Internet of Things intelligent detection method for electric power system

Abstract

The invention discloses a power system Internet of Things intelligent detection method, and relates to the technical field of power detection, and the system comprises the following components: S1, data collection and preprocessing, S2, big data analysis and feature extraction, S3, equipment life prediction model construction and training, S4, fault prediction and evaluation, S5, predictive maintenance strategy making, and S6, adaptive adjustment. According to the method, the equipment life prediction model of deep neural network self-learning modeling is adopted, the quantum optimization algorithm is used for optimization, accurate prediction of the residual life and the fault occurrence probability of the power equipment is achieved, and meanwhile, the power equipment state comprehensive evaluation model based on the combination of fuzzy logic and the evidence theory is combined, so that the power equipment state comprehensive evaluation efficiency is improved. The equipment health state grade can be comprehensively evaluated, a scientific basis is provided for formulating a predictive maintenance strategy, a maintenance plan can be reasonably arranged in advance, the maintenance cost is reduced, the service life of equipment is prolonged, and economic benefits and social benefits of a power system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power detection, and in particular to an intelligent detection method for the Internet of Things of a power system. Background Art

[0002] With the increasing complexity and intelligent demand of power systems, real-time monitoring and accurate evaluation of the operating status of power equipment are particularly important. Traditional manual inspections and regular maintenance methods can no longer meet the current power system's needs for efficient and precise management.

[0003] There are deficiencies in traditional technologies. On the one hand, it is impossible to accurately predict the remaining life of power equipment and the probability of failure. On the other hand, it is impossible to comprehensively evaluate the health status level of equipment.

[0004] In summary, the traditional power equipment life prediction method can no longer meet the requirements of modern power systems for high-precision and intelligent management. In order to overcome these shortcomings, it is particularly important to develop an intelligent detection method for the Internet of Things of power systems. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art and provide an intelligent detection method for the Internet of Things of power systems. The method processes and analyzes data through algorithms and models, thereby realizing accurate prediction of the remaining life and failure probability of power equipment. At the same time, the method is also combined with a comprehensive evaluation model of the power equipment status based on fuzzy logic and evidence theory, which can comprehensively evaluate the health status level of the equipment and provide a scientific basis for the formulation of predictive maintenance strategies.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: an intelligent detection method of the Internet of Things for power systems, the specific steps of the method are:

[0007] S1, data collection and preprocessing: The intelligent sensors with multi-parameter synchronous measurement and adaptive calibration functions deployed in various links of the power system collect the operating data of various types of power equipment, and transmit them to the data processing center through a low-latency, high-reliability dedicated communication protocol. The intelligent sensors can accurately measure voltage, current, power, equipment temperature, vibration and multiple parameters at the same time. The calibration is based on a multivariable dynamic calibration algorithm, which introduces the electromagnetic environment induction coefficient C inside the power equipment. em And the equipment material fatigue coefficient C fat , let the original measurement value be P raw , the calibration value is P calibrated , the ambient temperature is T, the humidity is H, the equipment has been running for t, and the calibration formula is P calibrated =P raw ×(1+C em ×T+Cfat ×t+k1H), where k1 is determined by performing multivariate linear regression analysis on historical calibration data of a large number of devices of the same model under different environments. The optimization algorithm of matrix singular value decomposition is used in the regression analysis to obtain accurate calibration coefficients under different working conditions. The collected data is cleaned, denoised, and format converted in the data processing center;

[0008] S2, Big data analysis and feature extraction: Use big data analysis technology to conduct in-depth mining of pre-processed data, use image recognition technology to monitor the appearance status of power equipment and fuse and analyze it with internal electrical parameter data, and extract key features that reflect the operating status of power equipment. Image recognition uses an adaptive feature enhancement convolutional neural network algorithm. Suppose the image pixel matrix is ​​I ij , the feature map is F mn , first calculate the local contrast of the pixel Where r is the neighborhood radius, M and N are the image sizes, and the contrast weight matrix W is constructed. C , and then calculate the gradient direction consistency of the pixel where θ i j is the gradient direction of pixel (i, j), and the gradient direction weight matrix W is constructed G , the final attention weight w ij =W C (i,j)×W G 9i,j), the feature extraction formula is k mn The convolution kernel is a learning algorithm that dynamically adjusts the convolution kernel based on data distribution during training. The shape and weight distribution of the convolution kernel are automatically adjusted according to the distribution density of data features to improve the effectiveness of feature extraction.

[0009] S3, Equipment life prediction model construction and training: Based on the extracted features, an equipment life prediction model is constructed using a deep neural network self-learning modeling method based on the fusion of massive historical power data and real-time operation data, and optimized using a quantum optimization algorithm. The number of hidden layer neurons in the deep neural network model is based on the data feature dimension d, the number of samples n, and the information entropy E of the data. info Determine the number of neurons in the hidden layer The information entropy E info The Shannon entropy is calculated on the data. The quantum optimization algorithm in the model training process adopts quantum simulated annealing and gradient descent collaborative optimization algorithm. The quantum bit state transition probability in the quantum simulated annealing algorithm where ΔE ij is the energy difference between quantum states i and j, k B is the Boltzmann constant, Tq is the quantum annealing temperature, T q During the annealing process, the model training is dynamically adjusted according to the convergence of the model training. When the model converges too fast, increase T q To escape from the local optimal solution, when the convergence speed is too slow, reduce T q In order to speed up convergence, the quantum simulated annealing algorithm is used to determine the initial search area of ​​the deep neural network, and then the gradient descent method is used to perform a fine search in the area. The weight update formula of the gradient descent method is: where w g d(t) is the weight from layer i to layer j at the tth iteration, η is the learning rate, E is the loss function, and the loss function uses the weighted mean square error function of data outlier detection. The predicted value is The true value is y, and the outlier detection uses the local outlier factor LOF algorithm. For the data point x i , its local outlier factor Where N k (x i ) is the data point x i The k nearest neighbor set, lrd i For data point x i The local reachable density, according to LOF i Weighting of data points, outlier weights Weighted mean square error function

[0010] S4, Fault prediction and assessment: Input the real-time collected data into the trained equipment life prediction model to predict the remaining life of the equipment and the probability of failure, and evaluate the equipment health status level through the comprehensive evaluation model of power equipment status based on fuzzy logic and evidence theory;

[0011] S5, predictive maintenance strategy formulation: formulate predictive maintenance strategy based on fault prediction results and equipment status assessment results;

[0012] S6, adaptive adjustment: The intelligent detection system adaptively adjusts the detection strategy and parameter thresholds according to the power system operation mode and load changes.

[0013] Furthermore, the low-latency, high-reliability communication protocol reduces communication delay by optimizing the data packet header and transmission process. The data packet header optimization adopts the data importance coding compression algorithm. The importance weight of the data type is set as w i , the packet header length is L h , the length of the original data type identification field is l i0 , the optimized length By analyzing the importance of different data types and encoding compression, the length of the data packet header can be shortened by 40%-60%, and the reliability of data transmission can be improved through error correction coding and retransmission mechanism. The error correction coding adopts the coding algorithm of polynomial transformation on the finite field. The order q of the finite field is determined according to the data block length n and the error correction capability requirement t. By constructing a specific generating polynomial G(x) and performing encoding operations on the finite field GF(q), the data transmission error rate is ensured to be less than 5×10 -7 .

[0014] Furthermore, during the construction and training of the equipment life prediction model, the quantum state update of the quantum simulated annealing algorithm in the quantum simulated annealing and gradient descent collaborative optimization algorithm adopts a strategy of dynamically adjusting the quantum gate rotation angle. Assuming the quantum gate rotation angle is θ, the error change rate of the model training is ΔE rate , θ=θ0×(1+ΔE rate ), where θ0 is the initial quantum gate rotation angle. By dynamically adjusting the quantum gate rotation angle according to the error change rate, it is possible to better balance the global search and local search capabilities, improve the efficiency and accuracy of model training, and avoid falling into the local optimal solution.

[0015] Furthermore, in the comprehensive evaluation model of power equipment status based on the combination of fuzzy logic and evidence theory, the membership function of fuzzy logic is based on the parameter range of the power equipment under different operating conditions and the historical stability coefficient C of the equipment operation. stab Determine, let a parameter of the power equipment be x, for the normal state, the membership function

[0016] For the membership function of abnormal state,

[0017] Where [a1, b1], [a2, b2] are the parameter intervals of normal and abnormal states obtained from a large amount of historical data statistics. The basic probability distribution function in the evidence theory is calculated based on the membership calculation results of fuzzy logic and the credibility weight w of the data. cred Determine, let the normal state membership obtained by fuzzy logic be m normal , the abnormal state membership is m abnormal, , the undetermined state membership is m uncertain =1-m normal -m abnormal , then the basic probability distribution function

[0018] Where θ represents the entire set, and the credibility weight w credBy comprehensively determining the reliability of data sources, the accuracy of data acquisition equipment, and the temporal freshness of data, this method can effectively integrate the uncertainty information in multi-source detection data and improve the accuracy of equipment status assessment.

[0019] Furthermore, the predictive maintenance strategy includes determining the maintenance time, maintenance content, and required resources. The maintenance time is determined based on the predicted value of the remaining life of the equipment L, the probability of failure P, and the maintenance time. f and the current operational risk tolerance R of the power system tol The risk decision model is used to determine When determining maintenance time in the future Within the time period, the maintenance content is determined based on the equipment type, fault prediction type, historical equipment maintenance records and the current load demand distribution of the power system. The required resources are determined based on the maintenance content, the geographical location of the equipment, the surrounding resource reserves and the cost-benefit analysis of resource allocation. In this way, maintenance plans can be reasonably arranged in advance to reduce maintenance costs and equipment failure rates.

[0020] Furthermore, the adaptive adjustment includes automatically increasing the detection frequency of key equipment and important areas during peak load or special operating conditions, and the adjustment multiple of the detection frequency is based on the load change rate ΔL, the equipment importance level I, and the topological complexity C of the power system. top OK, let the detection frequency adjustment multiple k f =ΔL×I×C top +1, reduce the sensitivity of parameter threshold, the adjustment of parameter threshold is based on the load-fault relationship model in historical data and the current power system operation stability index S op Determine, and set the parameter threshold value during normal operation to be T n , parameter threshold T at peak load h =T n ×(1-ΔL×k t ×S op ), where k t is the threshold adjustment coefficient. This adaptive adjustment can ensure that potential fault hazards are discovered in time under different operating conditions and ensure the safe and stable operation of the power system.

[0021] Furthermore, the multi-parameter synchronous measurement of the intelligent sensor adopts the optimized time-division multiplexing technology of clock synchronization and data cache, and the measurement period is T m , the number of parameters is n, and the length of the measurement time slice for each parameter is In each time slice, the measured data is first cached in a high-speed cache area. After all parameter measurements are completed, the cached data is uniformly transmitted based on a high-precision clock synchronization signal. By precisely controlling the time slice length, switching timing, and data caching and transmission strategies, the synchronization and accuracy of multi-parameter measurements are ensured while reducing the hardware complexity and energy consumption of the sensor.

[0022] Furthermore, the denoising in the data preprocessing stage adopts a denoising method that combines morphological filtering and wavelet transform. Assuming the original signal is s(t), a morphological opening operation s is first performed. open =γ λ (φ λ (s(t))), where φ λ is the erosion operation with the structural element λ, γ λ is a dilation operation with the structural element λ, and then the result of the opening operation is decomposed by wavelet. The high-frequency coefficient after wavelet decomposition is d j,k , threshold Where σ is the standard deviation of the noise, which can be estimated by preliminary analysis of the original signal, N is the signal length, and the high-frequency coefficient after denoising is Then the denoised signal is obtained through wavelet reconstruction. This denoising method can effectively remove noise interference in the signal, improve data quality, and provide a more reliable data basis for subsequent analysis.

[0023] Furthermore, the loss function in the deep neural network self-learning modeling method adopts a hybrid loss function that is adaptively adjusted based on data distribution characteristics. The true value is y, and the data distribution characteristic parameter is D feat , including the skewness S of the data k , Kurtosis K u , by statistical analysis of the data, the mean square error term Cross entropy term Mixed loss function L = α × MSE × D feat +(1-α)×CE×(1-D feat ), where α is the weight coefficient, which is determined according to the distribution characteristics of the data and the focus of the prediction target. By adopting a mixed loss function, the model's fitting ability and prediction accuracy for different types of data can be improved.

[0024] Furthermore, the convolutional neural network algorithm in the image recognition technology adopts the regularization operation of feature space mapping during the training process. Suppose the input data of a certain layer of neural network is x i , first map the data to the high-dimensional feature space Φ(x i ), and then calculate the covariance matrix in the feature space Perform eigenvalue decomposition on the covariance matrix C = UΛU T , where U is the eigenvector matrix, Λ is the eigenvalue diagonal matrix, and the regularized output This regularization operation can reduce overfitting and improve the generalization ability of the model.

[0025] Compared with the existing technology, an intelligent detection method of the Internet of Things for power systems has the following beneficial effects:

[0026] 1. This method uses a deep neural network self-learning model to predict the equipment life and optimizes it using a quantum optimization algorithm to accurately predict the remaining life and failure probability of power equipment. At the same time, combined with a comprehensive evaluation model of power equipment status based on fuzzy logic and evidence theory, it can comprehensively evaluate the health status level of equipment and provide a scientific basis for the formulation of predictive maintenance strategies, which helps to reasonably arrange maintenance plans in advance, reduce maintenance costs, extend equipment service life, and improve the economic and social benefits of the power system.

[0027] 2. This method realizes multi-parameter synchronous measurement and adaptive calibration by deploying intelligent sensors in various links of the power system, and uses low-latency, high-reliability communication protocols to transmit data, which significantly improves the accuracy and real-time performance of data acquisition. Combined with big data analysis and feature extraction technology, it can deeply mine the key features in the operation data of power equipment, and provide strong data support for equipment life prediction and fault prediction, which helps to timely discover potential fault hazards, reduce equipment failure rate, and improve the stability and reliability of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 An operation flow chart of an intelligent detection method for the Internet of Things of power systems. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] Embodiment 1

[0032] This embodiment describes a substation in a city that is responsible for power distribution and conversion in the area. There are many devices in the substation and the operating environment is complex. It is necessary to monitor the operating status of the equipment in real time to detect potential faults in time and ensure the stability of power supply.

[0033] Intelligent sensors with multi-parameter synchronous measurement and adaptive calibration functions are deployed on transformers, circuit breakers, and busbar key equipment in substations. For example, the sensors installed on the transformer can accurately measure the voltage on the high-voltage side and the low-voltage side in real time (set as V raw1 、V raw2 ), current (I raw1 ,I raw2 ), power (P r aw1、P raw2 ), and monitor the device temperature (T d ev) and vibration conditions (Vib). These sensors transmit data to the data processing center in the substation through a low-latency, high-reliability dedicated communication protocol. The data packet header uses a data importance coding compression algorithm. For example, a higher weight is given to the key data of voltage and current (assuming that the voltage weight w V =0.8, current weight w I =0.7), the lengths of the original data type identification fields are l V0 =10 bits, l I0 =8 bits, optimized voltage data type identification field length bit, current data type identification field length The data packet header length is shortened by about 50%, effectively reducing communication delay;

[0034] In the data processing center, the collected data is cleaned to remove obvious errors or abnormal data points; the denoising method of morphological filtering and wavelet transform fusion is used to remove noise interference. For example, the current signal I(t) is denoised. Let the original signal be I(t). First, perform morphological opening operation in is the erosion operation with the structural element λ, γ λ is a dilation operation with the structural element λ, and then the result of the opening operation is decomposed by wavelet. The high-frequency coefficient after wavelet decomposition is d j,k , threshold (Assuming that the noise standard deviation σ is estimated to be 0.5 through preliminary analysis of the original signal, and the signal length N is 1000), the high-frequency coefficient after denoising is Then the denoised signal is obtained through wavelet reconstruction, and finally format conversion is performed to uniformly convert the data into a format that is convenient for subsequent analysis.

[0035] The pre-processed data is deeply mined by using big data analysis technology, and the appearance status of the equipment is monitored by combining image recognition technology (such as using a camera to take pictures of the transformer to see if there is oil leakage or shell deformation), and then fused and analyzed with the internal electrical parameter data. The adaptive feature enhancement convolutional neural network algorithm is used for image recognition. Suppose the image pixel matrix is ​​I ij , the feature map is F mn , first calculate the local contrast of the pixel (Assuming the neighborhood radius r = 3, image size M = 200, N = 200), construct the contrast weight matrix W C , and then calculate the gradient direction consistency of the pixel where θ i j is the gradient direction of pixel (i, j)), and the gradient direction weight matrix W is constructed G , the final attention weight w i j=W C (i,j)×W G (i,j), the feature extraction formula is (k mn The convolution kernel is a convolution kernel, and the convolution kernel is dynamically adjusted based on the learning algorithm of data distribution during training. The shape and weight distribution of the convolution kernel are automatically adjusted according to the distribution density of data features, which improves the effectiveness of extracting the operating status features of the transformer. Taking the transformer oil temperature data and appearance image data as an example, big data analysis shows that there is a correlation between the oil temperature and certain features in the appearance image (such as the color change of the heat sink). The operating status of the transformer can be judged more accurately through fusion analysis.

[0036] Based on the extracted features, an equipment life prediction model is constructed using a deep neural network self-learning modeling method based on the fusion of massive historical power data and real-time operation data. For the transformer in the substation, the number of neurons in the hidden layer of the deep neural network model is determined according to its historical operation data and real-time monitoring data. Assume that the data feature dimension d = 10, the number of samples n = 500, and the information entropy E i nfo=0.8, according to the formula The number of neurons in the hidden layer can be obtained The quantum optimization algorithm is used to optimize the model training process. When training the transformer life prediction model, the quantum bit state transition probability in the quantum simulated annealing algorithm is where ΔE ij is the energy difference between quantum states i and j, k B is the Boltzmann constant, T q is the quantum annealing temperature, quantum annealing temperature T qDuring the annealing process, adjustments are made according to the convergence of model training. For example, when the initial convergence speed is too fast, the quantum annealing temperature is increased to jump out of the local optimal solution. When the convergence speed is too slow, the quantum annealing temperature is reduced to speed up convergence. The initial search area of ​​the deep neural network is determined by the quantum simulated annealing algorithm, and then the gradient descent method is used to perform a fine search in this area. The weight update formula of the gradient descent method is: where w gd (t) is the weight from layer i to layer j at the tth iteration, η is the learning rate, E is the loss function, and the loss function uses the weighted mean square error function of data outlier detection. The predicted value is The true value is y, and the outlier detection uses the local outlier factor LOF algorithm. For the data point x i , its local outlier factor Where N k (x i ) is the data point x i The k nearest neighbor set, lrd i For data point x i The local reachable density, according to LOF i Weighting of data points, outlier weights Weighted mean square error function Finally, the trained transformer life prediction model is obtained.

[0037] The real-time collected transformer data is input into the trained equipment life prediction model to predict the remaining life of the transformer (such as the remaining life is predicted to be 5 years) and the probability of failure (such as the probability of failure within the next year is predicted to be 5%). The health status level of the transformer is evaluated by the comprehensive evaluation model of the power equipment status based on fuzzy logic and evidence theory. The membership function of fuzzy logic is determined according to the oil temperature, load parameter range and historical stability coefficient of the equipment operation under different operating conditions of the transformer. For example, if the transformer oil temperature is x, the membership function for the normal state is (Assuming [a1, b1] = [50, 70], [a2, b2] = 70, 90], the anti-historical stability coefficient C stab =0.9), when the oil temperature x = 65°C, (This is just an example calculation. In practice, the membership function may need to be adjusted based on more factors). For the abnormal state, the membership function

[0038]

[0039] The basic probability distribution function in evidence theory is determined according to the membership calculation results of fuzzy logic and the credibility weight of the data. Suppose the normal state membership obtained by fuzzy logic is m normal , the abnormal state membership is mabnormal, , the undetermined state membership is m uncertain =1-m normal -m abnormal , then the basic probability distribution function m({θ})=m uncertain , where θ represents the entire set and the credibility weight w cred The reliability of the data source, the accuracy of the data acquisition equipment, and the time freshness of the data are comprehensively determined (assuming that the reliability of the data source is 0.8, the accuracy of the data acquisition equipment is 0.9, and the time freshness of the data is 0.95, and the weighted average method is used to obtain w cred =0.88), and finally the transformer health status level is judged to be good according to the evaluation results.

[0040] The predictive maintenance strategy of the transformer is formulated based on the fault prediction results and the equipment status assessment results. The maintenance time is determined by the risk decision model based on the predicted value of the remaining life of the equipment, the probability of failure, and the current operating risk tolerance of the power system. For example, when the probability of failure P f =5%, remaining life L = 5 years, operating risk tolerance R t When ol=0.01, calculate Approaching the operational risk tolerance, the maintenance time is determined within the next month. The maintenance content is determined based on the transformer type (such as oil-immersed transformer), fault prediction type (such as possible insulation aging problems), historical equipment maintenance records (such as the last maintenance time and maintenance content), and the current load demand distribution of the power system. For example, insulation performance testing and replacement of some aging components are arranged. The required resources are determined based on the maintenance content (such as testing equipment, insulating materials), the geographical location of the equipment (such as the specific location in the substation), the surrounding resource reserves (such as whether there are spare parts in nearby warehouses), and the cost-benefit analysis of resource allocation to ensure that the maintenance work is carried out efficiently and economically.

[0041] During the peak load period in summer, the detection frequency of key transformer equipment is automatically increased. The detection frequency adjustment multiple is based on the load change rate (for example, the load increases by 30%), the equipment importance level (the transformer importance level is high, set to 3), and the topological complexity of the power system (the substation topological complexity is moderate, set to C top =1.5) is determined according to the formula k f =ΔL×I×C top +1, the detection frequency adjustment multiple k can be calculated f=0.3×3×1.5+1=2.35, that is, the detection frequency is increased by about 2.35 times, and the parameter threshold sensitivity is reduced. The adjustment of the parameter threshold is determined according to the load-temperature relationship model in the historical data and the operation stability index of the current power system. For example, the upper limit threshold of the transformer oil temperature during normal operation is T n = 85℃, at peak load, according to the current operation stability index (such as small system voltage fluctuation, stability index S σp =0.9) and threshold adjustment coefficient (assuming k t =0.5), according to the formula T h =T n ×(1-ΔL×k t ×S op ), calculate the upper limit threshold of oil temperature during peak load T h =85×(1-0.3×0.5×0.9)≈79.8℃, so as to discover potential failure risks more timely.

[0042] Embodiment 2

[0043] This embodiment describes that the power transmission line of a power grid in a certain city stretches for hundreds of kilometers, and there are multiple pole towers distributed along the line. Smart sensors are installed on the pole towers to monitor the operating status of the power transmission line.

[0044] The intelligent sensor collects the current, voltage, conductor temperature, and tower vibration data of the transmission line every 5 milliseconds. For long-distance transmission lines, the ambient temperature and humidity at different locations vary greatly. The ambient temperature and humidity are obtained through the sensor's built-in environmental monitoring module. The equipment's operating time is obtained from the line commissioning record. These data are substituted into the calibration formula to calibrate the collected raw data. The data is transmitted through an optimized communication protocol. The communication protocol reduces the length of the data packet header by about 45% through importance analysis and coding compression of different data types. In the data processing center, the data is cleaned to remove erroneous data caused by communication interference. The denoising process uses the fusion method of morphological filtering and wavelet transform to process the conductor temperature signal. First, morphological opening operation is performed to remove part of the noise, and then wavelet decomposition and threshold processing of high-frequency coefficients are performed to reconstruct the signal, and finally the data format is converted.

[0045] The data is analyzed and processed using big data technology. Image recognition technology is used to combine the transmission line images taken by drone inspection with electrical parameter data for fusion analysis. The local contrast and gradient direction consistency of image pixels are calculated in the image recognition algorithm, and the weight matrix is ​​constructed. The features are extracted through the adaptive feature enhancement convolutional neural network. The image pixel matrix is ​​set as I ij , the feature map is F mn , first calculate the local contrast of the pixel Where r is the neighborhood radius, M and N are the image sizes, and the contrast weight matrix W is constructed. C , and then calculate the gradient direction consistency of the pixel where θ i j is the gradient direction of pixel (i, j), and the gradient direction weight matrix W is constructed G , the final attention weight w ij =W C (i,j)×W G (i,j), the feature extraction formula is k mn It is a convolution kernel, which is dynamically adjusted according to the data distribution during training. For example, when detecting whether an insulator is damaged or dirty, it can accurately extract relevant features and improve recognition accuracy.

[0046] Based on the extracted features, a life prediction model for transmission line equipment is constructed. The number of neurons in the hidden layer of the deep neural network is determined according to the data feature dimension, the number of historical samples, and the calculated data information entropy. The quantum simulated annealing and gradient descent collaborative optimization algorithm is used to train the model. The quantum bit state transition probability in the quantum simulated annealing algorithm is According to the quantum state energy difference ΔE ij , Boltzmann constant k B and dynamically adjusted quantum annealing temperature calculation T q , using quantum simulated annealing to determine the initial search area, The gradient descent method performs a detailed search in the area, and the weight update is calculated based on the loss function. The loss function uses the weighted mean square error function for data outlier detection. The LOF algorithm is used to detect outliers and weight them to enhance the model's ability to handle abnormal data.

[0047] The real-time data is input into the trained model to predict the remaining life of the transmission line and the probability of failure. For example, the remaining life of a certain line is predicted to be 8 years and the probability of failure is 3%. The evaluation model combining fuzzy logic and evidence theory is used to evaluate the equipment health status level. The fuzzy logic membership function is based on the parameter range and historical stability coefficient C under different operating conditions of the transmission line. stab Determine, let a parameter of the power equipment be x, for the normal state, the membership function

[0048] For the membership function of abnormal state,

[0049] Where [a1, b1], [a2, b2] are the parameter intervals of normal and abnormal states obtained from a large amount of historical data statistics. The basic probability distribution function in the evidence theory is calculated based on the membership calculation results of fuzzy logic and the credibility weight w of the data. cred Determine, let the normal state membership obtained by fuzzy logic be m normal , the abnormal state membership is m abnormal, , the undetermined state membership is m uncertain =1-m normal -m abnormal , then the basic probability distribution function m9{θ})=m uncertain , where θ represents the entire set and the credibility weight w cred It is determined through comprehensive evaluation of the reliability of the data source, accuracy analysis of the data acquisition equipment, and the temporal freshness of the data.

[0050] A maintenance strategy is formulated based on the prediction results. If a section of line has a high probability of failure and a short remaining life, the maintenance time is determined to be within the next month. The maintenance content is determined based on the line type, possible fault type, historical maintenance records and the current grid load demand distribution, such as replacing some insulators and checking the wire connection parts. The required resources are determined based on the maintenance content, the geographical location of the line, the surrounding resource reserves and the cost-benefit analysis of resource allocation, such as arranging maintenance personnel and transportation tools.

[0051] Under special operating conditions in severe weather, the detection frequency of key areas is increased. The detection frequency adjustment multiple is determined based on the load fluctuations caused by weather changes, the equipment importance level and the complexity of the power system topology, such as adjusting it to 1.5 times the usual level. At the same time, the parameter threshold sensitivity is reduced. For example, the conductor dancing amplitude alarm threshold is 0.5 meters under normal circumstances, and it is adjusted to 0.4 meters under special conditions.

[0052] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the same elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A power system Internet of Things intelligent detection method, characterized in that: The specific steps of this method are: S1, data collection and preprocessing: The intelligent sensors with multi-parameter synchronous measurement and adaptive calibration functions deployed in various links of the power system collect the operating data of various types of power equipment, and transmit them to the data processing center through a low-latency, high-reliability dedicated communication protocol. The intelligent sensors can accurately measure voltage, current, power, equipment temperature, vibration and multiple parameters at the same time. The calibration is based on a multivariable dynamic calibration algorithm, which introduces the electromagnetic environment induction coefficient C inside the power equipment. em And the equipment material fatigue coefficient C fat , let the original measurement value be P raw , the calibration value is P calibrated , the ambient temperature is T, the humidity is H, the equipment has been running for t, and the calibration formula is P calibrated =P raw ×(1+C em ×T+C fat ×t+k1H), where k1 is determined by performing multivariate linear regression analysis on historical calibration data of a large number of devices of the same model under different environments. The optimization algorithm of matrix singular value decomposition is used in the regression analysis to obtain accurate calibration coefficients under different working conditions. The collected data is cleaned, denoised, and format converted in the data processing center; S2, Big data analysis and feature extraction: Use big data analysis technology to conduct in-depth mining of pre-processed data, use image recognition technology to monitor the appearance status of power equipment and fuse and analyze it with internal electrical parameter data, and extract key features that reflect the operating status of power equipment. Image recognition uses an adaptive feature enhancement convolutional neural network algorithm. Suppose the image pixel matrix is ​​I ij , the feature map is F mn , first calculate the local contrast of the pixel Where r is the neighborhood radius, M and N are the image sizes, and the contrast weight matrix W is constructed. C , and then calculate the gradient direction consistency of the pixel where θ i j is the gradient direction of pixel (i, j), and the gradient direction weight matrix W is constructed G , the final attention weight w ij =W C (i, j)×W G (i, j), the feature extraction formula is k mn The convolution kernel is a learning algorithm that dynamically adjusts the convolution kernel based on data distribution during training. The shape and weight distribution of the convolution kernel are automatically adjusted according to the distribution density of data features to improve the effectiveness of feature extraction. S3, Equipment life prediction model construction and training: Based on the extracted features, an equipment life prediction model is constructed using a deep neural network self-learning modeling method based on the fusion of massive historical power data and real-time operation data, and optimized using a quantum optimization algorithm. The number of hidden layer neurons in the deep neural network model is based on the data feature dimension d, the number of samples n, and the information entropy E of the data. info Determine the number of neurons in the hidden layer The information entropy E info The Shannon entropy is calculated on the data. The quantum optimization algorithm in the model training process adopts quantum simulated annealing and gradient descent collaborative optimization algorithm. The quantum bit state transition probability in the quantum simulated annealing algorithm where ΔE ij is the energy difference between quantum states i and j, k B is the Boltzmann constant, T q is the quantum annealing temperature, T q During the annealing process, the model training is dynamically adjusted according to the convergence of the model training. When the model converges too fast, increase T q To escape from the local optimal solution, when the convergence speed is too slow, reduce T q In order to speed up convergence, the quantum simulated annealing algorithm is used to determine the initial search area of ​​the deep neural network, and then the gradient descent method is used to perform a fine search in the area. The weight update formula of the gradient descent method is: where w g d(t) is the weight from layer i to layer j at the tth iteration, η is the learning rate, E is the loss function, and the loss function uses the weighted mean square error function of data outlier detection. The predicted value is The true value is y, and the outlier detection uses the local outlier factor LOF algorithm. For the data point x i , its local outlier factor Where N k (x i ) is the data point x i The k nearest neighbor set, lrd i For data point x i The local reachable density, according to LOF i Weighting of data points, outlier weights Weighted mean square error function S4, Fault prediction and assessment: Input the real-time collected data into the trained equipment life prediction model to predict the remaining life of the equipment and the probability of failure, and evaluate the equipment health status level through the comprehensive evaluation model of power equipment status based on fuzzy logic and evidence theory; S5, predictive maintenance strategy formulation: formulate predictive maintenance strategy based on fault prediction results and equipment status assessment results; S6, adaptive adjustment: The intelligent detection system adaptively adjusts the detection strategy and parameter thresholds according to the power system operation mode and load changes.

2. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: The low-latency, high-reliability communication protocol reduces communication delay by optimizing the data packet header and transmission process. The data packet header optimization adopts the data importance coding compression algorithm. Suppose the importance weight of the data type is w i , the packet header length is L h , the length of the original data type identification field is l i0 , the optimized length By analyzing the importance of different data types and encoding compression, the length of the data packet header can be shortened by 40%-60%, and the reliability of data transmission can be improved through error correction coding and retransmission mechanism. The error correction coding adopts the coding algorithm of polynomial transformation on the finite field, and determines the order q of the finite field according to the data block length n and the error correction capability requirement t. The coding operation is performed on the finite field GF(q) by constructing a specific generating polynomial G(x).

3. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: During the construction and training of the equipment life prediction model, the quantum state update of the quantum simulated annealing algorithm in the quantum simulated annealing and gradient descent collaborative optimization algorithm adopts a strategy of dynamic adjustment of the quantum gate rotation angle. The quantum gate rotation angle is θ, and the error change rate of the model training is ΔE rate , θ=θ0×(1+ΔE rate ), where θ0 is the initial quantum gate rotation angle. By dynamically adjusting the quantum gate rotation angle according to the error change rate, the global search and local search capabilities can be better balanced.

4. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: In the comprehensive evaluation model of power equipment status based on the combination of fuzzy logic and evidence theory, the membership function of fuzzy logic is based on the parameter range of the power equipment under different operating conditions and the historical stability coefficient C of the equipment operation. stab Determine, let a parameter of the power equipment be x, for the normal state, the membership function For the membership function of abnormal state, Where [a1, b1], [a2, b2] are the parameter intervals of normal and abnormal states obtained from a large amount of historical data statistics. The basic probability distribution function in the evidence theory is calculated based on the membership calculation results of fuzzy logic and the credibility weight w of the data. cred Determine, let the normal state membership obtained by fuzzy logic be m normal , the abnormal state membership is m abnormal， , the undetermined state membership is m uncertain =1-m normal -m abnormal , then the basic probability distribution function m({θ})=m uncertain , where θ represents the entire set and the credibility weight w cred It is determined through comprehensive evaluation of the reliability of the data source, accuracy analysis of the data acquisition equipment, and the temporal freshness of the data.

5. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: The predictive maintenance strategy includes determining the maintenance time, maintenance content, and required resources. The maintenance time is based on the predicted value of the remaining life of the equipment L, the probability of failure P, and the maintenance time of the equipment. f And the current operational risk tolerance R of the power system tol The risk decision model is used to determine When determining maintenance time in the future Within the time period, the maintenance content is determined based on the equipment type, fault prediction type, historical maintenance record of the equipment and the current load demand distribution of the power system. The required resources are determined based on the maintenance content, the geographical location of the equipment, the surrounding resource reserves and the cost-benefit analysis of resource allocation.

6. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: The adaptive adjustment includes automatically increasing the detection frequency of key equipment and important areas during peak load or special operating conditions. The adjustment multiple of the detection frequency is based on the load change rate ΔL, the equipment importance level I and the topological complexity C of the power system. top OK, let the detection frequency adjustment multiple k f =ΔL×I×C top +1, reduce the sensitivity of parameter threshold, the adjustment of parameter threshold is based on the load-fault relationship model in historical data and the current power system operation stability index S op Determine, and set the parameter threshold value during normal operation to be T n , parameter threshold T at peak load h =T n ×(1-ΔL×k t ×S op ), where k t is the threshold adjustment factor.

7. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: The multi-parameter synchronous measurement of the intelligent sensor adopts the optimized time-division multiplexing technology of clock synchronization and data cache. The measurement period is T m , the number of parameters is n, and the measurement time slice length of each parameter is In each time slice, the measured data is first cached in the high-speed cache area. After all parameter measurements are completed, the cached data is uniformly transmitted based on the high-precision clock synchronization signal, by precisely controlling the time slice length, switching timing, and data caching and transmission strategy.

8. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: The denoising method in the data preprocessing stage adopts the denoising method of morphological filtering and wavelet transform fusion. Suppose the original signal is s(t), and the morphological opening operation s is first performed. open =γ λ (φ λ (s(t))), where φ λ is the erosion operation with the structural element λ, γ λ is a dilation operation with the structural element λ, and then the result of the opening operation is decomposed by wavelet. The high-frequency coefficient after wavelet decomposition is d j，k , threshold Where σ is the standard deviation of the noise, which can be estimated by preliminary analysis of the original signal, N is the signal length, and the high-frequency coefficient after denoising is Then the denoised signal is obtained through wavelet reconstruction.

9. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: The loss function in the deep neural network self-learning modeling method adopts a hybrid loss function that is adaptively adjusted based on data distribution characteristics. The predicted value is The true value is y, and the data distribution characteristic parameter is D feat , including the skewness S of the data k , Kurtosis K u , by statistical analysis of the data, the mean square error term Cross entropy term Mixed loss function L = α × MSE × D feat +(1-α)×CE×(1-D feat ), where α is the weight coefficient, which is determined according to the distribution characteristics of the data and the focus of the prediction target. By adopting a mixed loss function, the model's fitting ability and prediction accuracy for different types of data can be improved.

10. According to claim 1, a power system Internet of Things intelligent detection method is characterized in that: The convolutional neural network algorithm in the image recognition technology adopts the regularization operation of feature space mapping during the training process. Suppose the input data of a certain layer of neural network is x i , first map the data to the high-dimensional feature space Φ(x i ), and then calculate the covariance matrix in the feature space Perform eigenvalue decomposition on the covariance matrix C = UΛU T , where U is the eigenvector matrix, Λ is the eigenvalue diagonal matrix, and the regularized output

Images