A method for on-line monitoring of transformer insulation

Abstract

The application belongs to the technical field of transformer insulation monitoring, and discloses a transformer insulation online monitoring method, which comprises the following steps: arranging electrical, thermal, mechanical and chemical multidimensional sensing units at key positions of a transformer, synchronously collecting various state parameters and establishing spatial correlation; adopting a mutual information entropy screening combined with a redundant elimination mode to extract key features such as a dielectric loss tangent change rate and a characteristic gas concentration ratio in oil, and focusing on determining threshold intervals and change trend differences of thermal aging and environment stress related parameters; utilizing a mechanism and data double attention mechanism to strengthen feature discrimination, and then through an improved support vector machine and a Transformer integrated model, precise identification of thermal aging, sudden cooling and sudden heating stress and other deterioration types and subclasses is realized; using a time delay difference of signals collected by a distributed sensing array, combining spatial coordinates of each unit, and through a time difference positioning algorithm, a preliminary deterioration range is preliminarily circled, and a cross-correlation function method is used to improve signal synchronization accuracy.

Description

Technical Field

[0001] This invention belongs to the field of transformer insulation monitoring technology, specifically a method for online monitoring of transformer insulation. Background Technology

[0002] Transformers are crucial hub equipment in power systems, and their insulation performance directly affects the safety and stability of power transmission. Currently, power systems are developing towards higher voltage, larger capacity, and greater intelligence. Traditional offline insulation testing methods, such as dielectric loss measurement and insulation resistance testing, are no longer sufficient to meet the needs of real-time equipment condition assessment. While existing online transformer insulation monitoring technology has achieved the basic goal of real-time monitoring, the following issues remain regarding its adaptability to complex operating conditions and the further improvement of monitoring accuracy.

[0003] On the one hand, most existing technologies use single-dimensional sensing and monitoring, such as monitoring only partial discharge signals or dissolved gases in oil, which makes it difficult to effectively distinguish between insulation aging and environmental stress degradation. The latter is mostly caused by environmental factors such as sudden cooling and heating, and manifests as temporary performance fluctuations of the insulating medium. Moreover, these technologies do not relate to the multi-field action mechanism behind the degradation, which is prone to misjudgment and cannot provide accurate differentiated data support for insulation life prediction.

[0004] On the other hand, the monitoring system has limited spatial positioning accuracy for weak points in the internal insulation of transformers. Existing methods can usually only determine the general area of ​​deterioration, and fail to combine the comprehensive distribution characteristics of electric field, temperature field and stress field, so they cannot accurately locate specific weak points such as winding coils and lead insulation. This will lead to low efficiency when maintenance personnel carry out targeted maintenance, thereby increasing the cost of power outage maintenance and safety risks. Summary of the Invention

[0005] The purpose of this invention is to provide an online monitoring method for transformer insulation to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for online monitoring of transformer insulation, comprising the following specific steps: Preferably, during the sensor deployment stage, a dielectric sensing unit is embedded in key locations such as the gaps between different winding coils and the lead insulation layer using laser positioning-assisted installation technology. This allows for real-time monitoring of changes in the dielectric constant and conductivity of the insulating medium, as well as the adsorption of characteristic products such as furfural generated during insulation aging and the quantification of the adsorption amount. A fiber optic grating sensor array is evenly distributed every 5 cm along the winding axis and every 3 cm in the radial direction to simultaneously collect local temperature and mechanical stress data; eight MEMS vibration and electric field joint sensing units are evenly distributed around the inner wall of the oil tank and four are symmetrically arranged at the ends of the winding to capture the time-series correlation characteristics of local electric field distortion and vibration signal accompanying insulation degradation. By coordinating the layout of electrical, thermal, mechanical, and chemical multi-dimensional sensing units and combining laser ranging to calibrate the spatial coordinates of each sensing unit, the synchronous acquisition and spatial correlation of multiple parameters of insulation status are realized, and the acquired multi-dimensional raw signals are output to the signal processing stage.

[0007] Preferably, the signal processing stage uses a fusion model of convolutional neural network and gated recurrent unit to extract features from the multi-dimensional original signal, which is used to identify typical feature patterns of interference signals under different operating conditions, including the spike pulse feature of closing impact, the fixed frequency feature of harmonic interference, and the random fluctuation feature of oil flow vibration. This fusion model employs a concatenated structure of CNN feature extraction and GRU temporal modeling: The CNN part consists of 3 convolutional layers and 2 pooling layers, with ReLU activation function; the GRU part consists of 2 hidden layers, each with 128 hidden units, a dropout rate of 0.2, and tanh activation function.

[0008] The model training dataset contains 5,000 sets of three types of interference samples each: closing impact samples include closing scenarios of 110kV-500kV transformers, harmonic interference samples include 3rd to 13th harmonics, and oil flow vibration samples include fluctuations of 0.1Hz-100Hz. The sample labeling method is "interference type + feature parameter range". The training steps are as follows: ① The original signals of multiple dimensions, such as dielectric constant, temperature, vibration amplitude, and electric field intensity time series data, are sliced ​​according to a time window length of 60s and standardized to the [0,1] interval as input; ② The Adam optimizer is used, with an initial learning rate of 0.001 and a decay of 10% every 100 rounds; ③ Using the cross-entropy loss function as the optimization objective, iterate the training until the accuracy of the validation set stabilizes above 95% and then stop.

[0009] The model takes 4-dimensional time series data as input and outputs 3-dimensional feature vectors, which correspond to the feature matching degree of the three types of interference.

[0010] Based on the extracted interference features, the parameter matrix of the Kalman filter is dynamically adjusted to achieve adaptive suppression of different types of interference. Simultaneously, the empirical mode decomposition method is used to decompose non-stationary vibration interference into multiple intrinsic mode components and remove invalid components. Finally, a signal-to-noise ratio-based weighted fusion algorithm is used to perform multi-source fusion on the interference-free effective signal. The weights are dynamically allocated according to the real-time signal-to-noise ratio of each sensing unit. Then, the wavelet threshold enhancement algorithm is used to improve the signal amplitude and signal-to-noise ratio. The processed effective signal is output to the degradation identification stage.

[0011] Preferably, the degradation identification stage extracts key features from the preprocessed multi-dimensional signals by using a combination of mutual information entropy screening and redundancy removal, such as the rate of change of dielectric loss tangent, partial discharge pulse characteristic parameters, the ratio of characteristic gas concentrations in the oil, temperature gradient changes, and the amount of aging products adsorbed by the nanosensor unit. In particular, it strengthens the differentiation and extraction of thermal aging characteristic parameters and environmental stress characteristic parameters, and clarifies the threshold range and the difference in the changing trend of the two types of features. The threshold range for thermal aging characteristics is set as follows: the change rate of dielectric loss tangent is ≥0.002 / day, the CO / CO2 ratio is ≥0.05, the adsorption amount of aging products is ≥0.3mg / g, the temperature gradient is stable at 3-8℃ / cm, and the above characteristics continue to rise for more than 72 hours, which is judged as thermal aging deterioration. The environmental stress characteristic threshold range is set as follows: instantaneous fluctuation amplitude of dielectric constant ≥15%, stress mutation coefficient ≥50MPa / second, partial discharge pulse amplitude ≥300mV but no aging product accumulation, and the above characteristics recover to the healthy range within 24 hours to be judged as environmental stress deterioration.

[0012] A dual attention mechanism based on mechanism and data is established. On the one hand, a mechanism knowledge base is established to impose prior constraints on feature weights based on the chemical kinetic mechanism of insulation aging and the physical mechanism of environmental stress degradation. On the other hand, the weights are dynamically optimized through a data-driven multi-head attention mechanism to strengthen features that are strongly correlated with degradation type. The mechanism knowledge base is stored using a structured database. The core data consists of key-value pairs representing degradation type, feature type, weight constraint interval, and mechanistic basis, specifically including: ① Thermal aging type: CO / CO2 ratio weight constraint [0.2-0.3] (based on the chemical kinetics of CO / CO2 generation from cellulose thermal decomposition), and aging product adsorption amount weight constraint [0.25-0.35] (based on the positive correlation mechanism between aging product accumulation and insulation degradation). ② Electrical aging type: partial discharge pulse amplitude weight constraint [0.3-0.4] (based on the electrical breakdown mechanism of partial discharge on insulating medium), electric field distortion degree weight constraint [0.2-0.3] (based on the mechanism of electric field concentration accelerating insulation aging); ③ Environmental stress type: Dielectric constant fluctuation amplitude weight constraint [0.25-0.35] (based on the physical mechanism of thermal expansion and contraction of the medium caused by sudden temperature change), stress mutation coefficient weight constraint [0.2-0.3] (based on the influence mechanism of mechanical stress impact on insulation structure); The knowledge base supports dynamic updates. When new degradation cases are added to verify new mechanisms, the weight constraint range can be added or modified through the cloud backend.

[0013] Finally, the weighted fused features are input into the improved ensemble model of support vector machine and Transformer model. SVM is used for linear classification of features, and Transformer model captures long-term temporal correlation of features. The outputs of both are used to identify and classify two degradation types and subclasses through weighted voting decision. Finally, the degradation type and subclass identification results are output to the 3D localization stage.

[0014] The data-driven part adopts an 8-head multi-head attention mechanism with an attention dimension of 64, and calculates feature association weights by scaling dot products. The mechanism knowledge base adopts a key-value pair storage structure, where the key is the feature type and the value is the weight constraint range. For example, the weight constraint range for the "CO / CO2 ratio" associated with thermal aging is [0.2-0.3], and the weight constraint range for the dielectric constant fluctuation amplitude associated with environmental stress is [0.25-0.35]. The dual attention weight fusion method is "mechanism prior weight × 0.4 + data dynamic weight × 0.6".

[0015] The improved SVM uses the RBF kernel function, and the penalty coefficient C takes a value range of [1-10]. It is adaptively adjusted according to the feature dimension. When the feature dimension is >10, C=8, otherwise C=3. The Transformer model contains 3 encoder layers, each with 8 multi-head attention and 1 fully connected layer. The position encoding uses sine function encoding, and the sequence length is set to 30, corresponding to the feature sequence of 30 sampling periods. The weighted voting rule is "improved SVM prediction probability × 0.3 + Transformer prediction probability × 0.7", with the prediction probability threshold set at 0.8. If the probability is higher than the threshold, it is judged as the corresponding degradation type.

[0016] The training dataset contains 3000 sets of each of the five types of deterioration samples: thermal aging, electrical aging, sudden cooling and heating stress, vibration stress, and normal state. The dataset is divided into training set, validation set, and test set in a ratio of 7:2:1. The input consists of fused features such as the change rate of dielectric loss tangent, three pulse parameters of partial discharge, the ratio of two gases in the oil, temperature gradient, stress change, and adsorption amount of aging products. The SGD optimizer was used with a learning rate of 0.005 and 200 iterations. The training was terminated when the classification accuracy was ≥96%. The output consisted of degradation type labels and confidence scores. In the degradation type labels, 0 = normal, 1 = thermal aging, 2 = electrical aging, 3 = sudden cooling and heating stress, and 4 = vibration stress.

[0017] Preferably, in the three-dimensional positioning stage, the time delay difference of the signals collected by different sensing units in the distributed sensing array is used, combined with the spatial coordinates of each unit, and the range of the degraded area is initially determined by the time difference positioning algorithm. The cross-correlation function method is used to perform synchronous alignment processing on the signals. Based on the actual size parameters and material properties of the internal insulation structure of the transformer, a finite element forward model of the synergistic effect of electric field, temperature field and stress field is established. The characteristic parameters of the identified degradation type are used as input conditions. The equivalent parameters of the degradation region in the model are inverted and optimized by genetic algorithm, and the simulation results of the multi-physics distribution of the degradation region are iteratively corrected. Combining the signal amplitude attenuation law of multiple sensing units, the abnormal area of ​​multi-physical field distribution, and the preliminary positioning range, the Kriging space interpolation algorithm is used to realize the three-dimensional coordinate positioning of the weak point in the insulation. At the same time, the multi-physical field distribution cloud map of the positioning area is output, presenting information such as field strength distribution and temperature gradient of the deteriorated area. The obtained three-dimensional positioning results and deterioration type information are output to the lifetime prediction stage.

[0018] Preferably, the life prediction stage integrates transformer factory test data, historical operation and maintenance data, and real-time data from online monitoring. The data undergoes outlier removal, missing value completion, and normalization preprocessing to establish a full-state dataset including healthy, slightly deteriorated, moderately deteriorated, and severely deteriorated states. The dataset is then categorized and labeled according to transformer model, years of operation, and load characteristics. A temporal attention LSTM network is used to capture the evolution of insulation state parameters over time. An aging acceleration factor with the combined effect of multiple factors is embedded, and the loss function of the network is optimized based on the insulation aging mechanism to improve the model's fitting accuracy to the aging trend. Establish a transfer learning mechanism, based on a general model of the same type of transformer, fine-tune the model using historical data of a single transformer, generate a personalized life prediction model, and output insulation status assessment level, remaining life under different operation and maintenance strategies, and full life cycle data to the architecture building stage.

[0019] (1) Temporal attention LSTM structure: It contains 2 LSTM layers and 1 temporal attention layer. The attention window size is 12, corresponding to 12 months of state data. The input is the timed dielectric constant, conductivity, concentration of characteristic gas in oil, temperature, stress, adsorption amount of aging products and other insulation parameters. The time interval is 1 month, the sequence length is set to 60, corresponding to 5 years of historical data. The aging acceleration factor is quantified as "temperature weight 0.4 + electric field strength weight 0.3 + humidity weight 0.2 + load factor weight 0.1". The weights of each factor are determined through orthogonal experiments, where the temperature is 20-120℃, the electric field strength is 10-50kV / mm, the humidity is 30%-90%, and the load factor is 0.5-1.2 times the rated load. The loss function uses MSE loss combined with an aging mechanism penalty term, with the penalty term coefficient set to 0.1. The penalty is triggered when the predicted trend contradicts the Arrhenius aging law.

[0020] (2) Transfer learning steps: ① The general model training is based on the full life cycle data of 100 transformers of the same model, each containing 5-10 years of data. The Adam optimizer is used with an initial learning rate of 0.002 and 300 iterations. ② Using at least two years of historical data from a single transformer as fine-tuning samples, the learning rate was reduced to 0.0005, and the process was iterated for 50 rounds; ③ The adaptation standard is that the prediction error of the model after fine-tuning is ≤5%.

[0021] (3) The corresponding standards for output status levels 1-5 are as follows: Level 1 (healthy) = insulation parameter compliance rate ≥95%; Level 2 (slight deterioration) = 85%-94%; Level 3 (moderate deterioration) = 70%-84%; Level 4 (severe deterioration) = 55%-69%; Level 5 (severe deterioration) ≤54%.

[0022] Preferably, during the architecture building phase, a signal preprocessing algorithm, a preliminary degradation type identification model, and a personalized lifetime prediction sub-model are deployed at the edge. The computational overhead is reduced through model pruning and quantization optimization, thereby enabling real-time alarms and basic status assessments for abnormal signals. The preprocessed valid data, preliminary identification results, and model operation logs are encrypted and uploaded to the cloud platform. The cloud platform runs the 3D positioning model and the multiphysics simulation model, and uses a distributed database to store historical data. The cloud integrates monitoring data from multiple transformers of the same type through federated learning algorithms to iteratively optimize the general model, while providing updated parameters for personalized models at the edge. A model adaptability evaluation module is configured to monitor the prediction accuracy of the edge model in real time. When the prediction accuracy of the edge model is lower than the set threshold, the fine-tuning process is automatically triggered, and accurate evaluation, prediction, and positioning information are output to the operation and maintenance decision-making stage.

[0023] Federated learning employs the FedAvg aggregation strategy, with at least 10 transformers of the same model participating. Data sharding involves independent local data storage for each transformer, with only model parameters uploaded. The communication rounds are set to 50 rounds, with 20 iterations of local training per round. The local optimizer is SGD with a learning rate of 0.003. Parameter aggregation weights are allocated according to the proportion of data volume per transformer, with higher weights for larger data volumes, and the total weights are 1. Data privacy is protected using differential privacy technology, with a noise level set to 0.1.

[0024] The accuracy threshold of the model adaptability evaluation module is set at 92%. The accuracy of the edge model in identifying local abnormal signals is used as the evaluation index. After triggering the fine-tuning process, the edge model is initialized with the general model parameters optimized in the cloud and fine-tuned 30 times based on the latest local data of the past 3 months.

[0025] Preferably, the operation and maintenance decision-making stage receives the insulation status assessment results, remaining life prediction values ​​and location information output from the cloud, and combines the importance level of the transformer, the current load level, the power grid operating conditions and maintenance costs to set the weights of safety priority and economic priority through the analytic hierarchy process. The weights are dynamically adjusted according to the peak or off-peak periods of the power grid load. A differentiated operation and maintenance strategy library is generated based on weights, including strategies such as emergency power outage maintenance in case of severe degradation, load adjustment and insulation reinforcement in case of mild degradation, and environmental control and status tracking in case of environmental stress degradation. The implementation effect and potential risks of each strategy are predicted through a risk assessment model. Finally, the execution effect of the operation and maintenance strategy is fed back to the cloud platform to correct the parameters of the life prediction model and optimize the weight setting of the operation and maintenance decision model.

[0026] Specific measures for insulation strengthening treatment include: ① For slightly deteriorated winding coil insulation, the insulation strength is supplemented by vacuum injection of modified epoxy resin. The injection pressure is controlled at 0.1-0.2MPa, the injection temperature is 60-80℃, and the heat preservation and curing time is ≥24 hours. ② For areas with slight aging of the lead wire insulation layer, wrap aramid fiber reinforced insulation tape with 3-5 layers, each layer overlapping by 50%, and then encapsulate with heat shrink tubing. ③ The insulating oil of oil-immersed transformers is purified by a vacuum oil filter with a filtration accuracy of ≤5μm to remove moisture and impurities from the oil. At the same time, antioxidants are added to delay the aging of the insulating oil. Environmental control measures include: stabilizing the operating temperature at 40-80℃ through the transformer cooling system to avoid sudden cooling and heating; controlling the humidity of the surrounding environment at 30%-70% through the humidity control system; and reducing equipment vibration through shock-absorbing pads.

[0027] The risk assessment model adopts a fault tree combined with a Bayesian network architecture. The inputs are five types of parameters: insulation status level, remaining lifetime, equipment importance, grid load level, and maintenance cost. The outputs are two types of results: strategy implementation success rate and potential risk probability. For example, the emergency power outage maintenance strategy has a success rate of ≥98% and a potential risk probability of ≤1%, with risk types including grid load transfer failure and maintenance misoperation. The environmental control strategy has a success rate of ≥90% and a potential risk probability of ≤5%, with risk types including control effect not meeting expectations.

[0028] The model training data includes 1000 historical operation and maintenance cases, covering different state levels and power grid operating conditions. The training steps are as follows: ① Case data is standardized, with equipment importance quantified into levels 1-3 and load level quantified into 0.5-1.2; ② The Bayesian network structure learning uses the K2 algorithm, and the parameter learning uses maximum likelihood estimation; ③ Iterate training until the prediction error is ≤3% and then stop.

[0029] The feedback correction rules are as follows: if the actual remaining lifespan after strategy implementation deviates from the predicted value by more than 10%, the weight of the aging acceleration factor in the lifespan prediction model will be adjusted; if the risk probability of strategy implementation deviates from the predicted value by more than 2%, the safety / economic priority weights of the operation and maintenance decision will be recalculated.

[0030] The beneficial effects of this invention are as follows: 1. This invention arranges multi-dimensional sensing units (electrical, thermal, mechanical, and chemical) at key parts of the transformer to simultaneously collect various state parameters and establish spatial correlations. It uses a combination of mutual information entropy screening and redundancy elimination to extract key features such as the rate of change of dielectric loss tangent and the ratio of characteristic gas concentrations in the oil. It focuses on clarifying the threshold ranges and trend differences of parameters related to thermal aging and environmental stress. It uses a mechanism-data dual attention mechanism to enhance feature discrimination and then improves the support vector machine and Transformer ensemble model to achieve accurate identification of degradation types and subclasses such as thermal aging and sudden cooling / heating stress.

[0031] 2. This invention first utilizes the time delay difference of signals collected by a distributed sensor array, combined with the spatial coordinates of each unit, to initially delineate the degradation range using a time difference positioning algorithm, and then uses the cross-correlation function method to improve signal synchronization accuracy. Based on the actual parameters of the transformer insulation structure, a finite element forward model of electric field, temperature field, and stress field is established. After inputting the degradation type characteristic parameters, the simulation results are optimized by inversion using a genetic algorithm. Combining the signal amplitude attenuation law and the multi-physics field anomaly area, a Kriging spatial interpolation algorithm is used to achieve three-dimensional positioning, while simultaneously outputting a multi-physics field distribution cloud map, clearly presenting information such as field strength and temperature gradient, helping maintenance personnel to accurately locate weak parts such as winding coils and lead insulation, avoiding blind repairs, and reducing power outage costs and safety risks.

[0032] 3. This invention designs an edge-cloud collaborative architecture. The edge deploys preprocessing algorithms, preliminary identification models, and personalized lifespan prediction sub-models. Through model optimization, it achieves real-time alarms for abnormal signals and basic state assessments. The cloud stores historical data, runs 3D positioning and multiphysics simulation models, integrates monitoring data from multiple transformers using federated learning, iteratively optimizes the general model, and provides updated parameters to the edge. In the lifespan prediction stage, it integrates full lifecycle data, establishes a full-state dataset after preprocessing, captures state evolution patterns through a temporal attention LSTM network, and generates personalized models by combining transfer learning, outputting state level and remaining lifespan. In the operation and maintenance stage, it dynamically sets priorities based on factors such as equipment importance and grid conditions, generates differentiated strategies, and then corrects model parameters through closed-loop feedback to continuously improve operation and maintenance efficiency and transformer operational reliability. Attached Figure Description

[0033] Figure 1 This is the core flowchart of the insulation condition monitoring of the present invention; Figure 2 This is a flowchart of the edge-cloud collaborative architecture of the present invention; Figure 3 This is a flowchart of the closed-loop operation and maintenance decision-making process of this invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] like Figures 1 to 3 As shown, this embodiment of the invention provides a method for online monitoring of transformer insulation, including the following specific steps: In the sensor deployment stage, a micro dielectric sensing unit based on aramid nanopaper is embedded in key parts such as the gap between different winding coils and the lead wire insulation layer using laser positioning-assisted installation technology. This unit relies on the high temperature resistance and high specific surface area of ​​aramid nanopaper to monitor the changes in dielectric constant and conductivity of the insulating medium in real time. It can also adsorb characteristic products such as furfural produced by insulation aging and quantify the adsorption amount, thereby improving the stability and sensing sensitivity under high temperature conditions. A fiber optic grating sensor array is evenly distributed every 5 cm along the winding axis and every 3 cm in the radial direction to simultaneously collect local temperature and mechanical stress data; eight microelectromechanical systems (MEMS) vibration and electric field joint sensing units are evenly distributed around the inner wall of the oil tank and four are symmetrically arranged at the ends of the winding to capture the time-series correlation characteristics of local electric field distortion and vibration signal accompanying insulation degradation. By coordinating the layout of electrical, thermal, mechanical, and chemical multi-dimensional sensing units and combining laser ranging to achieve spatial coordinate calibration of each sensing unit with a positioning error ≤0.5mm, the synchronous acquisition and spatial correlation of multiple parameters of insulation status are realized, and the acquired multi-dimensional raw signals are output to the signal processing stage.

[0036] The performance parameters of the dielectric sensing unit are as follows: dielectric constant measurement range 2-10, conductivity measurement range 10-14-10-8 S / m, response time ≤50ms, and operating temperature range -40℃-150℃; the center wavelength of the fiber optic grating sensing array is 1550nm, temperature measurement range -20℃-120℃, measurement accuracy ±0.1℃, stress measurement range 0-1000MPa, and accuracy ±1MPa; the vibration measurement frequency range of the MEMS vibration and electric field combined sensing unit is 10Hz-10kHz, acceleration measurement range 0-50g, electric field measurement range 0-100kV / mm, and measurement accuracy ±0.5kV / mm.

[0037] During installation, the dielectric sensing unit is encapsulated and fixed with epoxy resin, the fiber optic grating sensing array is bonded to the winding surface with high-temperature adhesive, and the MEMS unit is fixed to the inner wall of the oil tank and the winding end with threaded clips to ensure that the sensing unit is in close contact with the monitoring part and does not affect the normal operation of the transformer.

[0038] To address the signal interference issues caused by complex electromagnetic environments such as closing impacts and harmonic interference in power systems, as well as the vibration of oil flow inside transformers, the signal processing stage first extracts features from the multi-dimensional original signals using a fusion model of convolutional neural networks and gated recurrent units. After pre-training with a large number of interference samples, the model can identify typical feature patterns of interference signals under different operating conditions, including the spike pulse characteristics of closing impacts, the fixed frequency characteristics of harmonic interference, and the random fluctuation characteristics of oil flow vibration. Based on the extracted interference features, the parameter matrix of the Kalman filter is dynamically adjusted to achieve adaptive suppression of different types of interference. Simultaneously, the empirical mode decomposition method is used to decompose non-stationary vibration interference into multiple intrinsic mode components and remove invalid components. Finally, a signal-to-noise ratio-based weighted fusion algorithm is used to perform multi-source fusion on the interference-free effective signal. The weights are dynamically allocated according to the real-time signal-to-noise ratio of each sensing unit. Then, the wavelet threshold enhancement algorithm is used to improve the signal amplitude and signal-to-noise ratio to ensure the accuracy of subsequent feature analysis. The processed effective signal is then output to the degradation identification stage.

[0039] The dynamic adjustment rules for Kalman filter parameters are as follows: when the closing impact spike pulse characteristic is identified, the process noise covariance Q is adjusted to 0.01 and the observation noise covariance R is adjusted to 0.1; when the fixed frequency characteristic of harmonic interference is identified, Q is adjusted to 0.005 and R is adjusted to 0.05; when the random fluctuation characteristic of oil flow vibration is identified, Q is adjusted to 0.02 and R is adjusted to 0.15.

[0040] The wavelet thresholding enhancement algorithm uses the sym8 wavelet basis and has a decomposition layer of 3. The threshold calculation adopts an adaptive method based on the signal standard deviation, with the threshold being 1.2 × signal standard deviation / √ln (number of sampling points). Soft thresholding is applied to the high-frequency coefficients after decomposition, while the low-frequency coefficients remain unchanged. The signal is then reconstructed through inverse wavelet transform.

[0041] The degradation identification stage constructs an insulation degradation type identification model that combines multi-feature fusion, attention mechanism, and aging mechanism constraints to solve the problem that existing technologies cannot accurately distinguish between aging degradation and environmental stress degradation. From the preprocessed multi-dimensional signals, a method combining mutual information entropy screening and redundancy removal is used to extract key features such as the change rate of dielectric loss tangent, partial discharge pulse characteristic parameters, the concentration ratio of characteristic gases in the oil, temperature gradient changes, and the adsorption amount of aging products of the nanosensor unit. Among these features, the distinction between thermal aging characteristic parameters (such as aging rate-related parameters based on the Arrhenius equation) and environmental stress characteristic parameters (such as instantaneous dielectric constant fluctuation amplitude and stress mutation coefficient) is strengthened, clarifying the threshold range and trend differences between the two types of features. The partial discharge pulse characteristic parameters include rise time, amplitude, and pulse frequency; the concentration ratio of characteristic gases in the oil includes CO / CO2 and CH4 / C2H2. Characteristic gases in the oil are collected using an embedded photoacoustic spectroscopy sensor with a sampling cycle of 1 hour / time. The sensor's detection limit is ≤0.1μL / L, and the measurement accuracy is ±2%. The concentration ratio is calculated using a volume ratio. The CO / CO2 ratio is calculated as "CO concentration (μL / L) ÷ CO2 concentration (μL / L)"; The CH4 / C2H2 ratio is calculated as "CH4 concentration (μL / L) ÷ C2H2 concentration (μL / L)"; When the C2H2 concentration is below the detection limit, this ratio is approximately calculated as "CH4 concentration ÷ 0.1" to ensure the continuity of characteristic parameters.

[0042] The specific implementation of mutual information entropy screening is as follows: each feature parameter is calculated by sliding calculation with a time window of 60 seconds. The mutual information entropy value is calculated based on the correlation between each feature and the degradation type label. The screening threshold is set to 0.6. Features with mutual information entropy values ​​≥ 0.6 are retained as candidate features. Redundancy removal employs the Pearson correlation coefficient method to calculate the correlation coefficient between candidate features. When the correlation coefficient between any two features is ≥0.85, the feature with the higher mutual information entropy value is retained, ultimately forming a non-repetitive, highly discriminative key feature set.

[0043] A dual attention mechanism based on mechanism and data is established. On the one hand, based on the chemical kinetic mechanism of insulation aging (e.g., the thermal decomposition law of cellulose) and the physical mechanism of environmental stress degradation (e.g., the thermal expansion and contraction law of the medium), a mechanism knowledge base is established to impose prior constraints on feature weights. On the other hand, the weights are dynamically optimized through a data-driven multi-head attention mechanism to strengthen features that are strongly correlated with the degradation type. For example, aging degradation corresponds to a stable increase in the CO / CO2 ratio and an increase in the adsorption amount of nanosensor units, while environmental stress degradation corresponds to instantaneous fluctuations in dielectric constant but no accumulation of aging products. Finally, the weighted fused features are input into an integrated model of an improved Support Vector Machine (SVM) and a Transformer model. The SVM is used for linear classification of features, while the Transformer model captures long-term temporal correlations of features. The outputs of both models are used to identify and classify two types of degradation and their subclasses through weighted voting decisions. The subclasses include thermal aging, electrical aging, sudden cooling and heating stress, and vibration stress. Finally, the degradation type and subclass identification results are output to the 3D localization stage.

[0044] The three-dimensional positioning stage proposes a three-dimensional spatial positioning method for insulation weak points by fusing signal time delay difference and multi-physics field forward and inverse modeling, which solves the problems of low positioning accuracy and failure to consider the collaborative effects of multiple fields in the existing technology. First, the time delay difference of the signals collected by different sensing units in the distributed sensing array is used. Combined with the spatial coordinates of each unit, the approximate range of the deteriorated area is initially determined by the time difference positioning algorithm. In order to improve the accuracy of time delay difference calculation, the cross-correlation function method is used to perform synchronous alignment processing on the signals, and the time delay difference calculation error is controlled within 1μs.

[0045] The specific operation of the cross-correlation function method is as follows: Select the signals collected by the two farthest sensing units in the distributed sensing array as references, calculate the cross-correlation function between the signals of the remaining sensing units and the reference signals, and take the time offset corresponding to the maximum value of the cross-correlation function as the signal delay correction value. To improve synchronization accuracy, the original signal is first subjected to a 50Hz low-pass filter to remove high-frequency noise, and then the cross-correlation function is calculated by sliding at a time step of 10μs. After synchronization and alignment, the timestamp deviation of the signals of all sensing units is ≤0.5μs, ensuring the accuracy of the time delay difference calculation.

[0046] Based on the actual size parameters and material properties of the transformer's internal insulation structure, a finite element forward model of the synergistic effects of electric field, temperature field, and stress field is established. The characteristic parameters of the identified degradation types, such as the thermal power change corresponding to aging degradation and the stress mutation parameters corresponding to environmental stress degradation, are used as input conditions. The equivalent parameters of the degradation region in the model are inverted and optimized through a genetic algorithm, and the simulation results of the multi-physics distribution of the degradation region are iteratively corrected to ensure the consistency between the simulation results and the actual monitoring data.

[0047] The model construction details are as follows: ① Tetrahedral unstructured mesh is adopted. The mesh size in the insulating medium area is ≤1mm, and the mesh size in the non-critical structural area is ≤5mm. The mesh quality standard is a twist degree of ≤0.3. ② First, perform steady-state calculations of the temperature field, and use the temperature distribution results as load inputs to the electric and stress fields. Then, solve the electric and stress fields through bidirectional coupling. Electric field distortion affects the dielectric properties of the medium, thereby changing the stress distribution, and vice versa. ③ The outer wall of the oil tank is set as a temperature boundary, with an ambient temperature of ±5℃. The winding ends are set as electric field Dirichlet boundaries, with the electric field strength value corresponding to the rated voltage. The contact surface between the insulation structure and the metal parts is set as a stress-fixed boundary to ensure that the model is consistent with the actual operating conditions of the transformer.

[0048] By combining the signal amplitude attenuation law of multiple sensing units, the abnormal area of ​​multi-physical field distribution, and the preliminary positioning range, the Kriging space interpolation algorithm is used to achieve accurate three-dimensional coordinate positioning of weak points in insulation, improving the positioning accuracy to the millimeter level. At the same time, the multi-physical field distribution cloud map of the positioning area is output, presenting information such as field strength distribution and temperature gradient of the deteriorated area. The obtained three-dimensional positioning results and deterioration type information are output to the lifetime prediction stage.

[0049] The specific steps of the time difference positioning algorithm are as follows: ① Select the three sensor units closest to the deteriorated area as positioning reference points; ② Calculate the signal delay differences Δt1, Δt2, and Δt3 between the target sensing unit and each reference point; ③ Based on the propagation speed v of electromagnetic waves in the transformer insulation medium, approximately 2 × 10⁻⁶ 8 m / s, converted into distance differences Δd1=Δt1×v, Δd2=Δt2×v, Δd3=Δt3×v; ④ Draw a sphere with the three reference points as the center and the corresponding distance difference as the radius. The intersection of the spheres is the preliminary positioning area.

[0050] The genetic algorithm parameters were set as follows: population size 50, number of iterations 100, crossover probability 0.8, mutation probability 0.05, and the fitness function was defined as "the sum of squared deviations between the simulated multiphysics distribution and the actual monitoring data". The iteration stopped when the deviation was less than 5%.

[0051] The Kriging spatial interpolation algorithm uses a spherical variogram, sets the search neighborhood to a 3×3×3 grid, and has an interpolation step size of 0.1 mm. Cross-validation is used to correct the interpolation accuracy, ensuring that the positioning error is ≤0.3 mm.

[0052] The life prediction stage integrates transformer factory test data, historical operation and maintenance data, and real-time data from online monitoring. It performs outlier removal, missing value completion, and normalization preprocessing on the data to establish a full-state dataset including healthy, slightly deteriorated, moderately deteriorated, and severely deteriorated states. The dataset is then classified and labeled according to transformer model, years of operation, and load characteristics. Outlier removal adopted the Grubbs criterion with a significance level of 0.05. Data was identified as outliers and removed when it deviated from the mean by more than three times the standard deviation. Missing value completion adopted the K-nearest neighbor interpolation method, selecting the 10 nearest neighbor time points as references and calculating the missing values ​​by distance weighting. Normalization adopted the min-max standardization method, mapping all data to the interval [0, 1]. The maximum value of parameters such as dielectric constant and conductivity was taken as the ultimate tolerance value of the transformer insulation medium, and the minimum value was taken as the initial measurement value of the new equipment, to ensure the consistency of the physical meaning of the data after standardization.

[0053] A temporal attention LSTM network is used to capture the evolution of insulation state parameters over time. An aging acceleration factor with the combined effect of multiple factors is embedded. The weights of the synergistic influence of temperature, humidity and electric field strength are quantified through orthogonal experiments. The loss function of the network is optimized based on the insulation aging mechanism to improve the model's fitting accuracy to the aging trend. A personalized transfer learning mechanism is established. Based on the general model of the same type of transformer, the model is fine-tuned by the historical data of a single transformer to generate a personalized life prediction model. The model outputs insulation condition assessment levels of 1-5 and remaining life under different operation and maintenance strategies, thereby improving the adaptability and prediction accuracy of the model on different individual transformers.

[0054] The remaining lifetimes listed below are predicted values ​​under normal operating conditions. The remaining lifetimes after maintenance need to be dynamically adjusted based on the effectiveness of the strategy implementation. Level 1 (Healthy): Dielectric constant 2.5-3.5, conductivity ≤10-12S / m, CO / CO2 ratio ≤0.03, aging product adsorption ≤0.1mg / g, partial discharge frequency ≤1 time / hour, remaining lifespan ≥15 years; Level 2 (slight degradation): Dielectric constant 3.6-4.5, conductivity 10⁻¹²-10⁻¹⁰ S / m, CO / CO₂ ratio 0.03-0.08, aging product adsorption 0.1-0.3 mg / g, partial discharge frequency 1-5 times / hour, remaining lifetime 10-15 years; Level 3 (moderate degradation): Dielectric constant 4.6-5.5, conductivity 10⁻¹⁰⁻¹⁰⁻⁸ S / m, CO / CO₂ ratio 0.08-0.15, aging product adsorption 0.3-0.6 mg / g, partial discharge frequency 5-10 times / hour, remaining lifetime 5-10 years; Level 4 (severe degradation): dielectric constant 5.6-6.5, conductivity 10⁻⁸-10⁻⁶ S / m, CO / CO₂ ratio 0.15-0.25, aging product adsorption 0.6-1.0 mg / g, partial discharge frequency 10-20 times / hour, remaining lifetime 1-5 years; Level 5 (Severe Deterioration): Dielectric constant ≥ 6.6, conductivity ≥ 10-6 S / m, CO / CO2 ratio ≥ 0.25, aging product adsorption ≥ 1.0 mg / g, partial discharge frequency ≥ 20 times / hour, remaining lifetime < 1 year.

[0055] In the architecture building phase, signal preprocessing algorithms, preliminary degradation type identification models, and personalized lifetime prediction sub-models are deployed at the edge (transformer local monitoring unit). Through model pruning and quantization optimization, computational overhead is reduced, enabling real-time alarms and basic status assessments of abnormal signals. The abnormal signal alarm threshold is set based on the health range of the full-state dataset: when the dielectric constant deviates from the health value by ±10%, the conductivity is higher than the health value by 50%, the amplitude of the partial discharge pulse is ≥500mV, the temperature gradient is ≥5℃ / cm, and the amount of aging products adsorbed is ≥0.5mg / g, a level 2 alarm is triggered. When any two of the above parameters exceed the limit simultaneously, or when the exceedance of a single parameter reaches twice the threshold, a Level 1 alarm is triggered. The alarm output includes local indicator lights and remote communication signals. A red light corresponds to a Level 1 alarm, and a yellow light corresponds to a Level 2 alarm. The alarm information includes the name of the abnormal parameter, the extent of exceedance, the time of occurrence, and the corresponding sensor unit number.

[0056] The model pruning adopts a structured channel pruning strategy: for the convolutional layers and fully connected layers of networks such as CNN and LSTM, the channels are sorted by importance (based on the sum of the absolute values ​​of the channel weights), and redundant channels with a weight summation value lower than 30% of the average value of all channels are removed. The quantization optimization adopts 8-bit integer quantization, which only quantizes the model weights and activation values. The bias parameters retain 32-bit floating-point type to balance the accuracy loss. After optimization, the storage size of the edge model is ≤50MB, the single inference time is ≤100ms, and the model accuracy decrease is controlled within 3%, which meets the real-time computing needs of the edge.

[0057] The preprocessed valid data, preliminary identification results, and model operation logs are encrypted and uploaded to the cloud platform. The cloud platform runs the complete 3D positioning model and multiphysics simulation model, and uses a distributed database to store historical data. The cloud integrates monitoring data from multiple transformers of the same type through federated learning algorithms, iteratively optimizes the general model while ensuring data privacy, and provides updated parameters for personalized models at the edge. A model adaptability evaluation module is configured to monitor the prediction accuracy of the edge model in real time. When the prediction accuracy of the edge model is lower than the set threshold, the fine-tuning process is automatically triggered to improve adaptability under different operating conditions and output accurate evaluation, prediction, and positioning information to the operation and maintenance decision-making stage.

[0058] The operation and maintenance decision-making stage receives insulation status assessment results, remaining life prediction values ​​and location information output from the cloud, and constructs an adaptive operation and maintenance decision-making model that combines multi-objective optimization and risk assessment. First, based on the importance level of the transformer (core hub / distribution network), the current load level, the power grid operating conditions and maintenance costs, the weights of safety priority and economic priority are set through the analytic hierarchy process, and the weights are dynamically adjusted according to the peak or low load periods of the power grid. A weighted, differentiated operation and maintenance strategy library is generated, including strategies such as emergency power outage maintenance for severe degradation, load adjustment and insulation reinforcement for mild degradation, and environmental control and status tracking for environmental stress degradation. The implementation effect and potential risks of each strategy are predicted through a risk assessment model. Finally, the execution effect of the operation and maintenance strategies, such as the recovery of insulation parameters after maintenance and the change in degradation rate after load adjustment, is fed back to the cloud platform to correct the parameters of the life prediction model, optimize the weight setting of the operation and maintenance decision model, improve operation and maintenance efficiency and transformer operation reliability, and realize the continuous iterative upgrade of the entire monitoring technology system.

[0059] The rules for constructing the judgment matrix for weight setting in the analytic hierarchy process are as follows: the comparison scale between safety priority and economic priority adopts the 1-9 scale method. During peak load periods (load factor ≥ 1.0), the scale of safety priority to economic priority is 7, meaning that safety priority is significantly important, with a weight of 0.8 and a weight of 0.2 for economic priority. During off-peak load periods (load factor ≤ 0.7), the scale is 3, meaning that safety priority is slightly more important, with a weight of 0.6 and a weight of 0.4 for economic priority. During normal load periods, the scale is 5, meaning that safety priority is relatively important, with a weight of 0.7 and a weight of 0.3 for economic priority.

[0060] The core evaluation indicators of the risk assessment model include: the rate of change of insulation degradation rate after strategy implementation, the probability of decline in power grid reliability, and the proportion of overrun of maintenance costs. The weights of each indicator are 0.5, 0.3, and 0.2, respectively. The comprehensive risk value is the sum of the actual values ​​of each indicator and their corresponding weights. When the comprehensive risk value is ≤0.1, it is judged as a low-risk strategy and should be implemented first.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online monitoring of transformer insulation, characterized in that, The specific steps include the following: Sensor deployment phase: Design a distributed intelligent array deployment scheme with multi-type sensor collaboration, deploy different functional sensor units in key insulation parts of the transformer, and realize synchronous acquisition and spatial correlation of multiple parameters of insulation status through multi-dimensional sensor collaborative layout and spatial coordinate calibration. Signal processing stage: A layered signal preprocessing mechanism is designed to address electromagnetic and vibration interference. Effective signals are obtained through interference feature identification, adaptive suppression, and multi-source signal fusion. Degradation identification stage: Establish an insulation degradation type identification model that combines multi-feature fusion, attention mechanism and aging mechanism constraints. Extract key features that distinguish aging from environmental stress degradation from preprocessed signals. Strengthen effective features through weight optimization driven by both mechanism and data. Use an integrated model to identify degradation types and subclasses and output the identification results. Three-dimensional positioning stage: Based on the time delay difference of distributed sensor signals, the deteriorated area is initially determined. A multi-physics field collaborative finite element forward model is established and inverted for optimization in combination with the insulation structure. The three-dimensional coordinate positioning of the weak point is realized through spatial interpolation algorithm, and the positioning results and deterioration type information are output. Lifetime prediction stage: Based on the insulation data of the whole life cycle, a condition assessment and lifetime prediction model is established. Through data preprocessing and time series feature capture, the condition level and remaining lifetime are output. Architecture building phase: Design an edge-cloud collaborative architecture, with the edge implementing real-time alarms for abnormal signals and basic status assessments, and the cloud completing in-depth model optimization and data storage; Operation and maintenance decision-making phase: Based on the cloud status assessment and location results, an adaptive operation and maintenance decision-making model is established, priorities are dynamically set and differentiated operation and maintenance strategies are generated, and model parameters are optimized through closed-loop feedback.

2. The online monitoring method for transformer insulation according to claim 1, characterized in that, During the sensor deployment phase, dielectric sensing units are embedded in key locations, including the gaps between winding coils and the lead insulation layer, to monitor changes in the dielectric constant and conductivity of the insulating medium and to quantify the adsorbed aging characteristic products. A fiber optic grating sensor array is arranged at a preset interval along the axial and radial directions of the winding to simultaneously collect local temperature and mechanical stress data; a MEMS vibration and electric field joint sensing unit is arranged on the inner wall of the oil tank and the end of the winding to capture the temporal correlation characteristics of electric field distortion and vibration signal. Spatial coordinate calibration is completed by combining multi-dimensional sensor collaborative layout with laser ranging, and the collected raw signals are output to the signal processing stage.

3. The online monitoring method for transformer insulation according to claim 2, characterized in that, The signal processing stage extracts typical feature patterns of interference signals through a fusion model, including the spike pulse characteristics of closing impact, the fixed frequency characteristics of harmonic interference, and the random fluctuation characteristics of oil flow vibration. Based on the extracted interference features and dynamic filtering parameters, adaptive interference suppression is achieved. Empirical mode decomposition is used to remove invalid vibration interference components. Multi-source fusion of the interference-removing signal is performed through a signal-to-noise ratio-based weighted fusion algorithm. The weights are dynamically allocated according to the real-time signal-to-noise ratio of the sensing unit. After signal enhancement processing, the effective signal is output to the degradation identification stage.

4. The online monitoring method for transformer insulation according to claim 3, characterized in that, The degradation identification stage employs a combination of mutual information entropy screening and redundancy removal to extract key features including the rate of change of dielectric loss tangent, partial discharge pulse characteristic parameters, the ratio of characteristic gas concentrations in the oil, temperature gradient changes, and the amount of aging products adsorbed. This strengthens the differentiation and extraction of thermal aging characteristic parameters and environmental stress characteristic parameters, and clarifies the threshold ranges and differences in the changing trends of the two types of features. A dual attention mechanism based on mechanism and data is established. Feature weights are constrained prior through a mechanism knowledge base, and weights are dynamically optimized by combining a data-driven attention mechanism. The fused features are input into an integrated model of an improved support vector machine and a Transformer model. The degradation type and subclass are identified through weighted voting decision-making, and the results are output to the 3D localization stage.

5. The online monitoring method for transformer insulation according to claim 4, characterized in that, The three-dimensional positioning stage utilizes the time delay difference of sensor signals combined with spatial coordinates to initially determine the deteriorated area through a time difference positioning algorithm, and uses the cross-correlation function method to achieve signal synchronization alignment. A finite element forward model of the synergistic effect of electric field, temperature field and stress field is established based on insulation structure parameters and material properties. The identified degradation type characteristic parameters are used as input, and the equivalent parameters of the degradation region are optimized by inversion through genetic algorithm. By combining the signal amplitude attenuation law of multiple sensing units, the abnormal distribution area of ​​multiple physical fields, and the preliminary positioning range, a spatial interpolation algorithm is used to realize the three-dimensional coordinate positioning of the weak point in insulation, and output the three-dimensional positioning result and the distribution information of multiple physical fields to the lifetime prediction stage.

6. The online monitoring method for transformer insulation according to claim 5, characterized in that, The life prediction stage integrates transformer life cycle data, performs outlier removal, missing value completion and normalization preprocessing on the data, establishes a full-state dataset and classifies and labels it. A temporal attention LSTM network is used to capture the evolution of insulation state parameters, and an aging acceleration factor is embedded to optimize the network's loss function. Establish a transfer learning mechanism to generate a lifetime prediction model, and output insulation status assessment level, remaining lifetime under different operation and maintenance strategies, and full life cycle data to the architecture building stage.

7. The online monitoring method for transformer insulation according to claim 6, characterized in that, During the architecture building phase, a signal preprocessing algorithm, a preliminary degradation type identification model, and a lifetime prediction sub-model are deployed at the edge. After optimization, real-time alarms for abnormal signals and basic status assessments are achieved. The pre-processed valid data, preliminary identification results and model running logs are encrypted and uploaded to the cloud platform. The cloud platform runs the 3D positioning model and multiphysics simulation model and stores historical data. The cloud-based system optimizes the general model through federated learning and provides updated parameters; it also configures a model adaptability evaluation module that automatically triggers a fine-tuning process when the prediction accuracy of the edge model falls below a set threshold.

8. The online monitoring method for transformer insulation according to claim 7, characterized in that, The operation and maintenance decision-making stage combines insulation condition assessment results, remaining life prediction values, location information, and power grid operating conditions, and sets safety priority and economic priority weights through the analytic hierarchy process. Based on weights, a differentiated operation and maintenance strategy library is generated, including emergency power outage maintenance, load adjustment, environmental control, and status tracking. The implementation effect and potential risks of the strategies are predicted through risk assessment. The execution effect of the operation and maintenance strategies is fed back to the cloud platform to correct the parameters of the life prediction model and optimize the weight settings of the operation and maintenance decision model.

Images