Line insulator health state analysis method based on Transform technology

Abstract

The invention relates to a line insulator health state analysis method based on a Transform technology, and belongs to the technical field of insulator monitoring. Comprising the steps of identifying input features from images and video data polled by an electric power line unmanned aerial vehicle, analyzing pixel distribution and colors of the images, calling an encoder assembly of a Transform technology, executing sequence embedding conversion, and obtaining feature embedding representation. According to the method, classification and regional clustering of the insulator state are carried out by combining a decoding technology, the monitoring precision and the response speed of the insulator state are improved, the change of the insulator state is predicted through deep learning of features and trend analysis of a time sequence, the maintenance period is effectively planned, and potential faults are prevented in advance; potential insulator operation abnormity can be identified more accurately, power failure events are reduced, and the stability and reliability of power grid operation are improved through real-time monitoring and dynamic threshold judgment.

Description

Technical Field

[0001] The invention belongs to the technical field of insulator monitoring, and in particular relates to a line insulator health status analysis method based on Transformer technology. Background Art

[0002] As a key support and insulation component in power lines, the status of insulators directly affects the reliability and safety of the power grid. The field of insulator monitoring technology involves the use of various methods and methods to evaluate and ensure the performance and safety of insulators in power transmission lines. The key is to prevent power failures and ensure the stable operation of the power grid.

[0003] Since the insulators used in the power system are exposed to the harsh environment of sun, rain, high temperature, strong wind, electromechanical load, etc. for a long time during operation, it is easy to have surface damage, resistance reduction, internal cracks, flashover and even breakdown of the insulators, which seriously affects the safety of the power system. Therefore, it is necessary to regularly analyze and evaluate the health status of the insulators to ensure their reliable operation. Existing technical methods include offline detection or online detection methods. Offline detection often requires power outage to remove the equipment for detection, which affects people's production and life, has low efficiency, and is mostly used for factory detection. Traditional online monitoring methods include ultraviolet, electromagnetic wave, pulse current, leakage current detection and other methods, which do not require power outages, but usually require the installation of equipment on the line, which is also inconvenient. Using drones equipped with visible light + infrared + ultraviolet lenses, the risk of insulation performance degradation or damage can be discovered in time through periodic inspections, especially through advanced algorithms and machine learning technologies for fault prediction and health status assessment, which has broad application prospects.

[0004] Insulator monitoring in existing technologies mainly relies on on-site detection and basic remote monitoring, which is limited in processing large-scale data and real-time analysis in complex environments. The lack of application of advanced algorithms limits the accuracy and efficiency of fault prediction and health status assessment. Traditional prediction methods are not sensitive enough to identify subtle changes in health status, resulting in the inability to detect problems in a timely manner, increasing maintenance costs and the risk of failure.

[0005] The line insulator statistical method based on Transformer technology refers to the application of the Transformer model to process and analyze the data of power line insulators for status monitoring and fault prediction. The powerful ability of Transformer is used to parse time series data and identify potential patterns and anomalies. The main purpose is to improve the operation efficiency and safety of the power grid through accurate data analysis, realize real-time monitoring of the health status of insulators, timely predict and prevent faults, reduce power outages and improve the overall reliability of electricity. Summary of the invention

[0006] The purpose of the present invention is to solve the shortcomings existing in the prior art and provide a line insulator health status analysis method based on Transformer technology.

[0007] To achieve the above object, the technical solution of the present invention is: a line insulator health status analysis method based on Transformer technology, comprising:

[0008] Identify input features from images and video data of power lines, analyze the pixel distribution and color of the images, and call the encoder component based on Transformer technology to perform sequence embedding transformation to obtain feature embedding representation;

[0009] Applying a multi-head self-attention mechanism to the feature embedding representation, calculating the attention weight by comparing and judging the correlation between the differentiated sequence elements, and obtaining a weighted feature map;

[0010] Through the weighted feature map, a decoder component based on Transformer technology is used to perform sequence-to-sequence decoding, pass the decoding vector through a multi-layer network, adjust and optimize the output layer by layer, and obtain a decoding feature sequence;

[0011] Based on the decoded feature sequence, classify the insulator state, compare the decoded feature with a preset state threshold, determine the state category of the insulator by the threshold, and generate an insulator health state distribution;

[0012] Using the insulator health status distribution, cluster analysis is performed on the insulator health status according to the power grid area, and the number and type of insulators are counted by area to obtain regional insulator statistics;

[0013] According to the statistical results of insulators in the area, trend analysis is performed, and the change of insulator status is predicted by time series analysis method. Through maintenance and real-time status records, the expected maintenance schedule of insulators is obtained.

[0014] In one embodiment of the present invention, the feature embedding representation includes a numerical sequence converted from pixel-level color, brightness and texture features, the weighted feature map includes feature maps of differentiated image regions re-weighted by criticality, the location and status of power lines and insulators, the decoded feature sequence includes a continuous feature vector sequence processed by a decoder component and feature adjustment and optimization from input image to output state classification, the insulator health status distribution includes health status categories of differentiated insulators: normal, moderately damaged and severely damaged, the regional insulator statistics include the number of insulators in multiple health states in each power grid area, type statistics and the overall status of regional equipment, and the expected insulator maintenance schedule includes a schedule, recommended maintenance and replacement cycles formulated based on predictive analysis of insulator state changes.

[0015] In one embodiment of the present invention, the steps of identifying input features from the image and video data of the power line, analyzing the pixel distribution and color of the image, and calling the encoder component based on the Transformer technology to perform sequence embedding conversion to obtain the feature embedding representation are as follows:

[0016] Extract raw pixel data from images and video data of power lines, identify and separate differentiated visual elements based on color intervals, optimize element differentiation through brightness threshold setting, and generate raw feature data;

[0017] Using the original feature data, converting image pixels into a numerical sequence, adjusting the sampling rate and resolution parameters in the sequence, optimizing processing efficiency through data compression and re-indexing, and generating serialized feature data;

[0018] The serialized feature data is used to perform feature embedding conversion. By setting the size and parameters of the embedding layer and utilizing an encoder component based on Transformer technology, the image feature vector is standardized to adapt to differentiated processing requirements and generate a feature embedding representation.

[0019] In one embodiment of the present invention, a multi-head self-attention mechanism is applied to the feature embedding representation, and the weight of attention is calculated by comparing and judging the correlation between the differentiated sequence elements. The steps of obtaining the weighted feature map are specifically as follows:

[0020] Based on the feature embedding representation, performing a difference comparison between elements, evaluating the difference of elements in the feature sequence, and setting a correlation parameter to measure the correlation of the differentiated features, thereby generating difference evaluation data;

[0021] The difference evaluation data is used to perform weight calculation using a genetic algorithm, and the influence of key features is optimized by adjusting the weight distribution strategy to generate a weight distribution record;

[0022] According to the weight distribution record, the calculated weights are applied to the original feature embedding representation, and the consistency and efficiency of key features are verified by weight reshaping and optimizing features to generate a weighted feature map.

[0023] In one embodiment of the present invention, the formula of the genetic algorithm is as follows:

[0024]

[0025] Among them, f(w) is the fitness function, w is the main weight vector, and w i is the main weight of the i-th feature, v is the auxiliary weight vector, and v i is the auxiliary weight of the i-th feature, e i is the evaluation index of the i-th feature, α is the main weight coefficient, and β is the auxiliary weight coefficient.

[0026] In one embodiment of the present invention, the weighted feature map is used to perform sequence-to-sequence decoding using a decoder component based on Transformer technology, and the decoding vector is passed through a multi-layer network, and the output is adjusted and optimized layer by layer to obtain a decoding feature sequence. Specifically, the steps are as follows:

[0027] Using the weighted feature map and the decoder component of the Transformer technology, sequence-to-sequence decoding is performed to convert the weighted feature map into an initial decoding vector, and initial decoding parameters are set to start the conversion of the sequence to generate an initial decoding vector;

[0028] According to the initial decoding vector, the decoding vector is passed through the multi-layer network layer by layer, and the processing parameters of each layer of the network are adjusted, including the selection of learning rate and activation function, so as to optimize the data flow and feature representation in decoding and generate an optimized decoding vector;

[0029] By using the optimized decoding vector, feature adjustment and output formatting are performed through multi-layer network processing, the consistency of the decoding sequence with the real-time changing processing requirements is checked, and a decoding feature sequence is generated.

[0030] In one embodiment of the present invention, based on the decoded feature sequence, the insulator state is classified, the decoded feature is compared with a preset state threshold, the state category of the insulator is determined by the threshold, and the steps of generating the insulator health state distribution are specifically as follows:

[0031] Initialize the state classification of the insulator based on the decoded feature sequence, map the decoded feature sequence to a preset state threshold, set the classification threshold standard, perform initialization determination of the insulator state, and generate state mapping data;

[0032] Using the state mapping data, performing threshold comparison on the features of the differentiated insulators, judging the state of the insulators by the set thresholds, applying the threshold judgment logic to distinguish the degree of damage, and generating an insulator state judgment result;

[0033] The insulator state judgment result is used to perform state classification, integrate the state information of the insulator, create a state distribution chart, identify the distribution of insulators in differentiated states in a visual form, and generate an insulator health state distribution.

[0034] In one embodiment of the present invention, the insulator health status distribution is used to perform cluster analysis on the health status of the insulators according to the power grid area, and the number and type of statistics are divided by area to obtain the regional insulator statistics results. Specifically, the steps are as follows:

[0035] Based on the insulator health status distribution, the insulators are grouped according to the power grid area, and the boundary and classification basis of each area are set to generate regional grouping data;

[0036] The regional grouping data is used, a cluster analysis method is used to count the number and type of insulators in differentiated health states in each region, and an initialization regional analysis result is generated by adjusting the counting method and type identification standard;

[0037] Based on the initialization regional analysis results, data from multiple regions are aggregated and verified, the integrity and synchronization of the data are checked, and regional insulator statistics are generated.

[0038] In one embodiment of the present invention, the formula of the cluster analysis method is as follows:

[0039]

[0040] Among them, J′ is the minimization of the cluster center, x i,z For each insulator health status data, c j,z is the health status cluster center, n is the total number of insulators z in the region, k is the preset number of health status categories, and w j,z is the weight coefficient of the jth health state, p j,z is the distance index, λ z is the penalty coefficient, c prev,j,z is the cluster center of the previous iteration.

[0041] In one embodiment of the present invention, the steps of performing trend analysis based on the regional insulator statistical results, using a time series analysis method to predict insulator state changes, and obtaining an expected insulator maintenance schedule through maintenance and real-time state records are specifically as follows:

[0042] Using the regional insulator statistical results, time series analysis is performed, historical maintenance records and real-time status records are collected, and the maintenance records are correlated with the insulator status to generate a time series analysis data set;

[0043] Utilizing the time series analysis data set, analyzing the trend of data changes, calculating the change trend of the insulator state by setting the window size and step size of the model parameters, predicting the changes in the short term by data pattern recognition, and generating a state change trend prediction result;

[0044] According to the state change trend prediction results, the maintenance cycle and time point of the insulator are set according to the predicted state change, the maintenance cycle is adjusted in real time to optimize the maintenance plan, and the expected maintenance schedule of the insulator is generated.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. In the present invention, by extracting key features from the image and video data of the power line and using the multi-head self-attention mechanism to conduct in-depth analysis of these features, the various states of the insulators can be understood more accurately, the processing capability and feature interpretability of complex data can be significantly improved, and the insulator state monitoring is made more accurate, which has high technical innovation significance and practical value.

[0047] 2. The present invention utilizes decoding technology and combines regional clustering methods to classify the status of insulators on power lines, which can quickly locate problem areas in large-scale data, improve the response speed and accuracy of insulator monitoring, combine classification with clustering, and enhance operational efficiency and fault prevention capabilities.

[0048] 3. The present invention identifies potential faults and maintenance needs in advance through deep learning analysis of insulator status characteristics and trend prediction of time series data, which not only optimizes maintenance cycle planning, but also reduces economic losses caused by sudden faults, making power grid maintenance more forward-looking and economically efficient.

[0049] 4. The present invention combines real-time monitoring and dynamic threshold judgment technology to enhance the continuous monitoring capability of the power grid operation status. Through real-time data analysis, it can automatically adjust the threshold and respond to changes in the power grid status, effectively improving the stability and security of the power grid, and playing a significant role in promoting the stability and economic benefits of power grid operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the main steps of the present invention;

[0051] Figure 2 It is a schematic diagram of the refinement of S1 of the present invention;

[0052] Figure 3It is a schematic diagram of the refinement of S2 of the present invention;

[0053] Figure 4 It is a schematic diagram of the refinement of S3 of the present invention;

[0054] Figure 5 It is a schematic diagram of the refinement of S4 of the present invention;

[0055] Figure 6 It is a detailed schematic diagram of S5 of the present invention;

[0056] Figure 7 It is a detailed schematic diagram of S6 of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0058] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0059] The present invention provides a line insulator health status analysis method based on Transformer technology, comprising:

[0060] Identify input features from images and video data of power lines, analyze the pixel distribution and color of the images, and call the encoder component based on Transformer technology to perform sequence embedding transformation to obtain feature embedding representation;

[0061] Applying a multi-head self-attention mechanism to the feature embedding representation, calculating the attention weight by comparing and judging the correlation between the differentiated sequence elements, and obtaining a weighted feature map;

[0062] Through the weighted feature map, a decoder component based on Transformer technology is used to perform sequence-to-sequence decoding, pass the decoding vector through a multi-layer network, adjust and optimize the output layer by layer, and obtain a decoding feature sequence;

[0063] Based on the decoded feature sequence, classify the insulator state, compare the decoded feature with a preset state threshold, determine the state category of the insulator by the threshold, and generate an insulator health state distribution;

[0064] Using the insulator health status distribution, cluster analysis is performed on the insulator health status according to the power grid area, and the number and type of insulators are counted by area to obtain regional insulator statistics;

[0065] According to the statistical results of insulators in the area, trend analysis is performed, and the change of insulator status is predicted by time series analysis method. Through maintenance and real-time status records, the expected maintenance schedule of insulators is obtained.

[0066] The following are specific implementation examples of the present invention.

[0067] See also Figure 1 An example of the present invention provides a line insulator health status analysis method based on Transformer technology, comprising the following steps:

[0068] S1: Identify input features from the image and video data of the power line, analyze the pixel distribution and color and record features of the image, and call the encoder component based on Transformer technology to perform sequence embedding conversion to obtain feature embedding representation;

[0069] S2: Apply a multi-head self-attention mechanism to the feature embedding representation, calculate the attention weight by comparing and judging the correlation between the differentiated sequence elements, and obtain the weighted feature map;

[0070] S3: Through the weighted feature map, the decoder component based on Transformer technology is used to perform sequence-to-sequence decoding, pass the decoding vector through the multi-layer network, adjust and optimize the output layer by layer, and obtain the decoded feature sequence;

[0071] S4: Based on the decoded feature sequence, classify the insulator state, compare the decoded feature with the preset state threshold, judge the state category of the insulator by the threshold, and generate the insulator health state distribution;

[0072] S5: Use the distribution of insulator health status to perform regional statistics, cluster the insulator health status according to the power grid area, and divide the statistics of quantity and type by region to obtain the regional insulator statistics results;

[0073] S6: Based on the regional insulator statistics, trend analysis is performed, and the change of insulator status is predicted using time series analysis methods. The expected maintenance schedule of insulators is obtained through maintenance records and real-time status records.

[0074] The feature embedding representation includes the numerical sequence converted from the pixel-level color, brightness and texture features. The weighted feature map includes the feature map of the differentiated image area re-weighted by criticality, the location and status of the power lines and insulators. The decoded feature sequence includes the continuous feature vector sequence processed by the decoder component and the feature adjustment and optimization from the input image to the output state classification. The insulator health status distribution includes the health status categories of differentiated insulators: normal, moderately damaged and severely damaged. The regional insulator statistics include the number and type statistics of insulators in various health states in each power grid area and the overall status of the regional equipment. The expected insulator maintenance schedule includes the schedule, recommended maintenance and replacement cycles formulated based on the predictive analysis of insulator state changes.

[0075] See also Figure 2 , identify input features from the image and video data of the power line, analyze the pixel distribution and color of the image, and call the encoder component based on Transformer technology to perform sequence embedding conversion. The specific steps to obtain the feature embedding representation are as follows:

[0076] S101: extracting raw pixel data from the image and video data of the power line, identifying and separating differentiated visual elements according to color intervals, optimizing the differentiation of elements by setting brightness thresholds, and generating raw feature data in the following execution flow;

[0077] Extract raw pixel data from the image and video data of power lines, identify and separate visually differentiated elements, such as different lines or background features, by analyzing the color range of the data. Optimize the distinction of elements by setting a suitable brightness threshold, make the visual features more obvious, and generate raw feature data. The formula used is:

[0078]

[0079] Among them, T a Represents the original feature data, P ia is the data of the ath pixel, L ia is the brightness threshold, n a is the total number of pixels, and extr represents the extraction function.

[0080] S102: Using the original feature data, converting the image pixels into a numerical sequence, adjusting the sampling rate and resolution parameters in the sequence, optimizing the processing efficiency through data compression and re-indexing, and generating the serialized feature data. The execution flow is as follows;

[0081] Using the original feature data, the image pixels are converted into a numerical sequence, and the sampling rate and resolution parameters in the sequence are adjusted to meet different analysis and storage requirements. By applying data compression technology and re-indexing methods, the processing efficiency is optimized, the data storage space is reduced, and the data processing speed is improved to generate serialized feature data. The formula used is:

[0082]

[0083] Among them, D b Represents serialized feature data, S jb is the original data of the ath pixel, R jb is the resolution parameter, n a is the total number of pixels, and trans represents the transformation function.

[0084] S103: Using serialized feature data, perform feature embedding conversion, set the size and parameters of the embedding layer, use the encoder component based on Transformer technology to standardize the image feature vector, adapt to the differentiated processing requirements, and generate the feature embedding representation. The execution process is as follows;

[0085] Using serialized feature data, we perform feature embedding conversion. By setting the size and parameters of the embedding layer and using the encoder component based on Transformer technology, we can standardize the image feature vector to meet differentiated processing requirements, such as different image recognition or classification tasks, and generate feature embedding representations. The formula used is:

[0086]

[0087] Among them, F c represents the feature embedding representation, V kc is the serialized feature of the kth data point, E kc is the embedding parameter, p c is the total number of data points, and embed represents the embedding function.

[0088] See also Figure 3 , a multi-head self-attention mechanism is applied to the feature embedding representation, and the attention weight is calculated by comparing and judging the correlation between the differentiated sequence elements. The specific steps for obtaining the weighted feature map are:

[0089] S201: Based on the feature embedding representation, the difference between elements is compared, the difference of elements in the feature sequence is evaluated, and the correlation parameter is set to measure the correlation of the differentiated features. The execution process of generating the difference evaluation data is as follows;

[0090] Based on feature embedding representation, we compare the differences between elements, conduct detailed difference evaluation on the elements in the feature sequence, and use correlation parameters to measure the correlation between different features to identify and highlight features with significant differences. This not only helps us understand the uniqueness and interaction of each element in the data, but also provides important basic data for subsequent feature processing and analysis, and generates element difference evaluation data.

[0091] S202: Using the differential evaluation data and the genetic algorithm to perform weight calculation, the influence of the key features is optimized by adjusting the weight distribution strategy, and the execution process of generating the weight distribution record is as follows;

[0092] Using differential evaluation data and genetic algorithms for weight calculation, the genetic algorithm simulates the process of natural selection, adjusts the weight distribution strategy, optimizes the influence of key features, and can effectively highlight the most critical features for analysis and decision-making. It not only enhances the accuracy of data processing, but also improves the efficiency of data analysis and generates weight distribution records.

[0093] The formula of genetic algorithm is as follows:

[0094]

[0095] Among them, f(w) is the fitness function, w is the main weight vector, and w i is the main weight of the i-th feature, v i is the auxiliary weight of the i-th feature, e i is the evaluation index of the i-th feature, α is the main weight coefficient, and β is the auxiliary weight coefficient.

[0096] The specific values ​​of the weight vectors w and v are determined based on expert evaluation or machine learning algorithms to ensure that they reflect the importance of the features and their contribution to the results. i It can be obtained through data monitoring and statistical methods, reflecting the performance or quality of the characteristics.

[0097] Parameter value setting:

[0098] A specific example is used to illustrate the calculation process of the formula. There are three characteristics set up, and the evaluation index e is collected through monitoring equipment to calculate the average value based on the data of the most recent month: e1 = 0.8, e2 = 0.5, e3 = 0.3.

[0099] The main weight w and auxiliary weight v are provided by a pre-trained machine learning model to ensure scientificity and accuracy. The weight values ​​are set as follows: w1=0.5, w2=0.3, w3=0.2, v1=0.4, v2=0.35, v3=0.25.

[0100] The coefficients α and β were determined through data analysis to reflect the influence of different feature combinations, and the set values ​​were as follows: α = 0.6, β = 0.4.

[0101] Calculation process:

[0102] 1. Calculate the contribution of the sovereign weight:

[0103]

[0104] 2. Calculate the contribution of auxiliary weights:

[0105]

[0106] 3. Calculate the fitness function value:

[0107] f(w)=α×0.61+β×0.366=0.6×0.61+0.4×0.366

[0108] =0.366+0.1464

[0109] =0.512

[0110] Result description:

[0111] The numerical result 0.5124 represents the fitness score under the given weight and evaluation index. This result shows the performance of the parameter combination in the model. A higher fitness score means a better parameter combination, which guides the adjustment of parameters and the formulation of strategies in the subsequent optimization steps.

[0112] S203: According to the weight distribution record, the calculated weight is applied to the original feature embedding representation, and the consistency and efficiency of the key features are verified by weight reshaping and optimizing the features. The execution process of generating the weighted feature map is as follows;

[0113] According to the weight distribution record, the calculated weights are applied to the original feature embedding representation, and the features are optimized through weight reshaping to verify the consistency and efficiency of key features. Through weighted adjustment, the performance of features in the model is optimized, the model's ability to capture key information is enhanced, and the feature representation after weight optimization is displayed to provide a more accurate and effective data representation for the next step of analysis and decision-making, and generate a weighted feature map using the formula:

[0114]

[0115] Among them, F g represents the weighted feature map, E kg is the original embedding representation of the g-th feature, W kg is the corresponding weight, p g is the total number of features, and reweight represents the reshaping weight function.

[0116] See also Figure 4 , through the weighted feature map, using the decoder component of the Transformer technology, performing sequence-to-sequence decoding, passing the decoded vector through the multi-layer network, adjusting and optimizing the output layer by layer, and obtaining the decoded feature sequence. The specific steps are:

[0117] S301: Using the weighted feature map and the decoder component of the Transformer technology, sequence-to-sequence decoding is performed, the weighted feature map is converted into an initial decoding vector, and the initial decoding parameters are set to start the conversion of the sequence. The execution process of generating the initial decoding vector is as follows;

[0118] Using weighted feature maps, the decoder component of Transformer technology is used for sequence-to-sequence decoding. This includes converting weighted feature maps into initial decoding vectors, which provide the basis for subsequent data processing. The sequence conversion is started by setting the initial decoding parameters, which include the start marker of decoding, the initialization of the attention mechanism, etc. The initial decoding vector is generated by converting the feature representation of image or video data into a format that is more suitable for specific applications, such as further data analysis or machine learning training, using the formula:

[0119]

[0120] Among them, V h represents the initial decoding vector, G ih is the hth weighted feature map, P ih is the initial decoding parameter, n h is the total number of weighted features, and decode represents the decoding function.

[0121] S302: According to the initial decoding vector, the decoding vector is passed through the multi-layer network layer by layer, and the processing parameters of each layer of the network are adjusted, including the selection of learning rate and activation function, and the data flow and feature representation in decoding are optimized. The execution process of generating the optimized decoding vector is as follows;

[0122] Based on the initial decoding vector, the decoding vector is passed through the multi-layer network layer by layer. The processing parameters of each layer of the network are carefully adjusted, including the selection of appropriate learning rate and activation function. The adjustment aims to optimize the data flow and feature representation in decoding, ensure that each layer of the network can effectively process and convert data, and enhance the performance of the entire network. Through layer-by-layer processing and optimization, an optimized decoding vector is generated, using the formula:

[0123]

[0124] Among them, D k represents the optimized decoding vector, X jkis the kth initial decoding vector, Y jk is the processing parameter of each layer of the network, m k is the total number of vectors, and transform represents the data transformation function.

[0125] S303: Through the optimized decoding vector, the features are adjusted and the output formatted through multi-layer network processing, and the consistency of the decoding sequence with the real-time changing processing requirements is checked. The execution process of generating the decoding feature sequence is as follows;

[0126] Through the optimized decoding vector, multi-layer network processing is performed to adjust the features and format the output. The core is to check the consistency of the decoding sequence with the real-time changing processing requirements to ensure that the output data format meets the expected application requirements. Through meticulous feature adjustment and formatting, the sequence is formatted into a specific format to meet different data usage scenarios, such as further data analysis, visualization or remaining data processing tasks, to generate a decoding feature sequence, using the formula:

[0127]

[0128] Among them, S l represents the decoding feature sequence, Z kl is the lth optimized decoding vector, W kl is the feature adjustment parameter, p c is the total number of data points, and finalize represents the processing and formatting function.

[0129] See also Figure 5 Based on the decoded feature sequence, the insulator state is classified, the decoded feature is compared with the preset state threshold, and the state category of the insulator is judged by the threshold. The steps of generating the insulator health state distribution are as follows:

[0130] S401: Based on the decoded feature sequence, initialize the state classification of the insulator, map the decoded feature sequence to a preset state threshold, set the classification threshold standard, perform initialization judgment on the insulator state, and generate the state mapping data. The execution process is as follows;

[0131] Based on the decoded feature sequence, the state classification of the insulator is initialized. The process includes mapping the decoded feature sequence to the preset state threshold and setting the classification threshold standard to initialize the insulator state. By identifying whether the insulator has potential damage or abnormality and accurately setting the threshold, the accuracy and consistency of the state judgment are ensured. After the operation is completed, the state mapping data is generated, and the formula used is:

[0132]

[0133] Among them, M mIndicates the state mapping data, F im is the mth decoding feature sequence, T im is the preset state threshold, p c is the total number of data points, and init represents the initialization function.

[0134] S402: Using the state mapping data, performing threshold comparison on the features of the differentiated insulators, judging the state of the insulators by the set thresholds, applying the threshold judgment logic to distinguish the degree of damage, and generating the insulator state judgment result. The execution process is as follows;

[0135] Using the state mapping data, the threshold comparison is performed on the features of differentiated insulators, the state of the insulator is judged by the set threshold, and the threshold judgment logic is applied to distinguish the degree of damage of the insulator. The key is to accurately apply the threshold logic to ensure that the state assessment of each insulator is both accurate and consistent, and to record the state and degree of damage of each insulator in detail to provide a basis for subsequent repair and maintenance, and to generate the insulator state judgment result. The formula used is:

[0136]

[0137] Among them, J o Indicates the result of insulator status judgment, S jo Mapping data for the oth state, U jo is the threshold standard, p c is the total number of data points, and compare represents the comparison function.

[0138] S403: Using the insulator state judgment result, classifying the state, integrating the insulator state information, creating a state distribution chart, identifying the distribution of insulators in differentiated states in a visual form, and generating the insulator health state distribution. The execution process is as follows;

[0139] Using the insulator status judgment results, the insulator status information is integrated to create a status distribution chart, which can identify the distribution of insulators in differentiated states in a visual form. It not only helps to quickly identify damaged areas, but also provides an intuitive way to show the health status of insulators. It provides important visual information for the maintenance team to help develop more effective maintenance strategies and preventive measures, and generates an insulator health status distribution chart using the formula:

[0140]

[0141] Among them, D p represents the insulator health status distribution, R kp is the initial judgment result of the pth state, V kp is the visualization parameter, p cis the total number of data points, and categorize represents the classification function.

[0142] See also Figure 6 , using the insulator health status distribution, cluster analysis is performed on the insulator health status according to the power grid area, and the number and type of statistics are divided by region. The specific steps to obtain the regional insulator statistical results are as follows:

[0143] S501: Based on the distribution of insulator health status, the insulators are grouped according to the power grid area, and the boundary and classification basis of each area are set. The execution process of generating regional grouping data is as follows;

[0144] Based on the distribution of insulator health status, insulators are grouped according to the geographical areas of the power grid, and clear boundaries and classification criteria are set for each area. The purpose is to conduct a more detailed regional analysis of the insulator status based on geographical location, operation history or maintenance records, and provide support for regional specific maintenance strategies. Regional grouping data is generated using the formula:

[0145]

[0146] Among them, G r Indicates regional grouping data, H ir is the health status distribution of the rth insulator, B ir is the region boundary parameter, p c is the total number of data points, and group represents the grouping function.

[0147] S502: using regional grouping data and cluster analysis methods, counting the number and type of insulators in differentiated health states in each region, and generating the initialization regional analysis results by adjusting the counting method and type identification standard. The execution process is as follows;

[0148] Using regional grouping data, cluster analysis methods are used to count the number and types of insulators in different health states in each region. By adjusting the counting method and type identification criteria, it is not only helpful to identify the main problem types in the region, but also to evaluate the maintenance needs of specific areas. Statistics provide quantitative data support, which helps to understand the urgency and focus of regional maintenance and generate initial regional analysis results.

[0149] The formula of the cluster analysis method is as follows:

[0150]

[0151] Among them, J′ is the minimization of the cluster center, x i,z is the health status data of insulator z, obtained from the detection instrument. The data is obtained by recording the readings of the detection equipment in real time. j,zis the cluster center of the jth health status, which is determined by the center update algorithm of K-means cluster analysis, n is the total number of insulators z in the region, k is the preset number of health status categories, and w j,z is the weight coefficient corresponding to the jth health state, which is calculated based on the frequency of occurrence of each health state and reflects the importance of different health states. j,z is the distance index, which is determined based on the distribution characteristics of the health status data. If the data distribution is relatively concentrated, a smaller index is selected, λ z is the penalty coefficient, which is adjusted according to the performance of the model on the validation dataset to optimize the generalization ability of the model. prev,j,z is the cluster center of the previous iteration, which comes from the cluster center value calculated last time.

[0152] Parameter value setting:

[0153] There are 2 insulators in the area (n=2), and 2 health status categories (k=2). The specific values ​​are as follows and calculated:

[0154] Set insulator health status data x i,1 =[10,20],x i ,2=[15,25].

[0155] Initial cluster center c 1,1 =[12,22],c 2,1 =[17,27].

[0156] According to the frequency of occurrence of various health states, the weight coefficient is obtained: w 1,1 =0.6, w 2,1 =0.4.

[0157] According to the data distribution, the distance index is: P 1,1 =2,P 2,1 =2.

[0158] The penalty coefficient is obtained based on model optimization: λ1=0.05.

[0159] Calculation process:

[0160]

[0161]

[0162] 3: Penalty term: λ1∥c 1,1 -c prev,1,1 ∥ 2

[0163] =0.05×0

[0164] =0

[0165] The above calculation results are added together to obtain the value of J′. The results show the stability of the model and the accuracy of health status clustering. The significance of this value for model optimization is further analyzed.

[0166] S503: According to the initialization regional analysis results, the data of multiple regions are summarized and verified, the integrity and synchronization of the data are checked, and the execution process of generating regional insulator statistical results is as follows;

[0167] According to the initial regional analysis results, the data of multiple regions are summarized and verified, and the integrity and synchronization of the data are checked to ensure that the statistical information accurately reflects the actual situation of each region. The report shows the health status and distribution of insulators in each region in detail, providing a basis for decision makers. Through the statistical report, the statistical results of regional insulators are generated, and the formula used is:

[0168]

[0169] Among them, S t Represents the regional insulator statistics, V kt is the kth t Initialization area analysis results of the area, D kt is the data validation parameter, m s is the total number of regions, and summarize represents the data summary function.

[0170] See also Figure 7 ,According to the statistical results of regional insulators, trend analysis is performed, and the time series analysis method is used to predict the change of insulator status. Through maintenance and real-time status records, the steps to obtain the expected maintenance schedule of insulators are as follows:

[0171] S601: Use the regional insulator statistical results to perform time series analysis, collect historical maintenance records and real-time status records, and perform correlation analysis between the maintenance records and the insulator status to generate a time series analysis data set. The execution process is as follows;

[0172] Using regional insulator statistics, time series analysis is performed, mainly collecting historical maintenance records and real-time status records, and correlating the data with the health status of the insulators. This helps understand the changing patterns of insulator status over time and identify key factors affecting insulator performance. By integrating historical and real-time data, a time series analysis data set is generated, using the formula:

[0173]

[0174] Among them, T u represents the time series analysis dataset, M iu is the maintenance record of the uth insulator, Siu is the corresponding state record, p c is the total number of data points, and correlate represents the correlation analysis function.

[0175] S602: Analyze the data set using the time series, analyze the trend of data changes, calculate the trend of insulator state changes by setting the window size and step size of the model parameters, predict the changes in the short term by data pattern recognition, and generate the state change trend prediction result. The execution process is as follows;

[0176] The time series analysis data set is used to analyze the trend of data changes. By setting the window size and step size of the model parameters, the trend calculation of the insulator state is performed. During the calculation process, the data pattern recognition technology is used to predict the state changes in the short term to foresee potential maintenance needs or failures, and generate the state change trend prediction results. The formula used is:

[0177]

[0178] Among them, P v Indicates the state change trend prediction result, X jv is the time series analysis data of the vth data point, Y jv is the model parameter, p c is the total number of data points, and trend represents the trend calculation function.

[0179] S603: According to the state change trend prediction result, the maintenance cycle and time point of the insulator are set according to the predicted state change, and the maintenance cycle is adjusted in real time to optimize the maintenance plan. The execution process of generating the expected maintenance schedule of the insulator is as follows;

[0180] According to the prediction results of the state change trend, the maintenance cycle and time point of the insulator are set, and the maintenance cycle is adjusted in real time according to the predicted state change to optimize the maintenance plan. Ensure that maintenance activities can be performed when they are most needed, maximize efficiency and effectiveness, reduce downtime and maintenance costs, and generate an expected maintenance schedule for insulators. The formula used is:

[0181]

[0182] Among them, S w represents the expected maintenance schedule for insulators, Q kw is the state prediction result of the kth insulator, R kw is the adjusted maintenance period, p w is the total number of insulators, and schedule represents the scheduling function.

[0183] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A line insulator health status analysis method based on Transformer technology, characterized in that: include: Identify input features from image and video data of power line drone inspections, analyze the pixel distribution and color of the image, and call the encoder component based on Transformer technology to perform sequence embedding conversion to obtain feature embedding representation; Applying a multi-head self-attention mechanism to the feature embedding representation, calculating the attention weight by comparing and judging the correlation between the differentiated sequence elements, and obtaining a weighted feature map; Through the weighted feature map, a decoder component based on Transformer technology is used to perform sequence-to-sequence decoding, pass the decoding vector through a multi-layer network, adjust and optimize the output layer by layer, and obtain a decoding feature sequence; Based on the decoded feature sequence, classify the insulator state, compare the decoded feature with a preset state threshold, determine the state category of the insulator by the threshold, and generate an insulator health state distribution; Using the insulator health status distribution, cluster analysis is performed on the insulator health status according to the power grid area, and the number and type of insulators are counted by area to obtain regional insulator statistics; According to the statistical results of insulators in the area, trend analysis is performed, and the change of insulator status is predicted by time series analysis method. Through maintenance and real-time status records, the expected maintenance schedule of insulators is obtained.

2. The method for analyzing the health status of line insulators based on Transformer technology according to claim 1, characterized in that: The feature embedding representation includes a numerical sequence converted from pixel-level color, brightness and texture features, the weighted feature map includes feature maps of differentiated image regions re-weighted by criticality, the location and status of power lines and insulators, the decoded feature sequence includes a continuous feature vector sequence processed by a decoder component and feature adjustment and optimization from input image to output state classification, the insulator health status distribution includes health status categories of differentiated insulators: normal, moderately damaged and severely damaged, the regional insulator statistics include the number of insulators in multiple health states in each power grid area, type statistics and the overall status of regional equipment, and the expected insulator maintenance schedule includes a schedule, recommended maintenance and replacement cycles formulated based on predictive analysis of insulator state changes.

3. The method for analyzing the health status of line insulators based on Transformer technology according to claim 1, characterized in that: The steps to identify input features from the image and video data of the power line, analyze the pixel distribution and color of the image, and call the encoder component based on Transformer technology to perform sequence embedding conversion to obtain the feature embedding representation are as follows: Extract raw pixel data from images and video data of power lines, identify and separate differentiated visual elements based on color intervals, optimize element differentiation through brightness threshold setting, and generate raw feature data; Using the original feature data, converting image pixels into a numerical sequence, adjusting the sampling rate and resolution parameters in the sequence, optimizing processing efficiency through data compression and re-indexing, and generating serialized feature data; The serialized feature data is used to perform feature embedding conversion. By setting the size and parameters of the embedding layer and utilizing an encoder component based on Transformer technology, the image feature vector is standardized to adapt to differentiated processing requirements and generate a feature embedding representation.

4. The method for analyzing the health status of line insulators based on Transformer technology according to claim 1, characterized in that: The multi-head self-attention mechanism is applied to the feature embedding representation, and the weight of attention is calculated by comparing and judging the correlation between the differentiated sequence elements. The steps of obtaining the weighted feature map are as follows: Based on the feature embedding representation, performing a difference comparison between elements, evaluating the difference of elements in the feature sequence, and setting a correlation parameter to measure the correlation of the differentiated features, thereby generating difference evaluation data; The difference evaluation data is used to perform weight calculation using a genetic algorithm, and the influence of key features is optimized by adjusting the weight distribution strategy to generate a weight distribution record; According to the weight distribution record, the calculated weights are applied to the original feature embedding representation, and the consistency and efficiency of key features are verified by weight reshaping and optimizing features to generate a weighted feature map.

5. The method for analyzing the health status of line insulators based on Transformer technology according to claim 4, characterized in that: The formula of the genetic algorithm is as follows: Among them, f(w) is the fitness function, w is the main weight vector, and w i is the main weight of the i-th feature, v is the auxiliary weight vector, and v i is the auxiliary weight of the i-th feature, e i is the evaluation index of the i-th feature, α is the main weight coefficient, and β is the auxiliary weight coefficient.

6. The method for analyzing the health status of line insulators based on Transformer technology according to claim 1, characterized in that: The weighted feature map is used to perform sequence-to-sequence decoding using a decoder component based on Transformer technology. The decoding vector is passed through a multi-layer network, and the output is adjusted and optimized layer by layer to obtain a decoding feature sequence. Specifically, the steps are as follows: Using the weighted feature map and the decoder component of the Transformer technology, sequence-to-sequence decoding is performed to convert the weighted feature map into an initial decoding vector, and initial decoding parameters are set to start the conversion of the sequence to generate an initial decoding vector; According to the initial decoding vector, the decoding vector is passed through the multi-layer network layer by layer, and the processing parameters of each layer of the network are adjusted, including the selection of learning rate and activation function, so as to optimize the data flow and feature representation in decoding and generate an optimized decoding vector; By using the optimized decoding vector, feature adjustment and output formatting are performed through multi-layer network processing, the consistency of the decoding sequence with the real-time changing processing requirements is checked, and a decoding feature sequence is generated.

7. The method for analyzing the health status of line insulators based on Transformer technology according to claim 1, characterized in that: Based on the decoded feature sequence, the insulator state is classified, the decoded feature is compared with a preset state threshold, and the state category of the insulator is determined by the threshold. The steps of generating the insulator health state distribution are specifically as follows: Initialize the state classification of the insulator based on the decoded feature sequence, map the decoded feature sequence to a preset state threshold, set the classification threshold standard, perform initialization determination of the insulator state, and generate state mapping data; Using the state mapping data, performing threshold comparison on the features of the differentiated insulators, judging the state of the insulators by the set thresholds, applying the threshold judgment logic to distinguish the degree of damage, and generating an insulator state judgment result; The insulator state judgment result is used to perform state classification, integrate the state information of the insulator, create a state distribution chart, identify the distribution of insulators in differentiated states in a visual form, and generate an insulator health state distribution.

8. The method for analyzing the health status of line insulators based on Transformer technology according to claim 1, characterized in that: The steps of using the insulator health status distribution to cluster the insulator health status according to the power grid area, and statistically analyzing the number and type of insulators by regional division to obtain the regional insulator statistical results are as follows: Based on the insulator health status distribution, the insulators are grouped according to the power grid area, and the boundary and classification basis of each area are set to generate regional grouping data; The regional grouping data is used, a cluster analysis method is used to count the number and type of insulators in differentiated health states in each region, and an initialization regional analysis result is generated by adjusting the counting method and type identification standard; Based on the initialization regional analysis results, data from multiple regions are aggregated and verified, the integrity and synchronization of the data are checked, and regional insulator statistics are generated.

9. The method for analyzing the health status of line insulators based on Transformer technology according to claim 8, characterized in that: The formula of the cluster analysis method is as follows: Among them, J′ is the minimization of the cluster center, x i,z For each insulator health status data, c j,z is the health status cluster center, n is the total number of insulators z in the region, k is the preset number of health status categories, and w j,z is the weight coefficient of the jth health state, p j,z is the distance index, λ z is the penalty coefficient, c prev,j,z is the cluster center of the previous iteration.

10. The method for analyzing the health status of line insulators based on Transformer technology according to claim 1, characterized in that: According to the statistical results of the regional insulators, trend analysis is performed, and the change of the insulator status is predicted by using the time series analysis method. The steps of obtaining the expected maintenance schedule of the insulators through maintenance and real-time status records are as follows: Using the regional insulator statistical results, time series analysis is performed, historical maintenance records and real-time status records are collected, and the maintenance records are correlated with the insulator status to generate a time series analysis data set; Utilizing the time series analysis data set, analyzing the trend of data changes, calculating the change trend of the insulator state by setting the window size and step size of the model parameters, predicting the changes in the short term by data pattern recognition, and generating a state change trend prediction result; According to the state change trend prediction results, the maintenance cycle and time point of the insulator are set according to the predicted state change, the maintenance cycle is adjusted in real time to optimize the maintenance plan, and the expected maintenance schedule of the insulator is generated.

Images