function, which means that the importance of the feature decreases with the distance from the target area. The increase of exponential decay; , , is the weighting coefficient, which indicates the weight of each part in the score and is adjusted dynamically according to the actual application scenario; Hidden danger features The response intensity indicates the degree of attention paid by the model to the hidden danger feature; Hidden danger features The importance weights are dynamically adjusted based on environmental factors; is the environmental complexity coefficient, which is used to normalize the diffusion risk and avoid overestimating the importance of a certain feature in a highly complex environment; Hidden danger features The rate of change in the time dimension, ,

[0015]

[0016] in Indicates hidden danger characteristics At the point in time The response value of is the smoothing term, is the logarithmic function of diffusion risk; is the potential danger diffusion coefficient of the region,

[0017]

[0018] in To spread hidden dangers from the current area to the area The probability of For the current region and region The distance between.

[0019] In another possible implementation, a hidden danger migration pattern signal is generated based on the historical hidden danger records of the power overhead line in combination with environmental characteristics, including: dividing the historical hidden danger records of the power overhead line into different hidden danger areas according to location distribution; constructing a hidden danger distribution matrix based on the characteristics of each hidden danger area, and the element value of the distribution matrix represents the frequency of hidden danger occurrence in the corresponding area; using environmental characteristics as an influencing factor for hidden danger migration, generating a migration matrix , the migration matrix middle Indicates hidden dangers from the area To the region Probability of expansion, hidden danger migration signal The generation formula is as follows:

[0020]

[0021] in is the periodic disturbance term; is the periodic disturbance intensity, which indicates the randomness or periodic fluctuation in the potential danger diffusion path; is the angular frequency of the periodic disturbance; For distance A periodic perturbation function is used to simulate the nonlinear effects in the diffusion path; Influencing factors of environmental characteristics, The weight of the physical conditions that represent the spread of hidden dangers between regions; the expansion trend signal is generated according to the migration matrix, the expansion trend of the hidden danger area is calculated, and a priority weight is generated for each expansion area, and the hidden danger migration pattern signal is obtained by combining them.

[0022] In another possible implementation, the extracted features and self-supervised training signals are input into the seed model, the weights and training strategies are adjusted according to the real-time feedback of the seed model, and the parameters of the seed model are dynamically optimized to obtain a proprietary model that adapts to the environmental characteristics of different locations, including: dynamically adjusting the weight distribution in the hidden danger feature extraction path in combination with the significance score, and designing the gradient update formula as follows:

[0023]

[0024] in, Represents the current model weight matrix; Score the significance of the hazard characteristics; Get a new weight matrix for gradient update; It is a risk-weighted gradient correction item, which enables the optimization of more concentrated features based on the spatial distribution characteristics of hidden danger features. represents the learning rate; represents the loss function, Represents the gradient of the loss function with respect to the weight; is the environmental weight of the hazard characteristics; Hidden danger features Response strength, is the variance attenuation factor. When the feature distribution is more dispersed, the weight decreases, and when the feature is concentrated, the weight increases. is the adaptive risk adjustment coefficient; Hidden danger features The distribution variance of is the variance attenuation coefficient; the target-oriented weight redistribution mechanism is used to improve the model's learning ability for the target area by assigning priority to the target output area error weight. According to the target area detection error, the activation value of the target area feature channel is dynamically adjusted. The target area activation adjustment formula is designed as follows:

[0025]

[0026] in, is the target area feature channel The current activation value of is the adjusted target region feature channel activation value; is the target area error weight; is the target area error, which indicates the optimization demand intensity of the potential danger area; is the global error, which is used to normalize the optimization weight of the target area; is the feature channel distribution priority value, is the total number of feature channels in the target area; The feature channel The weight of is calculated in combination with the significance score; is the dynamically changing attenuation coefficient; is the attenuation factor, based on the dynamic rate of change of the feature Reduce the weight of channels with large fluctuations; attenuate the weight of non-target areas, and design the weight adjustment formula for non-target area channels as follows:

[0027]

[0028] in, Non-target area feature channel The current weight of is the adjusted feature channel weight of the non-target area; Optimized attenuation coefficient for non-target areas; is the error weight of non-target area; is the sum of errors in the non-target area; is the error weight dynamic attenuation factor, For Channel The characteristic variance of is the feature variance benchmark of the non-target area.

[0029] In the second aspect, an intelligent monitoring device for power overhead lines based on self-supervision is provided, the device comprising: a front-end acquisition device and a self-supervision pre-training module mounted on the front-end acquisition device, the self-supervision pre-training module comprising a training inference module, the training inference module comprising a model iteration training unit and a model inference unit; the front-end acquisition device is used to collect real-time multi-dimensional image data of the power overhead lines, the front-end acquisition device comprises a multi-spectral camera and an infrared camera, the multi-spectral camera and the infrared camera support timed shooting and event-triggered shooting, and adjust the shooting frequency according to environmental conditions; the model iteration training unit is used to perform self-supervision pre-training on the seed model transmitted by the cloud server according to the historical multi-dimensional image data of the power overhead lines, and generate a proprietary model adapted to the characteristics of a specific environment, specifically comprising: receiving Receive the seed model transmitted by the cloud server, parse the seed model into three parts: weight matrix, hidden danger feature extraction path description, and pre-training task parameter configuration, and perform hierarchical feature consistency check; generate self-supervised training signals based on the environmental and hidden danger characteristics of the power overhead lines; perform hierarchical feature extraction on the intrinsic structure and characteristics of the historical multi-dimensional image data itself; input the extracted features and self-supervised training signals into the seed model, adjust the weights and training strategies according to the real-time feedback of the seed model, dynamically optimize the parameters of the seed model, and obtain a proprietary model that adapts to the environmental characteristics of different locations; the model inference unit is used to apply the proprietary model to real-time multi-dimensional image data for hidden danger identification, obtain hidden danger identification results, and transmit the hidden danger identification results to the visualization monitoring equipment for visualization and early warning.

[0030] On the third aspect, another intelligent monitoring method for overhead power lines based on self-supervision is provided, including: receiving multidimensional image data transmitted by a front-end acquisition device, the front-end acquisition device including a multispectral camera and an infrared camera, the multispectral camera and the infrared camera support timed shooting and event-triggered shooting, and adjust the shooting frequency according to environmental conditions; extracting features from the multidimensional image data at three levels: spatial domain, frequency domain and time domain, and generating feature information of different domains for characterizing the transmission channel environment, wherein a multi-scale convolution kernel is used to extract spatial features including edges and textures in the multidimensional image data to obtain spatial domain feature information; discrete Fourier transform is used to analyze the spectrum information of the multidimensional image data, frequency components and phase features are extracted, and frequency domain feature information is obtained; image sequences within a continuous time period are analyzed to capture dynamic change features and obtain time domain feature information; according to the actual conditions of the front-end environment, the weights of the feature information of each domain are adaptively adjusted, and the feature information of different domains is multimodally fused to form a unified feature representation with robustness, thereby forming a seed model.

[0031] In a fourth aspect, a cloud server is provided, including: a data receiving module, used to receive multidimensional image data transmitted by a front-end acquisition device, the front-end acquisition device includes a multispectral camera and an infrared camera, the multispectral camera and the infrared camera support timed shooting and event-triggered shooting, and adjust the shooting frequency according to environmental conditions; a feature extraction module, used to extract features of the multidimensional image data at three levels: spatial domain, frequency domain and time domain, and generate feature information of different domains for characterizing the transmission channel environment, wherein a multi-scale convolution kernel is used to extract spatial features including edges and textures in the multidimensional image data to obtain spatial domain feature information; a discrete Fourier transform is used to analyze the spectral information of the multidimensional image data, and the frequency components and phase features are extracted to obtain frequency domain feature information; an image sequence within a continuous time period is analyzed to capture dynamic change features and obtain time domain feature information; a feature fusion module, used to adaptively adjust the weights of the feature information of each domain according to the actual conditions of the front-end environment, and perform multimodal fusion of the feature information of different domains to form a unified feature representation with robustness to form a seed model.

[0032] In a fifth aspect, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, an intelligent monitoring method for overhead power lines based on self-supervision as provided in the first aspect is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in describing the embodiments of the present application are briefly introduced below.

[0034] Figure 1 A schematic diagram of the structure of an intelligent monitoring system for power overhead lines based on self-supervision provided by one embodiment of the present invention;

[0035] Figure 2 A schematic diagram of the structure of an intelligent monitoring device for power overhead lines based on self-supervision provided by one embodiment of the present invention;

[0036] Figure 3 A schematic diagram of the structure of a cloud server provided by an embodiment of the present invention;

[0037] Figure 4 A schematic flow chart of an intelligent monitoring method for power overhead lines based on self-supervision provided by an embodiment of the present invention;

[0038] Figure 5 A schematic flow chart of an intelligent monitoring method for power overhead lines based on self-supervision provided in another embodiment of the present invention;

[0039] Figure 6A schematic diagram of the physical structure of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0040] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar modules or modules with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be interpreted as limiting the present invention.

[0041] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present application refers to the presence of the features, integers, steps, operations, modules and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, modules, components and / or groups thereof. It should be understood that when we refer to a module as being "connected" or "coupled" to another module, it may be directly connected or coupled to the other modules, or there may be intermediate modules. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein includes all or any modules and all combinations of one or more associated listed items.

[0042] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation method of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0043] The technical solution of the present application and how to solve the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0044] An embodiment of the present invention provides an intelligent monitoring system for power overhead lines based on self-monitoring, such as Figure 1As shown, the intelligent monitoring system of the power overhead line based on self-supervision includes: an intelligent monitoring device of the power overhead line based on self-supervision, a visual monitoring device and a cloud server. The intelligent monitoring device of the power overhead line based on self-supervision includes a front-end acquisition device and a self-supervised pre-training module mounted on the front-end acquisition device. The front-end acquisition device is installed at multiple key positions of the power overhead line. The front-end acquisition device collects multi-dimensional image data through timed shooting and snapshot, and uploads it to the cloud server. The cloud server receives the multi-dimensional image data transmitted by the front-end acquisition device; and extracts features of the multi-dimensional image data at three levels of spatial domain, frequency domain and time domain, and generates feature information of different domains for characterizing the transmission channel environment; then the feature information of different domains is multi-modally fused to form a unified feature representation with robustness, and form a seed model. The cloud server sends the seed model to the self-supervised pre-training module. The self-supervised pre-training module performs self-supervised pre-training on the seed model transmitted by the cloud server according to the historical multi-dimensional image data of the power overhead line to obtain a trained proprietary model, which specifically includes: receiving the seed model transmitted by the cloud server, parsing the seed model into three parts: weight matrix, hidden danger feature extraction path description, and pre-training task parameter configuration, and performing hierarchical feature consistency verification; using the environmental and hidden danger characteristics of the power overhead line to generate a self-supervised training signal; performing hierarchical feature extraction on the intrinsic structure and characteristics of the historical multi-dimensional image data itself; inputting the extracted features and the self-supervised training signal into the seed model, adjusting the weights and training strategies according to the real-time feedback of the seed model, dynamically optimizing the parameters of the seed model, and obtaining a proprietary model that adapts to the environmental characteristics of different points. The self-supervised pre-training module obtains the real-time multi-dimensional image data of the power overhead line collected by the front-end acquisition device, and applies the proprietary model to identify hidden dangers in the real-time multi-dimensional image data, obtains hidden danger identification results, and transmits the hidden danger identification results to the visualization monitoring device for visualization display and early warning.

[0045] In this way, real-time monitoring and intelligent identification of hidden dangers in power overhead lines can be achieved. Through the combination of cloud processing servers and self-supervised pre-training modules, this embodiment can adaptively adjust the model in different environments to ensure efficient identification of hidden dangers in the line. At the same time, through the combined use of multi-spectral cameras and infrared cameras, as well as image data processed by the image enhancement module, the accuracy and efficiency of hidden danger identification are significantly improved.

[0046] like Figure 2 The structure diagram of an intelligent monitoring device for power overhead lines based on self-supervision provided by an embodiment of the present invention is shown, the device comprises: a front-end acquisition device and a self-supervision pre-training module mounted on the front-end acquisition device, the self-supervision pre-training module comprises a training reasoning module and an image enhancement module. The training reasoning module comprises a model iteration training unit and a model reasoning unit.

[0047] The front-end acquisition device is used to collect real-time multi-dimensional image data of overhead power lines; the model iteration training unit is used to perform self-supervised pre-training on the seed model transmitted by the cloud server according to the historical multi-dimensional image data of the overhead power lines to obtain a trained proprietary model; the model inference unit is used to apply the proprietary model to identify hidden dangers in the real-time multi-dimensional image data, obtain hidden danger identification results, and transmit the hidden danger identification results to the visualization monitoring equipment for visualization display.

[0048] The front-end acquisition device specifically includes a multispectral camera and an infrared camera. The multispectral camera can simultaneously collect multi-dimensional visible light image information in different bands. The different bands include but are not limited to red light, green light, blue light, and near infrared. The infrared camera is used to collect thermal imaging images of the line and its surrounding environment for detecting temperature anomalies. The working modes of the multispectral camera and infrared camera support timed shooting and event-triggered shooting. The shooting frequency can be adjusted according to environmental conditions. For example, in strong winds or extreme weather conditions, the system will automatically increase the shooting frequency to ensure that potential hidden dangers around the line can be captured in time. The captured image data is transmitted to the cloud server in real time via 4G / 5G network or satellite communication.

[0049] The self-supervised pre-training module sets the "training" and "inference" time periods. In the "training" time period, all images transmitted by the front-end acquisition device are used as real-time data sets, and the seed model sent by the cloud server is iteratively trained to generate a proprietary model that adapts to the characteristics of a specific environment. Specifically, it includes: receiving the seed model transmitted by the cloud server, parsing the seed model into three parts: weight matrix, hidden danger feature extraction path description, and pre-training task parameter configuration, and performing hierarchical feature consistency verification; using the environmental and hidden danger characteristics of the power overhead line to generate a self-supervised training signal; performing hierarchical feature extraction on the inherent structure and characteristics of the historical multi-dimensional image data itself; inputting the extracted features and the self-supervised training signal into the seed model, adjusting the weights and training strategies according to the real-time feedback of the seed model, and dynamically optimizing the parameters of the seed model to obtain a proprietary model that adapts to the environmental characteristics of different points. In the "inference" time period, the trained proprietary model is used to identify hidden dangers on the input image, and the identification results are pushed to the terminal visualization monitoring device for visualization and early warning. Specifically, the images collected by the front-end acquisition device will first undergo preliminary image processing steps, including denoising, contrast enhancement, and color correction, and the processed image data will be used as a real-time data set.

[0050] The self-supervised pre-training module includes a training and reasoning module and an image enhancement module. The training and reasoning module iteratively trains the seed model sent by the cloud processing server and uses the model to infer and recognize real-time images. The image enhancement module pre-processes different types of image data separately to improve the accuracy of subsequent feature extraction and hidden danger identification.

[0051] The image enhancement module includes the following functional units:

[0052] A noise removal unit, used to remove high-frequency noise in the image to improve the signal-to-noise ratio of the image;

[0053] The contrast enhancement unit enhances the contrast of the image through an adaptive histogram equalization algorithm to make the target area more prominent;

[0054] The color correction unit is used to correct the color cast of the image to ensure that the images acquired under different lighting conditions have consistent color performance.

[0055] The training and inference module also includes a model receiving unit for receiving a seed model sent from a cloud processing server. The model receiving unit ensures that the received seed model maintains data integrity during transmission and decodes and verifies it immediately after receipt to ensure the accuracy of the model in subsequent processing.

[0056] The model iteration training unit is started during the "training" time period, and is used to generate a data set using all the images transmitted by the front-end acquisition device, and iteratively train the received seed model. The model iteration training unit automatically generates training signals by utilizing the inherent structure and characteristics of the data itself, without relying on manually annotated labels. By extracting and adjusting the weights of a large amount of image data in the data set, the model parameters are dynamically optimized to generate a proprietary model that can adapt to the environmental characteristics of different points. Specifically, the seed model is dynamically adjusted in combination with the meteorological data and geographical features of the transmission line, and environmental adaptation is achieved through weight adjustment and structural optimization; based on the self-supervised learning framework, a multi-task collaborative training strategy is designed, including time series consistency prediction, cross-modal comparative learning, and feature restoration tasks; feature weights are adjusted and training strategies are optimized according to real-time feedback from the model, so that the seed model can dynamically adapt to the diverse line environment. During the iterative training process, the model iteration training unit will also adjust the training strategy according to the real-time feedback of the model to ensure that the model can accurately reflect the current environmental changes. When the model iteration training unit automatically generates training signals by utilizing the intrinsic structure and characteristics of the data itself, specifically, the model iteration training unit compares the similarities and differences between different samples to learn effective representations, maps different transformations of the same data sample (such as rotation, cropping, etc.) to similar representation spaces, and maps different samples to representation spaces that are as different as possible. The model learns the intrinsic characteristics of the data through this comparative relationship. The input data is compressed into a low-dimensional representation through a bottleneck layer, and then reconstructed back to the original data, so that the reconstructed data is as close to the original data as possible. Through the compression and reconstruction process, the autoencoder is used to enable the model to capture the main features and patterns of the data. The intrinsic structure of the data is the feature that can be represented in a low-dimensional space. Through the fill-in-the-blank task, the seed model needs to predict the covered or removed parts based on the context, so that the seed model understands the relationship between the various parts of the data to learn the intrinsic structure.

[0057] The model inference unit is started during the "inference" time period and uses the trained proprietary model to identify hidden dangers on the input real-time image data. The model inference unit performs image feature matching and hidden danger detection, packages the recognition result data, and transmits it to the terminal visual monitoring device. The unit also evaluates the confidence of the recognition result. When the confidence is lower than the threshold preset by the terminal visual monitoring device, the output recognition result will not have a hidden danger detection box.

[0058] During the "inference" period, the self-supervised pre-training module uses the trained proprietary model to identify hidden dangers in real-time image data. By comparing the current input image with the feature patterns learned by the model, the self-supervised pre-training module can identify potential hidden dangers in the line, such as tree branches touching the line, broken wires, damaged insulators, etc. The results of hidden danger identification include the type of hidden danger (such as tree branches touching, broken wires, damaged insulators, etc.), specific location, risk level and confidence. The hidden danger identification results are packaged into visual data and transmitted to the terminal visual monitoring device through the data transmission channel. The system will determine whether to trigger an alarm based on the confidence of the recognition result. When the confidence is higher than the preset threshold, the system will automatically trigger an alarm and convey the hidden danger information to the maintenance personnel to ensure that the line hidden dangers can be responded to and handled in a timely manner.

[0059] The visual monitoring device displays the recognition results in real time through the user interface. Users can view the type, location and risk level of the currently detected hidden dangers, and obtain corresponding alarm prompts. When the confidence of the recognition result exceeds the preset threshold, the system will automatically trigger an alarm and push the alarm information to the maintenance personnel to ensure that the hidden dangers can be handled in time. Users can also access historical data records through the monitoring interface and manually calibrate certain specific hidden dangers. These calibration information will be used for subsequent model iteration training to improve the accuracy of future model training, thereby improving the overall recognition accuracy of the system. The system allows users to mark false positives or missed positives, and use this information for subsequent model optimization through a feedback mechanism.

[0060] The embodiment of the present invention can automatically learn and identify the hidden danger characteristics of overhead power lines through the self-supervised pre-training module, thereby significantly improving the monitoring accuracy and reducing the cases of missed detection and false detection. At the same time, the self-supervised learning algorithm can automatically generate training samples in the absence of a large amount of labeled data, reducing the workload and cost of manual labeling. It is possible to monitor overhead power lines in real time, discover hidden dangers in a timely manner, and issue early warning information to help operation and maintenance personnel take measures quickly to reduce potential risks. The self-supervised pre-training module has good adaptive capabilities and can dynamically adjust and optimize according to actual data to improve the adaptability of the system in different environments and conditions. It can effectively deal with various noises and interferences, maintain high stability and robustness, and ensure the reliability of monitoring results. The present invention combines self-supervised pre-training and overhead power line monitoring technology to fill the gap in the existing technology in automatic identification and early warning of hidden dangers, and has high technical innovation and application prospects.

[0061] An embodiment of the present invention also provides a cloud server. Figure 3 , the cloud server includes:

[0062] A data receiving module, used for receiving multi-dimensional image data transmitted by a front-end acquisition device;

[0063] A feature extraction module is used to extract features from multi-dimensional image data in three levels: spatial domain, frequency domain and time domain, and generate feature information in different domains for characterizing the power transmission channel environment;

[0064] The feature fusion module is used to perform multimodal fusion of feature information from different domains to form a unified feature representation with robustness and form a seed model.

[0065] The feature extraction module adopts a multi-level convolutional neural network (CNN) structure, which is used to extract the spatial domain, frequency domain and time domain feature information of the image data respectively. The feature extraction module includes a spatial domain feature extraction unit, a domain feature extraction unit and a time domain feature extraction unit. Specifically, the spatial domain feature extraction unit uses a multi-scale convolution kernel to extract spatial features such as edges and textures in the image, and specifically identifies key structural features in the image, such as lines, insulators and surrounding vegetation, through a convolutional neural network (CNN). The frequency domain feature extraction unit uses discrete Fourier transform (DFT) to analyze the spectrum information of the image, extracts spectrum information such as frequency components and phase characteristics, and identifies periodic signal anomalies, especially the characteristics in the case of wire vibration or arc discharge. The time domain feature extraction unit analyzes the image sequence in a continuous time period, captures dynamic change features, such as the feature information of dynamic scenes such as wind blowing branches, and identifies dynamic hidden dangers of the line caused by external factors such as wind and rain.

[0066] The extracted feature information of each domain is multimodally fused through the feature fusion module to form a more robust unified feature representation. The feature fusion module includes a weighted fusion subunit, a feature selection subunit, and a feature compression subunit, where the control parameters are remotely set by the terminal visualization monitoring device.

[0067] Specifically, the weighted fusion subunit assigns weights to different modal data according to their importance, ensuring that key features dominate the final feature representation.

[0068] The feature selection subunit uses the principal component analysis algorithm to select the most representative features from multimodal data and eliminate redundant information.

[0069] The feature compression subunit compresses high-dimensional feature vectors to reduce the amount of computation for data processing and avoid model overfitting problems.

[0070] The generated seed model is then sent to the self-supervised pre-training module.

[0071] The cloud server adaptively adjusts the weights of feature information in each domain according to the actual conditions of the front-end environment to ensure the accuracy and robustness of feature representation. The generated unified feature representation is used as a seed model and sent to the self-supervised pre-training module for subsequent iterative training. Through the combination of the cloud server and the self-supervised pre-training module, this embodiment can adaptively adjust the model in different environments to ensure efficient identification of hidden dangers on the line. At the same time, through the joint use of multi-spectral cameras and infrared cameras, as well as image data processed by the image enhancement module, the accuracy and efficiency of hidden danger identification are significantly improved.

[0072] like Figure 4 The figure is a flow chart of an intelligent monitoring method for an overhead power line based on self-supervision provided by an embodiment of the present invention. The intelligent monitoring method is applied to the aforementioned intelligent monitoring device for overhead power lines based on self-supervision, and specifically includes:

[0073] Step S11, using a front-end acquisition device to acquire real-time multi-dimensional image data of the power overhead line.

[0074] Optionally, the front-end acquisition device includes a multi-spectral camera and an infrared camera, and the multi-spectral camera and the infrared camera support timed shooting and event-triggered shooting.

[0075] Multispectral cameras are used to collect multi-dimensional visible light image information in different bands. Infrared cameras are used to collect thermal images of power overhead lines and their surroundings, which are used to detect temperature anomalies. The multispectral camera and the infrared camera are controlled by hardware synchronization mechanism to achieve simultaneous shooting, and the synchronously shot data streams are preliminarily aligned and corrected to form a multi-dimensional image data set with time consistency.

[0076] Step S12, based on the historical multidimensional image data of the power overhead line, the seed model transmitted by the cloud server is self-supervised pre-trained to generate a proprietary model adapted to the characteristics of a specific environment, specifically comprising: receiving the seed model transmitted by the cloud server, parsing the seed model into three parts: a weight matrix, a description of a hidden danger feature extraction path, and a pre-training task parameter configuration, and performing a hierarchical feature consistency check; using the environmental and hidden danger characteristics of the power overhead line to generate a self-supervised training signal; performing hierarchical feature extraction on the intrinsic structure and characteristics of the historical multidimensional image data itself; inputting the extracted features and the self-supervised training signal into the seed model, adjusting the weights and training strategies according to the real-time feedback of the seed model, dynamically optimizing the parameters of the seed model, and obtaining a proprietary model adapted to the characteristics of the environment at different locations.

[0077] In an embodiment of the present invention, the seed model transmitted by the cloud server is first received, and the received seed model is decoded and verified. Specifically, the seed model is parsed into three parts: weight matrix, hidden danger feature extraction path description, and pre-training task parameter configuration, and a hierarchical feature consistency check is performed.

[0078] Specifically, it includes: comparing the characteristic response pattern of the weight matrix with the historical feature library of the power overhead line, and matching the historical high-risk hidden danger areas (such as historical wildfires and mechanical construction sites). Calculate the statistical indicators of the output feature distribution of each level of the convolution kernel, including the mean, variance and kurtosis, and use the chi-square distribution fitting method to compare the convolution output with the hidden danger feature distribution under the specific environment of the power line (such as different voltage levels and climatic conditions). In the first layer, verify whether the main target area of ​​the convolution kernel output covers specific hidden dangers (such as the surface damage area of ​​the conductor). In the second layer, evaluate whether the features are robust to environmental factors (such as channel wildfires caused by high heat), and verify the stability of hidden danger feature extraction by superimposing feature comparisons under different conditions. In the third layer, verify whether global features (such as line vibration amplitude) and local features (such as crack texture) appear in the detection results at the same time to ensure the comprehensive adaptability of the model.

[0079] Then, the environmental and hidden danger characteristics of the power overhead lines are used to generate self-supervised training signals, including: dividing the power overhead lines into mountainous areas, plains, and cities, obtaining the hidden danger characteristics of each area and performing significance scoring to obtain regional characteristic enhancement signals; generating hidden danger migration pattern signals based on the historical hidden danger records of the power overhead lines combined with environmental characteristics; dividing the historical hidden danger records of the power overhead lines into different hidden danger areas according to location distribution, determining the frequency of hidden danger characteristics in the corresponding areas based on the characteristics of each hidden danger area, using environmental characteristics as the influencing factors of hidden danger migration, generating a migration matrix, and obtaining hidden danger migration pattern signals; performing time-frequency characteristic analysis on the hidden danger characteristics of periodic signals, and obtaining frequency domain signals according to the following relationship:

[0080]

[0081] in is the short-time Fourier transform, is the frequency enhancement factor, is the frequency characteristic weight function, which is used to increase the significance of specific hidden danger frequency. is the enhancement coefficient, which indicates the enhancement intensity of the high-significance frequency component. is the time attenuation factor, which dynamically adjusts the signal strength by combining time and frequency factors. is a time disturbance function, which indicates the intensity change of hidden danger characteristics in a specific time period. is the time attenuation coefficient, which is used to control the overall impact of time disturbance on the frequency domain signal;

[0082] The amplitude and phase information of the frequency domain signal are mapped into a two-dimensional spectrum diagram, where each pixel represents the intensity of the frequency component at the hidden danger location, and the potential hidden dangers are automatically calibrated through the significant area of ​​the spectrum diagram.

[0083] The training signals are automatically generated by using the intrinsic structure and characteristics of the historical multidimensional image data of the power overhead lines. Specifically, based on the historical multidimensional image data samples of the power overhead lines, the similarities and differences between different samples are compared to learn effective representations, and different transformations of the same data sample are mapped to similar representation spaces, while different samples are mapped to representation spaces that are as different as possible. The intrinsic features of the learning data used for representation in low-dimensional space are compared; a bottleneck layer is applied to compress the input data into a low-dimensional representation, and then reconstruct it back to the original data. Through the compression and reconstruction process, the main features of the data are captured by the autoencoder; the gap filling task predicts the covered or removed parts according to the context, determines the relationship between the various parts in the data, and obtains the intrinsic structure of the data; the training signals are automatically generated according to the intrinsic structure and main features of the data.

[0084] Regional characteristic enhanced signal generation: The transmission line environment is divided into mountainous areas, plains, cities and other areas, and regional specific feature sets are generated based on actual data. For mountainous areas, priority is given to extracting the sag and tension distribution characteristics of conductors caused by snow and ice accumulation, and the bottom-up diffusion characteristics of channel smoke caused by wildfires; for plain areas, priority is given to extracting the threat characteristics of large-scale vegetation interference areas such as grass and tree growth to conductors; for urban areas, high-frequency interference sources such as construction machinery characteristics such as tower cranes and cement pump trucks, and industrial emission smoke characteristics are extracted.

[0085] For each area’s hidden danger characteristics, the significance score is calculated according to the following relationship: :

[0086]

[0087] in is the correction term for environmental impact factors, , through the coupled calculation of environmental impact and distance attenuation, the close-range hidden danger characteristics in the critical environment are given priority, The environmental sensitivity coefficient of the characteristic indicates the degree to which a specific hazard characteristic is dynamically affected by the environment (for example, the enhancing effect of humidity on discharge hazards). is the distance decay function, which means that the importance of the feature decreases with the distance from the target area. The increase of exponential decay; , , is the weighting coefficient, which indicates the weight of each part in the score and is adjusted dynamically according to the actual application scenario; Hidden danger features The response intensity indicates the degree of attention paid by the model to the hidden danger feature; Hidden danger features The importance weights are dynamically adjusted based on environmental factors (such as humidity, wind speed, and terrain); is the environmental complexity coefficient, which is used to normalize the diffusion risk and avoid overestimating the importance of a certain feature in a highly complex environment; Hidden danger features The rate of change in the time dimension, , used to assess the sudden trend or long-term stability of hidden dangers, is defined as:

[0088]

[0089] in Indicates hidden danger characteristics At the point in time The response value of is a smoothing term to avoid the denominator being zero. To diffuse the risk logarithmic function, a logarithmic function is used to compress the risk and enhance the adaptability of the model to high-risk areas; is the potential risk diffusion coefficient of the region, which is calculated based on the probability of hidden danger occurrence and the propagation distance. It indicates the potential risk of hidden danger spreading from this region to other regions and is defined as:

[0090]

[0091] in To spread hidden dangers from the current area to the area The probability of For the current region and region The distance between

[0092] Incorporate the time dimension to calculate the rate of change of hidden danger characteristics at different times and capture the sudden trend or stable characteristics of hidden dangers. For example, for flame jumping characteristics, the rapidly changing edge response will obtain a higher For vegetation disturbance characteristics, slowly changing disturbance areas will receive lower value;

[0093] Considering the risk of hidden dangers spreading between regions, the probability of hidden dangers occurring is and propagation distance Quantification enhances the regional adaptability of the rating. For example, in a high humidity environment, insulation aging High, neighboring areas Higher score; in the mechanical interference scenario, the hidden danger propagation distance Directly affects its rating;

[0094] By weight , , Dynamically allocate the proportions of each part of the score so that the model can flexibly adjust to different scenarios, such as the increase in the number of fire-prone areas and value, reduce The value of makes the dynamic change of hidden dangers and the influence of diffusion coefficient more significant; the sag risk area caused by snow accumulation increases and The value of makes the model pay more attention to the static characteristic response and dynamic change rate.

[0095] The hidden danger migration pattern signal is the concretization and modeling of the expansion trend signal, which expresses the probability and influencing factors of hidden danger expansion through matrix and formula.

[0096] Hidden danger migration pattern signal generation: divide the historical hidden danger records of power overhead lines into different hidden danger areas (such as insulators, conductors, and channels) according to their location distribution; construct a hidden danger distribution matrix based on the characteristics of each hidden danger area (such as crack width and temperature anomaly values), and the element value of the matrix represents the frequency of hidden dangers in the corresponding area; use environmental characteristics (such as temperature, humidity, and wind speed) as influencing factors for hidden danger migration to generate a migration matrix , in the matrix Indicates hidden dangers from the area To the region The extended probability and design hidden danger migration signal generation formula are as follows:

[0097]

[0098] in It is a periodic disturbance term, which can truly reflect the dynamic behavior of hidden danger migration; is the intensity of periodic disturbances, which indicates the randomness or periodic fluctuations in the diffusion path of hidden dangers, such as the uncertainty of smoke diffusion caused by changes in wind direction; is the angular frequency of the periodic disturbance; For distance The periodic perturbation function is used to simulate the nonlinear effects in the diffusion path (such as the high-frequency diffusion of hidden dangers in short distances and gradual attenuation over long distances); factors that represent the influencing characteristics of the environment (e.g., high humidity increases the likelihood of migration), The weight of the physical conditions that indicate the spread of hidden dangers between regions (such as the distance between two regions and the impact of wind direction on smoke diffusion);

[0099] The expansion trend signal is generated according to the migration matrix, the expansion trend of the hidden danger area is calculated, and a priority weight is generated for each expansion area. For example, the migration trend of insulation aging in a high-humidity environment preferentially extends from the wire joint to the insulator surface; the migration of wildfire hazards preferentially spreads from the foot of the mountain to the channel area near the line.

[0100] Frequency domain signal generation under specific circumstances: Considering the time-frequency characteristics analysis of periodic signals (such as mechanical sound wave vibration), the frequency domain signal generation formula is designed as follows:

[0101]

[0102] in is the short-time Fourier transform; is the frequency enhancement factor, which makes the model more sensitive to the key frequencies in the hidden danger characteristics; is a frequency feature weight function, which is used to increase the significance of specific hidden danger frequencies; is the enhancement coefficient, which indicates the enhancement intensity of the high-significance frequency component; The time attenuation factor is used to dynamically adjust the signal strength by combining time and frequency factors; is a temporal disturbance function, which represents the intensity change of hidden danger characteristics in a specific time period, such as the weakening of low-light signals in a night environment; is the time attenuation coefficient, which is used to control the overall impact of time disturbance on the frequency domain signal;

[0103] The amplitude and phase information of the frequency domain signal are mapped into a two-dimensional spectrum diagram, in which each pixel represents the intensity of a specific frequency component at the hidden danger location. For example, the frequency of the discharge hidden danger signal is concentrated at 50kHz, and the vibration hidden danger signal is concentrated at 10Hz. The potential hidden danger is automatically calibrated through the significant area of ​​the spectrum diagram.

[0104] Then, the intrinsic structure and characteristics of the historical multi-dimensional image data are extracted by hierarchical features, including:

[0105] Global feature extraction of hidden dangers: Design multi-scale convolution kernels to extract hidden danger features in different ranges. The convolution kernel captures the hidden dangers of large-scale smoke and vegetation cover; The convolution kernel captures the hidden danger features of small-scale construction machinery and wire cracks; the void convolution expands the receptive field and captures the contextual information of long-distance hidden dangers;

[0106] Feature sparse filtering: A sparse regularization method based on feature distribution that removes irrelevant or low-relevant features to prevent redundant information from affecting model optimization.

[0107] The extracted features and self-supervised training signals are input into the seed model, and the weights and training strategies are adjusted according to the real-time feedback of the seed model. The parameters of the seed model are dynamically optimized to obtain a proprietary model that adapts to the characteristics of different point environments, including:

[0108] Combined with the significance score, the weight distribution in the hidden danger feature extraction path is dynamically adjusted, and the gradient update formula is designed as follows:

[0109]

[0110] in, Represents the current model weight matrix, which indicates the current parameter status of each layer of the model network (such as the convolution kernel), and is the core goal of optimization; is the significance score of the hidden danger feature, indicating the importance of the current hidden danger feature; Get a new weight matrix for gradient update; It is a risk-weighted gradient correction item, which enables the optimization of more concentrated features based on the spatial distribution characteristics of hidden danger features. represents the learning rate; represents the loss function, Represents the gradient of the loss function with respect to the weight; is the environmental weight of the hidden danger characteristic, which represents the effect of environmental factors (such as humidity and temperature) on the hidden danger characteristic. The importance of the impact; Hidden danger features Response intensity, which indicates the sensitivity of the model to the hidden danger characteristics; is the variance attenuation factor. When the feature distribution is more dispersed, the weight decreases, and when the feature is concentrated, the weight increases. Hidden danger features The distribution variance of is the variance attenuation coefficient, which controls the weight of the distribution variance in the gradient update; It is an adaptive risk adjustment coefficient that automatically adjusts the risk impact ratio of gradient update according to the detection accuracy of the target area. For example, it increases when the detection error is large. value, give priority to deep optimization of key areas; reduce it when the detection error is small value, ensuring that the optimization focuses on the global feature distribution; the formula combines the significance score, feature distribution and environmental weight to extend the optimization from simple loss minimization to adaptive optimization of key hidden danger features.

[0111] Using the target-oriented weight redistribution mechanism, the model's learning ability for the target area is improved by assigning priority to the target output area error weight. According to the target area detection error, the activation value of the target area feature channel is dynamically adjusted. The target area activation adjustment formula is designed as follows:

[0112]

[0113] in, is the target area feature channel The current activation value of represents the current response strength of the model to the target area features; is the adjusted target region feature channel activation value; is the target area error weight, which dynamically adjusts the target area activation value according to the relative error ratio between the target and the global value; is the target area error, which indicates the optimization demand intensity of the potential danger area; is the global error, which is used to normalize the optimization weight of the target area; Distribute priority values ​​for feature channels, and improve the optimization of stable feature channels through dynamic priority allocation, while avoiding the interference of fluctuating features on optimization; is the total number of feature channels in the target area; The feature channel The weight of is calculated in combination with the significance score; is the dynamically changing attenuation coefficient; is the attenuation factor, based on the dynamic rate of change of the feature Reduce the weighting of channels with greater volatility.

[0114] For the weight attenuation of non-target areas, reduce the model's computing resource allocation in low-risk areas, focus on optimizing key hidden danger areas, and design the non-target area channel weight adjustment formula as follows:

[0115]

[0116] in, Non-target area feature channel The current weight of is the adjusted feature channel weight of the non-target area; The optimized attenuation coefficient of the non-target area controls the magnitude of weight attenuation in the non-target area; The error weight of the non-target area is used to reduce the optimization priority of the non-target area so that more computing resources can be concentrated in the target area. is the sum of errors in the non-target area; is the error weight dynamic attenuation factor, which is based on the dynamic adjustment of feature variance to enhance the ability to retain important features in non-target areas and significantly attenuate redundant features. For Channel The characteristic variance of is the feature variance benchmark of the non-target area.

[0117] Step S13, applying a proprietary model to the real-time multi-dimensional image data to identify hidden dangers, obtain hidden danger identification results, and transmit the hidden danger identification results to the visual monitoring equipment for visual display and early warning.

[0118] Before step S13, different types of image data are preprocessed separately to improve the accuracy of subsequent feature extraction and hidden danger identification. Optionally, high-frequency noise in real-time multi-dimensional image data is removed to improve the signal-to-noise ratio of the image and ensure the accuracy of subsequent feature extraction; the contrast of real-time multi-dimensional image data is enhanced by an adaptive histogram equalization algorithm to highlight the target area, which is helpful for the identification of hidden danger features; the color cast of real-time multi-dimensional image data is corrected to ensure that the real-time multi-dimensional image data acquired under different lighting conditions has consistent color performance, thereby improving the accuracy of feature extraction and hidden danger identification.

[0119] The embodiment of the present invention carries out intelligent monitoring of overhead power lines based on self-supervision, which can realize real-time monitoring and intelligent identification of hidden dangers of overhead power lines, and reduce false alarms and missed alarms by intelligently adjusting model parameters to adapt to the complex and changeable natural environment, thereby ensuring the safe operation of overhead power lines. Through the training and application of proprietary models, while ensuring high-quality and efficient completion of real-time monitoring tasks, the individual accuracy of the proprietary models is continuously optimized.

[0120] The method of the above embodiment is applied to the corresponding intelligent monitoring device for power overhead lines based on self-supervision in the above embodiment, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0121] like Figure 5 The figure is a flow chart of an intelligent monitoring method for a power overhead line based on self-monitoring provided by an embodiment of the present invention. The intelligent monitoring method is applied to the aforementioned cloud server and specifically includes:

[0122] Step S21, receiving multi-dimensional image data transmitted by a front-end acquisition device.

[0123] The front-end acquisition device includes a multi-spectral camera and an infrared camera. The multi-spectral camera and the infrared camera support timed shooting and event-triggered shooting, and adjust the shooting frequency according to environmental conditions.

[0124] Step S22, extracting features in three levels, namely, spatial domain, frequency domain and time domain, from the multi-dimensional image data to generate feature information in different domains for characterizing the power transmission channel environment.

[0125] Optionally, a multi-scale convolution kernel is used to extract spatial features including edges and textures in multi-dimensional image data to obtain spatial domain feature information; discrete Fourier transform is used to analyze the spectral information of multi-dimensional image data, frequency components and phase features are extracted, and frequency domain feature information is obtained; image sequences within a continuous time period are analyzed to capture dynamic change characteristics and obtain time domain feature information.

[0126] Step S23, according to the actual conditions of the front-end environment, the weights of the feature information of each domain are adaptively adjusted, and the feature information of different domains is multimodally fused to form a unified feature representation with robustness to form a seed model.

[0127] The method of the above embodiment is applied to the corresponding cloud server in the above embodiment, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0128] Figure 6 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 6 As shown, the electronic device may include: a processor (processor) 601, a communication interface (Communications Interface) 602, a memory (memory) 603 and a communication bus 604, wherein the processor, the communication interface and the memory communicate with each other through the communication bus. The processor may call the logic instructions in the memory to execute an intelligent monitoring method for a power overhead line based on self-supervision.

[0129] The method includes: using a front-end acquisition device to collect real-time multi-dimensional image data of the power overhead line; performing self-supervised pre-training on a seed model transmitted by a cloud server based on historical multi-dimensional image data of the power overhead line to obtain a trained proprietary model; using the proprietary model to identify hidden dangers in the real-time multi-dimensional image data, obtaining hidden danger identification results, and transmitting the hidden danger identification results to a visual monitoring device for visual display and early warning.

[0130] Alternatively, the method includes: receiving multidimensional image data transmitted by a front-end acquisition device; performing feature extraction on the multidimensional image data at three levels, namely, spatial domain, frequency domain and time domain, to generate feature information of different domains for characterizing the power transmission channel environment; adaptively adjusting the weight of the feature information of each domain according to the actual conditions of the front-end environment, performing multimodal fusion on the feature information of different domains, forming a unified feature representation with robustness, and forming a seed model.

[0131] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0132] On the other hand, an embodiment of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute an intelligent monitoring method for power overhead lines based on self-supervision provided by the above-mentioned method embodiments.

[0133] The method includes: using a front-end acquisition device to collect real-time multi-dimensional image data of the power overhead line; performing self-supervised pre-training on a seed model transmitted by a cloud server based on historical multi-dimensional image data of the power overhead line to obtain a trained proprietary model; using the proprietary model to identify hidden dangers in the real-time multi-dimensional image data, obtaining hidden danger identification results, and transmitting the hidden danger identification results to a visual monitoring device for visual display and early warning.

[0134] Alternatively, the method includes: receiving multidimensional image data transmitted by a front-end acquisition device; performing feature extraction on the multidimensional image data at three levels, namely, spatial domain, frequency domain and time domain, to generate feature information of different domains for characterizing the power transmission channel environment; adaptively adjusting the weight of the feature information of each domain according to the actual conditions of the front-end environment, performing multimodal fusion on the feature information of different domains, forming a unified feature representation with robustness, and forming a seed model.

[0135] On the other hand, an embodiment of the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute an intelligent monitoring method for overhead power lines based on self-supervision provided in the above-mentioned embodiments.

[0136] The method includes: using a front-end acquisition device to collect real-time multi-dimensional image data of the power overhead line; performing self-supervised pre-training on a seed model transmitted by a cloud server based on historical multi-dimensional image data of the power overhead line to obtain a trained proprietary model; using the proprietary model to identify hidden dangers in the real-time multi-dimensional image data, obtaining hidden danger identification results, and transmitting the hidden danger identification results to a visual monitoring device for visual display and early warning.

[0137] Alternatively, the method includes: receiving multidimensional image data transmitted by a front-end acquisition device; performing feature extraction on the multidimensional image data at three levels, namely, spatial domain, frequency domain and time domain, to generate feature information of different domains for characterizing the power transmission channel environment; adaptively adjusting the weight of the feature information of each domain according to the actual conditions of the front-end environment, performing multimodal fusion on the feature information of different domains, forming a unified feature representation with robustness, and forming a seed model.

[0138] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0139] The above is only a partial implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An intelligent monitoring method for power overhead lines based on self-supervision, characterized in that: The intelligent monitoring method comprises: A front-end acquisition device is used to collect real-time multi-dimensional image data of the power overhead line. The front-end acquisition device includes a multi-spectral camera and an infrared camera. The multi-spectral camera and the infrared camera support timed shooting and event-triggered shooting, and the shooting frequency is adjusted according to environmental conditions. The seed model transmitted by the cloud server is self-supervised pre-trained according to the historical multi-dimensional image data of the power overhead line to generate a proprietary model adapted to the characteristics of a specific environment, specifically including: receiving the seed model transmitted by the cloud server, parsing the seed model into three parts: a weight matrix, a description of a hidden danger feature extraction path, and a pre-training task parameter configuration, and performing a hierarchical feature consistency check; using the environmental and hidden danger characteristics of the power overhead line to generate a self-supervised training signal; performing hierarchical feature extraction on the intrinsic structure and characteristics of the historical multi-dimensional image data itself; inputting the extracted features and the self-supervised training signal into the seed model, adjusting the weight and training strategy according to the real-time feedback of the seed model, dynamically optimizing the parameters of the seed model, and obtaining a proprietary model adapted to the characteristics of the environment at different locations; Applying the proprietary model to identify hidden dangers in the real-time multi-dimensional image data, obtaining hidden danger identification results, and transmitting the hidden danger identification results to a visual monitoring device for visual display and early warning; The receiving of the seed model transmitted by the cloud server parses the seed model into three parts: weight matrix, hidden danger feature extraction path description, and pre-training task parameter configuration, and performs hierarchical feature consistency verification, specifically including: Compare the characteristic response pattern of the weight matrix with the historical characteristic library of power overhead lines to match the historical high-risk hidden danger areas; Calculate the statistical indicators of the output feature distribution of each level of convolution kernel, including mean, variance and kurtosis, and use the chi-square distribution fitting method to compare the convolution output with the hidden danger feature distribution under the specific environment of the power line; In the first layer, it is verified whether the main target area of ​​the convolution kernel output covers specific hidden dangers; in the second layer, it is evaluated whether the features are robust to environmental factors, and the stability of hidden danger feature extraction is verified by superimposing features of different conditions; in the third layer, it is checked whether global features and local features appear in the detection results at the same time to ensure the comprehensive adaptability of the model.

2. The method according to claim 1, characterized in that The method of utilizing the environment and hidden danger characteristics of the power overhead line to generate a self-supervised training signal specifically includes: Divide the power overhead lines into mountainous areas, plains and cities, obtain the hidden danger characteristics of each area and perform significance scoring to obtain regional characteristic enhancement signals; Generate hidden danger migration pattern signals based on historical hidden danger records of power overhead lines and environmental characteristics; The historical hidden danger records of power overhead lines are divided into different hidden danger areas according to their location distribution. The frequency of occurrence of hidden danger characteristics in the corresponding areas is determined based on the characteristics of each hidden danger area. Environmental characteristics are used as the influencing factors of hidden danger migration to generate a migration matrix and obtain hidden danger migration pattern signals. The time-frequency characteristics of the periodic signal are analyzed for the hidden danger characteristics, and the frequency domain signal is obtained according to the following relationship: in is the short-time Fourier transform, is the frequency enhancement factor, is the frequency characteristic weight function, which is used to increase the significance of specific hidden danger frequency. is the enhancement coefficient, which indicates the enhancement intensity of the high-significance frequency component. is the time decay factor, is a time disturbance function, which indicates the intensity change of hidden danger characteristics in a specific time period. is the time attenuation coefficient, which is used to control the overall impact of time disturbance on the frequency domain signal; The amplitude and phase information of the frequency domain signal are mapped into a two-dimensional spectrum diagram, where each pixel represents the intensity of the frequency component at the hidden danger location.

3. The method according to claim 2, characterized in that The power overhead lines are divided into mountainous areas, plains and cities, and the hidden danger characteristics of each area are obtained and scored for significance, so as to obtain regional characteristic enhancement signals, including: The transmission line environment is divided into mountainous areas, plains and cities, and a set of hidden danger characteristics of the region is generated based on actual data. For mountainous areas, the sag and tension distribution characteristics of the conductors caused by snow and ice accumulation, as well as the bottom-up diffusion characteristics of channel smoke caused by wildfires are extracted; for plain areas, the threat characteristics of vegetation interference areas to the conductors are extracted; for urban areas, the construction machinery characteristics of high-frequency interference sources and the characteristics of industrial smoke emissions are extracted; For each area’s hidden danger characteristics, the significance score is calculated according to the following relationship: : in is the correction term for environmental impact factors, , is the environmental sensitivity coefficient of the characteristic, which indicates the degree to which a specific hidden danger characteristic is dynamically affected by the environment. is the distance decay function, which means that the importance of the feature decreases with the distance from the target area. The increase of exponential decay; , , is the weighting coefficient, which indicates the weight of each part in the score and is adjusted dynamically according to the actual application scenario; Hidden danger features The intensity of the response; Hidden danger features The importance weights are dynamically adjusted based on environmental factors; is the environmental complexity coefficient; Hidden danger features The rate of change in the time dimension, , in Indicates hidden danger characteristics At the point in time The response value of is the smoothing term, is the logarithmic function of diffusion risk; is the potential danger diffusion coefficient of the region, in To spread hidden dangers from the current area to the area The probability of For the current region and region The distance between.

4. The method according to claim 2, characterized in that The generating of the hidden danger migration mode signal according to the historical hidden danger records of the power overhead line and the environmental characteristics includes: Divide the historical hidden danger records of power overhead lines into different hidden danger areas according to their location distribution; Constructing a hidden danger distribution matrix based on the characteristics of each hidden danger area, wherein the element value of the distribution matrix represents the frequency of occurrence of hidden dangers in the corresponding area; Use environmental characteristics as influencing factors for hidden danger migration to generate a migration matrix , the migration matrix middle Indicates hidden dangers from the area To the region Probability of expansion, hidden danger migration signal The generation formula is as follows: in is the periodic disturbance term; is the periodic disturbance intensity, which indicates the randomness or periodic fluctuation in the potential danger diffusion path; is the angular frequency of the periodic disturbance; For distance A periodic perturbation function is used to simulate the nonlinear effects in the diffusion path; Influencing factors of environmental characteristics, The weight of the physical conditions that represent the spread of hidden dangers between regions; An expansion trend signal is generated according to the migration matrix, the expansion trend of the hidden danger area is calculated, and a priority weight is generated for each expansion area, and the hidden danger migration mode signal is obtained by combining them.

5. The method according to claim 1, characterized in that The extracted features and the self-supervised training signal are input into the seed model, the weights and training strategies are adjusted according to the real-time feedback of the seed model, and the parameters of the seed model are dynamically optimized to obtain a proprietary model adapted to the characteristics of different point environments, including: Combined with the significance score, the weight distribution in the hidden danger feature extraction path is dynamically adjusted, and the gradient update formula is designed as follows: in, Represents the current model weight matrix; Score the significance of the hazard characteristics; Get a new weight matrix for gradient update; It is a risk-weighted gradient correction item, which enables the optimization of more concentrated features based on the spatial distribution characteristics of hidden danger features. represents the learning rate; represents the loss function, Represents the gradient of the loss function with respect to the weight; is the environmental weight of the hazard characteristics; Hidden danger features Response intensity; is the variance attenuation factor. When the feature distribution is more dispersed, the weight decreases, and when the feature is concentrated, the weight increases. is the adaptive risk adjustment coefficient; Hidden danger features The distribution variance of is the variance attenuation coefficient; Using the target-oriented weight redistribution mechanism, the model's learning ability for the target area is improved by assigning priority to the target output area error weight. According to the target area detection error, the activation value of the target area feature channel is dynamically adjusted. The target area activation adjustment formula is designed as follows: in, is the target area feature channel The current activation value of is the adjusted target region feature channel activation value; is the target area error weight; is the target area error, which indicates the optimization demand intensity of the potential danger area; is the global error, which is used to normalize the optimization weight of the target area; is the feature channel distribution priority value, is the total number of feature channels in the target area; The feature channel The weight of is calculated in combination with the significance score; is the dynamically changing attenuation coefficient; is the attenuation factor, based on the dynamic rate of change of the feature Reduce the weight of channels with greater volatility; The weight of the non-target area is attenuated, and the weight adjustment formula of the non-target area channel is designed as follows: in, Non-target area feature channel The current weight of is the adjusted feature channel weight of the non-target area; Optimized attenuation coefficient for non-target areas; is the error weight of non-target area; is the sum of errors in the non-target area; is the error weight dynamic attenuation factor, For Channel The characteristic variance of is the feature variance benchmark of the non-target area.

6. An intelligent monitoring device for power overhead lines based on self-monitoring, characterized in that: The intelligent monitoring device comprises: a front-end acquisition device and a self-supervised pre-training module mounted on the front-end acquisition device, wherein the self-supervised pre-training module comprises a training reasoning module, and the training reasoning module comprises a model iteration training unit and a model reasoning unit; The front-end acquisition device is used to collect real-time multi-dimensional image data of the power overhead line. The front-end acquisition device includes a multi-spectral camera and an infrared camera. The multi-spectral camera and the infrared camera support timed shooting and event-triggered shooting, and adjust the shooting frequency according to environmental conditions; The model iteration training unit is used to perform self-supervised pre-training on the seed model transmitted by the cloud server according to the historical multi-dimensional image data of the power overhead line, and generate a proprietary model adapted to the characteristics of a specific environment, specifically comprising: receiving the seed model transmitted by the cloud server, parsing the seed model into three parts: a weight matrix, a description of a hidden danger feature extraction path, and a pre-training task parameter configuration, and performing a hierarchical feature consistency check; using the environmental and hidden danger characteristics of the power overhead line to generate a self-supervised training signal; performing hierarchical feature extraction on the intrinsic structure and characteristics of the historical multi-dimensional image data itself; inputting the extracted features and the self-supervised training signal into the seed model, adjusting the weight and training strategy according to the real-time feedback of the seed model, dynamically optimizing the parameters of the seed model, and obtaining a proprietary model adapted to the characteristics of the environment at different locations; The model inference unit is used to apply the proprietary model to the real-time multi-dimensional image data to identify hidden dangers, obtain hidden danger identification results, and transmit the hidden danger identification results to the visual monitoring device for visual display and early warning; The receiving of the seed model transmitted by the cloud server parses the seed model into three parts: weight matrix, hidden danger feature extraction path description, and pre-training task parameter configuration, and performs hierarchical feature consistency verification, specifically including: Compare the characteristic response pattern of the weight matrix with the historical characteristic library of power overhead lines to match the historical high-risk hidden danger areas; Calculate the statistical indicators of the output feature distribution of each level of convolution kernel, including mean, variance and kurtosis, and use the chi-square distribution fitting method to compare the convolution output with the hidden danger feature distribution under the specific environment of the power line; In the first layer, it is verified whether the main target area of ​​the convolution kernel output covers specific hidden dangers; in the second layer, it is evaluated whether the features are robust to environmental factors, and the stability of hidden danger feature extraction is verified by superimposing features of different conditions; in the third layer, it is checked whether global features and local features appear in the detection results at the same time to ensure the comprehensive adaptability of the model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, an intelligent monitoring method for power overhead lines based on self-supervision as described in any one of claims 1-5 is implemented.

Images