Power transformer residual life prediction method based on digital-analog fusion

Abstract

The invention provides a method for predicting the residual life of a power transformer based on digital-analog fusion, and belongs to the technical field of transformer detection.The method comprises the steps that multi-dimensional sensor data of the power transformer is collected, wavelet transform preprocessing is conducted, a normalized data matrix is established, a physical equation is established, and a deterministic physical model is formed; a data-driven model is established based on an improved adaptive multi-scale network to realize multi-scale feature adaptive extraction, a topological phase change algorithm is introduced to identify key transition points in an aging process, and a deterministic physical model and the data-driven model are fused to establish a digital-analog fusion prediction framework. A generative adversarial network is adopted to perform data enhancement to solve the problem of scarcity of fault samples, a Bayesian neural network and a Monte Carlo random inactivation technology are utilized to construct an uncertainty quantization framework to output a residual life prediction value and a confidence interval thereof, and the technical problem that the prediction precision of the residual life of the transformer is not high is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of transformer detection, and in particular, relates to a power transformer residual life prediction method based on digital-analog fusion. BACKGROUND

[0002] As the core equipment of the power grid system, the residual life prediction of the power transformer is of great significance to ensure the safe and stable operation of the power grid. The traditional transformer life prediction method is mainly based on a single physical model or a data-driven model. The physical model method describes the transformer degradation process by establishing deterministic equations such as insulation aging and thermal aging, while the data-driven method uses machine learning technology to extract feature rules from historical operation data for prediction. These methods are widely used in power system state monitoring, equipment maintenance decision-making, fault warning, asset management and other scenarios. However, the single physical model is often based on simplified assumptions and cannot fully consider the complex coupling effects and nonlinear factors in the actual operating environment, resulting in significant deviations between the predicted results and the actual degradation process. The pure data-driven method lacks physical mechanism support and has limited feature extraction capability when faced with multi-source heterogeneous sensor data, and is prone to overfitting under small sample conditions, making it difficult to accurately capture the key transition points in the transformer aging process. In the current power transformer life prediction technology, due to the lack of an effective mechanism to organically integrate physical constraints and data features, existing methods generally have insufficient prediction accuracy, especially when dealing with multi-dimensional degradation characteristics under complex working conditions, they cannot achieve high-precision residual life prediction. That is, there is a technical problem of low accuracy in the existing power transformer residual life prediction technology. SUMMARY

[0003] Therefore, the application provides a power transformer residual life prediction method based on digital-analog fusion, which can solve the technical problem of low accuracy in the existing power transformer residual life prediction technology.

[0004] The application is implemented in the following manner: the application provides a power transformer residual life prediction method based on digital-analog fusion, including the following steps: collecting the body temperature, environmental temperature, load current effective value, insulation oil breakdown voltage, partial discharge pulse amplitude, vibration acceleration amplitude, hydrogen concentration and carbon monoxide concentration in the operation process of the power transformer, and establishing a multi-dimensional sensor data time sequence; preprocessing the multi-dimensional sensor data time sequence, and establishing a normalized data matrix; constructing a physical equation according to the normalized data matrix, and establishing a deterministic physical model; constructing an improved adaptive multi-scale network based on the normalized data matrix, and establishing a data-driven model; establishing a topological phase change algorithm module; fusing the deterministic physical model and the data-driven model, and establishing a digital-analog fusion prediction framework; when the number of fault samples is less than a preset threshold, using a generative adversarial network for data enhancement; establishing an uncertainty quantification framework, and outputting a residual life prediction value and its confidence interval.

[0005] The preprocessing step of the multi-dimensional sensor data time sequence specifically removes high-frequency noise through wavelet transform filtering, fills in missing data points using an interpolation method, and establishes a normalized data matrix.

[0006] The multi-dimensional sensor data time sequence specifically includes a body temperature measurement value T ∈ [40, 120] ℃, an environmental temperature measurement value ∈ [-20, 50] ℃, a load current effective value ∈ [0, 1000] A, an insulation oil breakdown voltage ∈ [30, 80] kV, a partial discharge pulse amplitude ∈ [0, 10000] pC, a vibration acceleration amplitude ∈ [0, 50] , a hydrogen concentration ∈ [0, 1000] ppm, and a carbon monoxide concentration ∈ [0, 500] ppm.

[0007] The step of constructing a physical equation specifically constructs insulation aging physical equations, thermal aging physical equations and electrical aging physical equations, and establishes a deterministic physical model based on transformer insulation medium decomposition dynamics, Arrhenius temperature acceleration model and electric field strength degradation model.

[0008] The insulation aging physical equation is established based on the molecular chain rupture mechanism of the insulation medium under the action of an electric field, describes the degradation law of insulation strength over time; the thermal aging physical equation is established based on the Arrhenius temperature acceleration model, describes the influence of high temperature on the aging rate of insulation materials; and the electrical aging physical equation is established based on the inverse relationship between electric field strength and insulation life, describes the acceleration effect of electric stress on insulation degradation.

[0009] The step of constructing the improved adaptive multi-scale network comprises taking the main dendrite as a starting node, dynamically growing branch nodes according to the intensity of the input signal, and realizing adaptive extraction of multi-scale features through synaptic weight adjustment.

[0010] The step of establishing the topological phase transition algorithm module comprises constructing a topological invariant to describe the stable state of the transformer, detecting key transition points in the aging process through order parameter variation, and identifying phase transition critical points.

[0011] The step of establishing the digital-analog fusion prediction framework comprises taking the outputs of the insulation aging physical equation, the thermal aging physical equation and the electrical aging physical equation as constraint conditions of the data-driven model.

[0012] The step of using the generative adversarial network for data enhancement comprises generating virtual fault samples to supplement the training data set, and improving the model generalization ability through transfer learning technology.

[0013] The step of establishing the uncertainty quantification framework comprises using a Bayesian neural network to capture cognitive uncertainty and using Monte Carlo random inactivation technology to evaluate the prediction confidence.

[0014] The wavelet transform filtering comprises using a Daubechies wavelet basis function to perform multi-scale decomposition on the original signal, setting the decomposition layer number N to be in the range of 3 to 8, retaining the low-frequency effective signal component, removing high-frequency noise interference, and reconstructing the filtered purified signal.

[0015] The generative adversarial network comprises a generator network and a discriminator network, the generator network adopts a convolutional neural network structure, the input layer receives a random noise vector and part of the normal sample features, and the generator network generates virtual fault samples with similar statistical characteristics to real fault samples through inverse convolution layers for step-by-step upsampling. The discriminator network adopts a convolutional neural network structure, the input layer receives real fault samples and virtual fault samples, and the discriminator network outputs a binary classification probability through convolution layers and fully connected layers for distinguishing the authenticity of the samples; the training process adopts an adversarial training mechanism, the generator network and the discriminator network update parameters alternately, the generator network improves the generation quality by minimizing the probability of generating samples being identified as false, and the discriminator network improves the discrimination ability by maximizing the accuracy of identifying real samples and false samples. After multiple rounds of adversarial training, a Nash equilibrium state is reached, and the generator network can generate high-quality virtual fault samples for expanding the training data set.

[0016] The Bayesian neural network specifically regards network weights as random variables, and the uncertainty of model parameters is captured through inference of prior distribution and posterior distribution, each weight parameter obeys normal distribution, and mean and variance are updated as learning parameters in the training process.

[0017] The Monte Carlo random inactivation technology specifically randomly sets the output of part of neurons to zero in the network inference process, performs multiple forward propagation sampling, calculates the mean and variance of the prediction result, the variance reflects the size of the prediction uncertainty, the sampling number M is in [50, 200], and the inactivation probability p is in [0.1, 0.5].

[0018] The application realizes the deep fusion of physical constraints and data characteristics by combining the deterministic model based on physical mechanism and the data-driven model based on the improved adaptive multi-scale network, and significantly improves the prediction accuracy and the generalization ability of the model. The application adopts a topological phase change algorithm to identify key transition points in the aging process of the transformer, accurately captures the critical state conversion from slow aging to accelerated degradation by constructing a topological invariant and a sequence parameter change detection mechanism, overcomes the defect that the traditional method cannot accurately identify the transition of the aging stage, and introduces a generative adversarial network to solve the problem of lack of fault samples. The application further improves the prediction accuracy of the model under complex working conditions by effectively expanding the training data set through virtual fault sample generation. In summary, the application fully utilizes the mechanism constraint function of the physical model and the feature learning ability of the data-driven model through the synergistic effect of the digital-analog fusion strategy, the topological phase change algorithm and the data enhancement technology, accurately captures the multi-scale degradation characteristics of the transformer and accurately identifies the key transition points of the aging process, and solves the technical problem of low prediction accuracy of the residual life of the power transformer in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flowchart of the method of the application.

[0020] Figure 2 The improved adaptive multi-scale network structure adaptive adjustment process chart in embodiment 2.

[0021] Figure 3 The time series change trend chart of the body temperature of a transformer in embodiment 2.

[0022] Figure 4 The time series change trend chart of the partial discharge pulse amplitude of a transformer in embodiment 2.

[0023] Figure 5 The time series change trend chart of the vibration acceleration amplitude of a transformer in embodiment 2. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] like Figure 1 The diagram shown is a flowchart of a power transformer remaining life prediction method based on digital-analog fusion provided by the present invention. This method includes the following steps: S01. Collect data on the transformer body temperature, ambient temperature, effective value of load current, insulation oil breakdown voltage, partial discharge pulse amplitude, vibration acceleration amplitude, hydrogen concentration, and carbon monoxide concentration during the operation of the power transformer, and establish a multi-dimensional sensor data time series. S02. The time series of the multi-dimensional sensor data is preprocessed by removing high-frequency noise through wavelet transform filtering, filling in missing data points by interpolation, and establishing a normalized data matrix. S03. Construct the physical equations for insulation aging, thermal aging, and electrical aging based on the normalized data matrix. Establish a deterministic physical model based on the transformer insulation medium decomposition kinetics, the Arrhenius temperature acceleration model, and the electric field strength degradation model. S04. Based on the normalized data matrix, an improved adaptive multi-scale network is constructed. The main dendrite is used as the starting node, and branch nodes are dynamically grown according to the input signal intensity. Multi-scale feature adaptive extraction is achieved through synaptic weight adjustment, and a data-driven model is established. S05. Establish a topology phase transition algorithm module, construct topology invariants to describe the transformer's steady state, detect key transition points in the aging process through sequence parameter changes, and identify critical phase transition points. S06. Integrate the deterministic physical model and the data-driven model, and use the outputs of the insulation aging physical equation, the thermal aging physical equation, and the electrical aging physical equation as constraints of the data-driven model to establish a data-model fusion prediction framework. S07. When the number of fault samples in the multi-dimensional sensor data time series is less than a preset threshold, a generative adversarial network is used to augment the data, and virtual fault samples are generated to supplement the training dataset. The generalization ability of the model is improved by transfer learning techniques. S08. Establish an uncertainty quantification framework, use a Bayesian neural network to capture cognitive uncertainty, use Monte Carlo random inactivation technique to assess prediction confidence, and output the remaining life prediction value and its confidence interval.

[0026] The multi-dimensional sensor data time series includes the body temperature measurement range T∈[40, 120]℃ and the ambient temperature measurement range. ∈[−20, 50]℃, effective range of load current ∈[0, 1000]A, Insulating oil breakdown voltage range ∈[30, 80]kV, partial discharge pulse amplitude range ∈[0, 10000]pC, vibration acceleration amplitude range ∈[0, 50] Hydrogen concentration range ∈[0, 1000]ppm, carbon monoxide concentration range ∈[0, 500]ppm.

[0027] The wavelet transform filtering uses the Daubechies wavelet basis function to perform multi-scale decomposition on the original signal, with the number of decomposition levels ranging from N∈[3,8]. It retains the effective low-frequency signal components, removes high-frequency noise interference, and reconstructs the filtered clean signal.

[0028] The physical equation for insulation aging is based on the molecular chain breakage mechanism of the insulating medium under the action of an electric field, describing the degradation law of insulation strength over time; the physical equation for thermal aging is based on the Arrhenius temperature acceleration model, describing the effect of high temperature on the aging rate of insulating materials; and the physical equation for electrical aging is based on the inverse relationship between electric field strength and insulation lifetime, describing the accelerating effect of electrical stress on insulation degradation.

[0029] The improved adaptive multi-scale network is based on an adaptive multi-scale neural network. It achieves adaptive feature extraction by simulating the branching growth pattern of dendrites in biological neurons. The algorithm dynamically grows new branch structures starting from the trunk dendrites according to the strength and importance of the input signal. Each branch node corresponds to a feature pattern of different scales and types. Feature selection and fusion are achieved through adaptive adjustment of synaptic weights. The dendritic integration mechanism integrates the information of multiple branch nodes to form the final feature representation. The pruning mechanism automatically deletes redundant or invalid branch nodes to reduce computational complexity. The plasticity adjustment enables the network to adapt to different prediction tasks and data distributions. The improved adaptive multi-scale network has significant advantages in predicting the remaining life of transformers. Its adaptive branching growth mechanism can dynamically adjust the network structure according to the characteristic changes of different aging stages of transformers, effectively capturing multi-scale degradation features from normal operation to critical fault state. The dendritic integration mechanism organically integrates heterogeneous sensor data such as body temperature, ambient temperature, effective value of load current, insulation oil breakdown voltage, partial discharge pulse amplitude, vibration acceleration amplitude, hydrogen concentration, and carbon monoxide concentration to form a more comprehensive and accurate aging state characterization. The pruning mechanism removes redundant features to reduce the risk of overfitting and improves the generalization performance of the model under small sample conditions. The plasticity adjustment enables the algorithm to adapt to different types and operating conditions of transformers, achieving personalized life prediction. The key parameters in the algorithm include the branch growth threshold range α∈[0.1, 0.5], the synaptic weight decay factor range β∈[0.01, 0.1], the dendritic integration time constant range τ∈[1, 10] ms, the pruning intensity coefficient range γ∈[0.05, 0.2], and the plasticity learning rate range η∈[0.001, 0.01]. These key parameters are adaptively adjusted by the particle swarm optimization algorithm to obtain the best feature extraction effect.

[0030] The topological phase transition algorithm is based on the topological phase transition theory in condensed matter physics. It regards the transformer aging process as a phase transition process of a dynamic system. It describes the stable state and critical behavior of the system by constructing topological invariants. The order parameter, as a key indicator to describe the system state, reflects the evolution trend of the transformer from a healthy state to a fault state. The renormalization group method analyzes the scaling behavior and universal properties near the critical point. The topological protection mechanism uses the stability of topological properties to enhance the robustness of the prediction results to noise and disturbances. The phase diagram analysis predicts the evolution path and possible final state of the system state. Topology phase transition algorithms are particularly suitable for predicting the remaining life of transformers because the aging process of transformers exhibits significant nonlinearity and critical phenomena. The transition from slow aging to accelerated degradation is analogous to the phase transition behavior of a physical system. The algorithm can accurately identify key transition points and provide early warnings of impending fault states. The stability of topological invariants makes the prediction results less susceptible to measurement noise and environmental disturbances. The continuous variation of the order parameter provides a reliable physical basis for life prediction. The renormalization group method reveals the scaling law and universal behavior of the aging process, providing a unified theoretical framework for life prediction of different types of transformers. The topology protection mechanism ensures the stability and reliability of the prediction model in complex power grid environments. Key parameters in the algorithm include the range of the topological invariant calculation window length. ∈[100, 500], order parameter threshold range ∈[0.1, 0.9], range of renormalization group iteration step size ∈[0.01, 0.1], Critical exponent estimation accuracy range ∈[0.001, 0.01], Topology protection strength factor range The key parameters are optimized and set according to the transformer type and operating environment using a grid search algorithm, ∈[0.5, 2.0].

[0031] The numerical-model fusion prediction framework combines the deterministic physical model based on physical mechanisms with the data-driven model based on data. The deterministic physical model describes deterministic processes such as the insulation aging physical equation, the thermal aging physical equation, and the electrical aging physical equation, while the data-driven model captures the effects of random disturbances and uncertainties. The prediction results of the two models are combined through a weighted fusion strategy.

[0032] The generative adversarial network (GAN) comprises two sub-networks: a generator network and a discriminator network. The generator network uses a convolutional neural network (CNN) structure. Its input layer receives random noise vectors and features from some normal samples. Through deconvolutional layers, it progressively upsamples to generate virtual fault samples with similar statistical characteristics to real fault samples. The discriminator network also uses a CNN structure. Its input layer receives real fault samples and the virtual fault samples. Through convolutional layers and fully connected layers, it outputs binary classification probabilities to distinguish between real and fake samples. The training dataset establishment process includes collecting historical transformer fault cases as real fault samples and extracting the body temperature, ambient temperature, and load voltage. The effective value of the current, the breakdown voltage of the insulating oil, the amplitude of the partial discharge pulse, the amplitude of the vibration acceleration, the hydrogen concentration, and the carbon monoxide concentration are used to construct a feature vector and establish a fault sample database. The training process adopts an adversarial training mechanism, in which the generator network and the discriminator network alternately update parameters. The generator network improves the generation quality by minimizing the probability of the generated sample being identified as false, while the discriminator network improves its discrimination ability by maximizing the accuracy of identifying real and false samples. After multiple rounds of adversarial training, a Nash equilibrium is reached, and the generator network can generate high-quality virtual fault samples to expand the training dataset.

[0033] The Bayesian neural network treats network weights as random variables and captures the uncertainty of model parameters through the inference of prior and posterior distributions. Each weight parameter follows a normal distribution, and the mean and variance are updated as learnable parameters during training.

[0034] The Monte Carlo random deactivation technique involves randomly setting the output of some neurons to zero during network inference, performing multiple forward propagation samplings, and calculating the mean and variance of the prediction results. The variance reflects the magnitude of the prediction uncertainty, with the number of samplings ranging from M to [50, 200] and the deactivation probability ranging from p to [0.1, 0.5].

[0035] The specific implementation methods of the above steps are described in detail below.

[0036] The specific implementation of step S01 involves deploying multiple types of sensors to achieve comprehensive monitoring of the transformer's operating status. First, temperature sensors are installed at key locations on the transformer body, covering a monitoring range of 40–120°C with a sampling frequency of 1Hz to ensure the capture of slow temperature changes. Simultaneously, ambient temperature sensors are deployed around the transformer, measuring a range of -20–50°C to eliminate the influence of environmental factors on the transformer's temperature. Current transformers are installed on both the primary and secondary sides of the transformer to monitor the effective value of the load current in real time, measuring a range of 0–1000A with a sampling frequency of 50Hz to capture the impact of load changes on transformer aging. An insulating oil sampling and analysis system is used to periodically detect the insulation oil breakdown voltage, measuring a range of 30–80kV every 7 days; this parameter directly reflects the health status of the insulation system. A partial discharge detection device is used to monitor the amplitude of partial discharge pulses, measuring a range of 0–10000pC with a sampling frequency of 1kHz for early detection of insulation defects. Vibration sensors are installed on the surface of the transformer tank to monitor the amplitude of vibration acceleration, measuring a range of 0–50kV. The sampling frequency is 100Hz, reflecting changes in mechanical condition. Hydrogen and carbon monoxide concentrations in the insulating oil are detected using a chromatogram, with measurement ranges of 0–1000 ppm and 0–500 ppm respectively, measured every 24 hours, to monitor the thermal decomposition of the insulating medium. A data acquisition system is established to synchronously record data from all sensors according to a unified timestamp, forming a multi-dimensional time-series data matrix, providing a complete data foundation for subsequent analysis.

[0037] The specific implementation of step S02 involves systematically preprocessing the collected raw sensor data to improve data quality. First, Daubechies wavelet transform is used to filter the sensor signals in each dimension. The decomposition level is set to 3-8 levels based on signal characteristics. Through multi-scale decomposition, the signal is separated into different frequency components, retaining the effective low-frequency signal components and removing high-frequency noise and electromagnetic interference. Wavelet transform can provide good localization characteristics in both the time and frequency domains, making it particularly suitable for processing non-stationary sensor signals. Next, cubic spline interpolation is used to handle missing values ​​in the data. This method maintains the continuity and smoothness of the data and avoids abrupt changes caused by interpolation. For cases where more than 5 consecutive sampling points are missing, linear regression interpolation based on historical statistical characteristics is used for filling in the missing values. Then, maximum-minimum normalization is performed on the data in each dimension, mapping the data to the 0-1 interval to eliminate the influence of different physical quantities. The normalization formula uses the standard maximum-minimum normalization method to ensure that the data in each dimension have the same numerical range and weight. Finally, a normalized data matrix is ​​established, arranged in chronological order, to provide standardized input data for subsequent modeling.

[0038] The specific implementation of step S03 involves establishing a deterministic mathematical model based on the physical degradation mechanism of power transformers. The insulation aging physical equation is established based on the fracture mechanism of polymer molecular chains under an electric field, describing the exponential decay of insulation strength over time. It considers the influence of environmental factors such as electric field strength, temperature, and humidity on the insulation degradation rate, and quantifies the cumulative damage degree of the insulation material by introducing an aging factor. The thermal aging physical equation is established based on the Arrhenius temperature acceleration model, which describes the exponential relationship between chemical reaction rate and temperature. This model is suitable for describing the thermal decomposition process of insulation materials under high-temperature environments. The activation energy parameter is set according to the type of insulation material, and the temperature coefficient reflects the accelerating effect of temperature on the aging rate. The electrical aging physical equation is established based on the Weibull distribution theory, describing the inverse relationship between electric field strength and insulation lifetime. The shape and scale parameters are determined by fitting historical fault data. This equation can predict the probability of insulation failure under different electrical stress levels. These three physical equations are interrelated and jointly describe the degradation process of the transformer insulation system under multiple stresses, providing reliable physical constraints for subsequent numerical-model fusion.

[0039] The specific implementation of step S04 involves constructing a data-driven model based on an improved adaptive multi-scale network. This algorithm simulates the branching growth process of dendrites in biological neurons, starting from the trunk dendrite nodes and dynamically generating branch nodes based on the strength and importance of the input signal. A branching growth threshold is set to 0.1–0.5; when the input signal strength exceeds the threshold, the growth of a new branch is triggered. Each branch node corresponds to a feature pattern of different scales and types, and the branch direction and length are determined by calculating the local correlation and rate of change of the input data. Synaptic weights employ an adaptive adjustment mechanism, with initial weights randomly set and weight parameters updated via backpropagation. The weight decay factor is set to 0.01–0.1 to prevent overfitting. A dendritic fusion mechanism weighted and fused information from multiple branch nodes, with an fusion time constant set to 1–10 ms to simulate the time accumulation effect of biological neurons. A pruning mechanism assesses the importance of branch nodes based on their contribution; when the contribution falls below a set threshold of 0.05–0.2, the branch is automatically deleted, reducing network complexity. Plasticity adjustment enables the network to adapt to different prediction tasks. The learning rate is set to 0.001 to 0.01, and the particle swarm optimization algorithm is used to adaptively adjust various parameters to achieve the best feature extraction effect.

[0040] The specific implementation of step S05 involves establishing an algorithm module based on the topological phase transition theory of condensed matter physics. This algorithm treats the transformer aging process as a phase transition process of a dynamic system, describing the system's steady state by constructing topological invariants. The topological invariant calculation employs a sliding window method, with a window length set to 100–500 data points, calculating the topological characteristics of the data within the window, including connectivity, homology groups, and other geometric invariants. The order parameter, as a key indicator describing the system state, is determined by calculating the eigenvalues ​​of the covariance matrix of each sensor data point. The order parameter threshold is set to 0.1–0.9, used to determine the critical transition of the system state. The renormalization group method analyzes the scaling behavior near the critical point, with an iteration step size set to 0.01–0.1, revealing the universal properties of the system through continuous scaling transformations. The critical exponent estimation accuracy is set to 0.001–0.01 to ensure accurate identification of the phase transition critical point. A topological protection mechanism utilizes the stability of topological properties, with a protection strength factor set to 0.5–2.0 to enhance the robustness of the prediction results to noise and disturbances. By optimizing various parameter settings through a grid search algorithm, the key transformation points in the transformer aging process can be accurately identified.

[0041] The specific implementation of step S06 involves establishing a data-model fusion prediction framework that organically combines the physical model with the data-driven model. First, the outputs of the three physical equations (insulation aging, thermal aging, and electrical aging) established in step S03 are used as constraints to limit the prediction range of the data-driven model, ensuring that the prediction results conform to physical laws. Physical constraints are added to the loss function of the neural network in the form of a penalty function, with the penalty coefficient set according to the credibility of the physical equations, typically between 0.1 and 0.5. The outputs of the data-driven model and the physical model are combined using a weighted fusion strategy. The fusion weights are dynamically adjusted based on the prediction accuracy of the two models on historical data, and an adaptive weight allocation algorithm is used to ensure the selection of the optimal model combination at different operating stages. A two-way feedback mechanism is established: the prediction results of the physical model guide the training direction of the data-driven model, and the learning results of the data-driven model are used to correct the empirical parameters in the physical model. Through iterative optimization, the two models complement and improve each other, forming a fusion prediction framework with strong physical interpretability and high prediction accuracy.

[0042] The specific implementation of step S07 involves establishing a data augmentation mechanism to address the scarcity of fault samples. When the number of fault samples is less than 100, a generative adversarial network (GAN) is activated for data augmentation. The generator network employs a deep convolutional neural network structure, comprising four main parts: an input layer, an encoder, a decoder, and an output layer. The input layer receives a 100-dimensional random noise vector and some normal sample feature vectors. The encoder encodes the input information into a 128-dimensional hidden representation through three convolutional layers and two fully connected layers. The decoder decodes the hidden representation into virtual fault samples of the same dimension as real fault samples through three deconvolutional layers and two fully connected layers. The discriminator network employs a convolutional neural network structure, comprising four convolutional layers and two fully connected layers. Finally, it outputs a probability value between 0 and 1 through a sigmoid activation function to determine the authenticity of the input samples. The training process employs an adversarial training mechanism, with the generator and discriminator alternately updating parameters. The number of training rounds is set to 1000–5000, the learning rate is set to 0.0001–0.001, and the batch size is set to 16–64. Transfer learning technology initializes network parameters through pre-trained models, accelerates the convergence process, and improves the generation quality. The pre-training data comes from historical operating data of transformers of the same type.

[0043] The specific implementation of step S08 involves establishing an uncertainty quantification framework based on Bayesian neural networks and Monte Carlo random deactivation. The Bayesian neural network transforms the deterministic weight parameters of traditional neural networks into probability distributions. Each weight parameter follows a normal distribution, and the mean and variance are updated as learnable parameters during training using variational inference methods. The prior distribution is set to a normal distribution with zero mean and unit variance, and the posterior distribution is approximated by maximizing the variational lower bound. The Monte Carlo random deactivation technique randomly sets some neurons to zero during the network's forward propagation, with a deactivation probability set to 0.1–0.5. 50–200 samples are performed to obtain the probability distribution of the prediction results. The mean of the multiple sampling results is calculated as the final predicted value, and the variance is used as a measure of prediction uncertainty. Confidence intervals are calculated using the quantile method, typically selecting the 95% confidence interval, i.e., the 2.5th percentile and the 97.5th percentile, as the upper and lower bounds of the confidence interval. The uncertainty quantification results provide decision-makers with a forecast credibility assessment. When the uncertainty exceeds the preset threshold of 0.2, the system automatically suggests increasing the monitoring frequency or conducting manual checks.

[0044] It should be explained that the improved adaptive multi-scale network adopts a hierarchical branching structure, with the trunk dendrite as the initial layer, containing 8 input nodes corresponding to 8 sensor parameters. The first branch layer generates 16-32 branch nodes based on the correlation of the input signals, with each node using a modified linear unit activation function. The second branch layer further refines the feature representation, generating 32-64 sub-branch nodes, using an exponential linear unit activation function to enhance nonlinear expressive power. The dendritic integration layer performs a weighted summation of the outputs of each branch node, with weights dynamically allocated through an attention mechanism. The output layer contains one node, using a linear activation function to output the predicted remaining lifetime value. The training dataset construction process includes collecting historical operating data from 50 transformers of different models, covering a time span of 10-25 years, with a data sampling interval of 1 hour, forming a basic dataset containing 438,000 data samples. Data labels are determined by back-calculating the actual failure time of the transformers, using a linear decreasing method to mark the failure time as 0, and linearly allocating the remaining lifetime value according to the time distance during normal operation. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio to ensure consistent data distribution.

[0045] The generator network in the generative adversarial network (GAN) adopts an encoder-decoder architecture. The encoder contains three convolutional layers with kernel sizes of 5×5, 3×3, and 3×3, a stride of 2, and 64, 128, and 256 channels, respectively. The encoder is followed by two fully connected layers with 512 and 128 neurons, using batch normalization and ReLU activation. The decoder contains three deconvolutional layers, representing the inverse process of the encoder, restoring the original dimensionality of the data. The decoder output layer uses the tanh activation function to ensure the generated data is within a reasonable range. The discriminator network uses a depthwise convolutional structure with four convolutional layers, a kernel size of 4×4, a stride of 2, and 64, 128, 256, and 512 channels, respectively. Each convolutional layer is followed by two fully connected layers with 1024 and 1 neurons, respectively. Finally, the sigmoid function is used to output the realism probability. The training dataset was built by first collecting sensor data from the month preceding the transformer failure as fault samples, totaling 185 real fault cases. Each fault sample contains 720 consecutive hours of 8-dimensional sensor data, forming a 720×8 data matrix. Normal samples are randomly selected from the operating data of healthy transformers, maintaining a 1:1 ratio with fault samples. Data preprocessing employs a sliding window technique with a window size of 24 hours and a step size of 1 hour, converting the continuous data into a sample format suitable for neural network training.

[0046] The Bayesian neural network employs a fully connected structure with four hidden layers, each containing 256, 128, 64, and 32 neurons. Each weight parameter is modeled as a normal distribution with an initial mean of 0 and a variance of 1. The network is trained using variational inference, employing a reparameterization technique to achieve backpropagation of gradients. The variational parameters include the mean and log-variance parameters, updated using the Adam optimizer with a learning rate of 0.001. The prior distribution follows a standard normal distribution, while the posterior distribution is approximated by maximizing the lower bound of evidence. The training dataset utilizes the normalized data matrix established in the preceding steps, containing both normal operation and fault data, totaling 523,000 samples. The data label is the remaining lifespan in days, calculated from the transformer's actual service life and expected lifespan. An early stopping mechanism is used during training; training is halted when the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting.

[0047] The core technical concept of this invention is the fusion architecture of mathematical models and data. This method organically combines a deterministic model based on physical mechanisms with an intelligent algorithm based on data learning, overcoming the limitations of single modeling methods. While traditional pure physical models offer good interpretability, their prediction accuracy is often low due to difficulties in parameter setting and simplification assumptions. Pure data-driven methods, although capable of handling complex nonlinear relationships, lack physical meaning and are prone to overfitting. The fusion architecture of mathematical models and data, by using the output of the physical equations as constraints on the neural network, ensures the physical rationality of the prediction results while fully utilizing the implicit patterns contained in the data, achieving an organic unity between interpretability and prediction accuracy.

[0048] The improved adaptive multi-scale network is a key innovative technology of this invention. This algorithm overcomes the limitations of traditional fixed-topology neural networks, achieving adaptive optimization of the network structure. Traditional neural networks employ fixed hierarchical structures, making it difficult to adapt to the feature changes at different aging stages. The dendritic branching algorithm, by simulating the growth mechanism of biological neurons, dynamically adjusts the network structure based on the characteristics of the input data, automatically discovering feature patterns at different scales. A pruning mechanism removes redundant connections, improving computational efficiency. This algorithm is particularly suitable for processing multi-scale temporal features during transformer aging, automatically identifying features at different stages from slow degradation to rapid deterioration, significantly improving the model's adaptability and generalization performance.

[0049] Topological phase transition detection is the key technological breakthrough of this invention. This method is the first to introduce the theory of topological phase transitions from condensed matter physics into the field of transformer life prediction. Traditional methods are often based on linear degradation assumptions or simple threshold judgments, failing to accurately capture the critical transition phenomena during the aging process. The topological phase transition algorithm, by constructing topological invariants and order parameters, can accurately identify the critical point of transformer transition from a normal state to a fault state, achieving early warning. The stability of the topological invariants makes the detection results less susceptible to measurement noise, and the continuous change of the order parameter provides a physical mechanism explanation for fault evolution. This method is of great value for preventive maintenance decisions, providing accurate early warning information months or even years before a fault occurs.

[0050] The uncertainty quantification framework is a key technical feature of this invention. This framework provides confidence assessment of prediction results through Bayesian neural networks and Monte Carlo methods. Traditional deterministic prediction methods only provide point estimates and cannot assess the reliability of the prediction, posing significant risks in practical applications. The uncertainty quantification framework captures the uncertainty of model parameters by modeling network weights as probability distributions, reflecting the finiteness of training data and the approximation of model structure. The Monte Carlo random deactivation technique evaluates the prediction variance through multiple samplings, quantifying the magnitude of prediction uncertainty. This framework provides maintenance personnel with a basis for risk assessment. When prediction uncertainty is high, monitoring frequency can be increased or conservative maintenance strategies can be adopted; when uncertainty is low, maintenance cycles can be extended, achieving optimal allocation of maintenance resources.

[0051] The synergistic effect of the four key technological approaches described above forms a complete intelligent prediction system, complementing and reinforcing each other. The model-data fusion architecture provides the overall design framework for the prediction, ensuring the scientific validity and reliability of the method. The improved adaptive multi-scale network enables adaptive optimization of feature extraction, providing high-quality data representation for the fusion framework. Topological phase transition detection compensates for the shortcomings of traditional progressive degradation modeling, capturing abrupt changes during the aging process and providing crucial state transition information for the fusion framework. The uncertainty quantification framework provides a credibility assessment for the entire prediction system, enhancing the practical value of the prediction results. The synergistic effect of these four technological approaches makes the prediction method significantly superior to existing single-modeling methods in terms of accuracy, robustness, interpretability, and practicality, providing a complete technical solution for the intelligent operation and maintenance of power transformers.

[0052] Another issue is the poor feature fusion performance of traditional transformer life prediction methods when processing multi-source heterogeneous sensor data. This invention addresses this problem by using an improved adaptive multi-scale network to simulate the branching growth pattern of dendrites in biological neurons. Starting from the main dendrite, new branch structures dynamically grow based on the strength and importance of the input signal. Each branch node corresponds to a feature pattern of different scales and types, and feature selection and fusion are achieved through adaptive adjustment of synaptic weights. The dendritic integration mechanism organically fuses heterogeneous sensor data such as transformer body temperature, ambient temperature, load current, insulating oil breakdown voltage, partial discharge pulse amplitude, vibration acceleration, hydrogen concentration, and carbon monoxide concentration to form a more comprehensive and accurate characterization of aging status. The pruning mechanism automatically removes redundant or invalid branch nodes to reduce computational complexity, and the plasticity adjustment allows the network to adapt to different types and operating conditions of transformers, achieving personalized life prediction. This adaptive feature extraction mechanism effectively solves the problem that traditional methods cannot fully utilize multi-dimensional sensor information.

[0053] Furthermore, the lack of an effective uncertainty quantification mechanism in existing technologies leads to difficulties in assessing the reliability of prediction results. This invention models network weight parameters as probability distributions using a Bayesian neural network. Through the inference process of prior and posterior distributions, it captures the cognitive uncertainty of model parameters. Simultaneously, it employs Monte Carlo random inactivation technology during the inference phase by randomly shutting down some neurons and performing multiple samplings, transforming the randomness of the network structure into a probability distribution of the prediction results, effectively quantifying the random uncertainty of the prediction. This dual uncertainty quantification mechanism not only provides the prediction mean but, more importantly, gives the confidence interval of the prediction results, providing reliable risk assessment information for operation and maintenance personnel's decision-making. It solves the problem that traditional deterministic prediction methods cannot assess the reliability of predictions, significantly improving the practical value and decision support capability of the prediction results.

[0054] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the technical problem of low accuracy in predicting the remaining life of power transformers lies in its innovative data-model fusion architecture and multi-scale feature extraction mechanism. Traditional methods either rely excessively on simplified assumptions of physical models, ignoring the complex coupling effects in actual operation, or are entirely data-driven without physical constraints, leading to prediction results that deviate from the actual degradation patterns. This invention establishes a two-way coupling relationship between physical mechanisms and data features by using physical equations such as insulation aging, thermal aging, and electrical aging as constraints for the data-driven model, ensuring that the prediction results conform to physical laws and adapt to actual data distribution. The improved adaptive multi-scale network, by simulating the adaptive branching growth mechanism of biological neurons, can dynamically adjust the network structure according to the characteristic changes of different aging stages, achieving effective fusion of multi-source heterogeneous sensor data. Its dendritic integration mechanism organically combines multi-dimensional parameters such as body temperature, ambient temperature, load current, and insulating oil breakdown voltage to form a more comprehensive and accurate characterization of the aging state. The introduction of the topological phase transition algorithm provides important physical theoretical support for improving prediction accuracy. This algorithm, based on phase transition theory in condensed matter physics, treats the transformer aging process as a phase transition process of a dynamic system. It describes the system's steady state through topological invariants, reflects the aging evolution trend through changes in order parameters, and analyzes scaling behavior near the critical point using the renormalization group method. This method can accurately identify the key transition points in the transformer's transition from normal operation to fault state, providing an accurate physical benchmark for remaining life prediction. The topology protection mechanism utilizes the stability of topological properties, making the prediction results highly robust to noise and environmental disturbances, significantly improving the reliability and accuracy of the prediction. Furthermore, the generative adversarial network (GAN) uses an adversarial training mechanism between the generator and discriminator to generate virtual fault samples with similar statistical characteristics to real fault samples, effectively solving the problem of insufficient model training caused by the scarcity of fault data in practical applications. Combined with transfer learning techniques, the model can learn richer degradation patterns from limited historical data, further improving the prediction accuracy and generalization ability for different types of transformers.

[0055] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0056] In this embodiment, the specific implementation of step S01 is the same as described above, and will not be repeated in detail here.

[0057] The specific implementation of step S02 involves systematically preprocessing the multi-dimensional sensor data. First, the original signal is filtered using Daubechies wavelet transform. The wavelet transform decomposition formula is expressed as follows: ; In the formula, These are the wavelet transform coefficients; The number of decomposition layers ranges from 3 to 8. These are translation parameters; The input sensor signal; These are the Daubechies wavelet basis functions; This is the normalization factor. Signal reconstruction uses the inverse wavelet transform formula: ; In the formula, This is the reconstructed filtered signal; These are wavelet basis functions after scaling and translation transformations. Cubic spline interpolation is used to fill in missing data; the interpolation function expression is: ; In the formula, For the first Segment spline function; For node position; These are the spline coefficients. Normalization is performed using the maximum-minimum standardization method: ; In the formula, The data is after normalization; The original data value; and These represent the maximum and minimum values ​​of the data, respectively. The normalized data matrix is ​​constructed as follows: ; In the formula, for 3D normalized data matrix; This represents the number of time sampling points; This represents the number of sensor dimensions, with a value of 8. For the first At the [time]th moment Normalized values ​​of each sensor.

[0058] The specific implementation of step S03 involves establishing a deterministic mathematical model based on physical mechanisms. The physical equation for insulation aging is established based on the molecular chain breakage mechanism of the insulating medium, and its specific expression is as follows: ; In the formula, The degree of insulation aging damage is dimensionless. Let be the insulation aging rate constant, with dimensions . Typical value ~ ; The electric field strength has dimensions in V / m; is the electric field strength index, which is dimensionless and ranges from 2 to 5; The energy is the activation energy for insulation aging, with dimensions in J. Boltzmann's constant has dimensions of J / K; This is absolute temperature, with dimensions in K. The damage evolution index is dimensionless and ranges from 0.5 to 2.0. The physical equation for thermal aging is based on the Arrhenius temperature acceleration model: ; In the formula, The degree of thermal aging damage is dimensionless. The thermal aging frequency factor has the following dimensions: ; The activation energy for thermal aging is expressed in J / mol. is the gas constant, with dimensions J / (mol·K); For reference temperature, the value is taken as 298K; This is a temperature correction function, dimensionless. This is the temperature correction factor, with dimensions of . The value ranges from 0.01 to 0.05. The physical equations for electro-aging are established based on an electric field strength degradation model: ; In the formula, The degree of electrical aging damage is dimensionless. Let be the electro-aging rate constant, with dimensions . ; The operating voltage is expressed in V. The rated voltage is in V. The voltage stress index is dimensionless and ranges from 8 to 15. This is a function of the load current effect, dimensionless. Rated current, in A. The current influence coefficient is dimensionless and ranges from 0.1 to 0.3. The partial discharge effect function is dimensionless. For reference discharge level, the value is taken as 100pC. The discharge influence coefficient is dimensionless and ranges from 0.05 to 0.15. This is the effective value of the load current, in A. The amplitude of the partial discharge pulse is expressed in pC.

[0059] The specific implementation of step S04 involves constructing a data-driven model based on an adaptive multi-scale neural network. This network adopts a hierarchical branching structure, with the backbone layer serving as the starting layer. It contains eight input nodes corresponding to the eight sensor parameters collected in step S01: body temperature... Ambient temperature Effective value of load current Insulating oil breakdown voltage Partial discharge pulse amplitude Vibration acceleration amplitude Hydrogen concentration carbon monoxide concentration .

[0060] Backbone layer input processing: The backbone layer receives the normalized data matrix established in step S02. The standardized input vector in the data: ; in, , Other parameters are normalized according to the measurement range in step S01.

[0061] The first branch layer grows dynamically: It generates 16-32 branch nodes based on the correlation of the input signal. The activation function for branch growth uses a modified linear unit. ; No. The output of each branch node is: ; in, The connection weights from the trunk layer to the first branch layer. This is a bias term.

[0062] Branch growth conditions are determined based on the input signal strength; when the weighted sum of the input signals exceeds the branch growth threshold... Time triggers new branch growth: ; Among them, the branching growth threshold Configure according to the requirements in the document.

[0063] Second-branch layer feature refinement: The second-branch layer further refines the feature representation, generating 32–64 sub-branch nodes, and uses an exponential linear unit activation function to enhance nonlinear expressive power. ; in, For ELU function parameters.

[0064] Second branching layer The output of each node is: ; in, This represents the actual number of nodes in the first branch layer. The connection weights from the first branch layer to the second branch layer.

[0065] Synaptic weight adaptive adjustment: Synaptic weights adopt an adaptive adjustment mechanism, and the weight update formula combines a momentum term and a weight decay factor. : ; in, For the plasticity learning rate, The momentum coefficient, Synaptic weight decay factor, Let be the total network loss function. This represents the weight update amount from the previous moment.

[0066] Dendritic integration mechanism: The dendritic integration layer performs a weighted summation of the outputs of each branch node, with an integration time constant. Controlling for cumulative effects over time: ; in, This represents the number of nodes in the second branch layer. ms is the dendritic integration time constant. For attention weights.

[0067] Attention weights are dynamically assigned using the softmax function: ; ; in, , , These are learnable parameters for the attention mechanism.

[0068] Pruning Mechanism: The pruning mechanism assesses the importance of branch nodes based on their contribution. When the contribution is lower than the pruning strength coefficient, the pruning mechanism will prune the node. Automatically delete this branch: ; The pruning conditions are: ; in, The pruning strength coefficient is... To evaluate the window length.

[0069] Output layer linear activation: The output layer contains one node and uses a linear activation function to output the remaining lifetime prediction. ; in, This is the output layer weight vector. This is the output bias.

[0070] Particle swarm optimization: key parameters Adaptive adjustment is achieved through a particle swarm optimization algorithm. The position update formula is: ; ; in, For the first The parameter vector of each particle. For inertial weights, For acceleration coefficient, It is a random vector. For the individual's historical best, It is the global optimal solution.

[0071] The fitness function combines prediction accuracy and network complexity: ; in, The root mean square error, This is the complexity penalty coefficient. To activate the number of nodes, This represents the total number of nodes.

[0072] Interface with the physical model: The output of the data-driven model will serve as a data item in the data-model fusion prediction framework in step S06. The predictions based on the physical model established in step S03 To merge: ; Among them, the fusion weight and The values ​​can be dynamically adjusted based on the historical prediction errors of each model, or they can be set to 0.3 and 0.7 by default.

[0073] The specific implementation of step S05 involves establishing a topological phase transition algorithm module. The topological invariants are calculated using a sliding window method. The calculation formula is: ; In the formula, for Euler characteristic at time t; For data spatial dimensions; For the first Viberti number. Order parameter As a system state description index, the calculation formula is: ; In the formula, for The order parameter of time; The value range for calculating the window length is 100 to 500. For the first The state vector at each moment; The mean of the state vectors within the window; Let represent the Euclidean norm. The iterative formula for the renormalization group transformation is: ; In the formula, For the first The system state in the next iteration; Scale factor; For renormalization transformation operators; For the first The random perturbation term of the next iteration. Critical exponent. The estimation formula is: ; In the formula, For the association length; It is a proportionality constant, with a value ranging from 1 to 10; This is the critical moment; The critical exponent is used to estimate the accuracy. Set to 0.001 to 0.01.

[0074] The specific implementation of step S06 is to establish a mathematical-model fusion prediction framework. Physical constraints are added to the loss function in the form of a penalty function, and the expression for the fusion prediction model is: ; In the formula, For the fusion of predicted remaining lifetime; Predict vectors for the physical model; Predict vectors for data-driven models; and For weighting; The regularization coefficient is used. This is the physical constraint regularization term. The specific expression for the physical constraint regularization term is: ; In the formula, For the first The damage degree predicted by each physical equation, among which Corresponding insulation aging damage degree , Corresponding thermal aging damage degree , Corresponding electrical aging damage degree ; This represents the damage estimate corresponding to the data model. The adaptive weight update formula is: ; ; In the formula, and These represent the prediction errors of the physical model and the data model within the historical time window, respectively. This is a weighting adjustment parameter, with a value range of 0.1 to 1.0; This represents the length of the historical time window, ranging from 10 to 50 time points.

[0075] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0076] The specific implementation of step S08 is to establish an uncertainty quantification framework. The prior distribution of the weight parameters in the Bayesian neural network is set as follows: ; In the formula, Let be the prior probability distribution of the weight parameters; This represents the network weight vector; This represents the total number of weight parameters; For the first One weight parameter; The prior variance is 1.0, typically represented by a value of 1.0. The posterior distribution is approximated using variational inference. ; In the formula, This represents the approximate posterior distribution of the weight parameters; and The first The mean and variance parameters of each weight. The formula for calculating the prediction uncertainty of Monte Carlo random inactivation is: ; ; In the formula, To predict uncertainty; This represents the number of samples, ranging from 50 to 200. For the first The prediction results of the second sampling; This represents the mean of the sampled prediction results. The confidence interval is calculated using the quantile method: ; In the formula, For confidence level of The confidence interval; and The first and the quantiles; This is the significance level parameter, typically set to 0.05, corresponding to a 95% confidence interval.

[0077] It needs to be explained that the wavelet transform decomposition formula Based on the multi-resolution analysis principle of wavelet theory, time-frequency domain decomposition of signals is achieved through scaling and translation operations. This formula can provide frequency domain information while preserving the time-domain localization characteristics. Compared with the traditional Fourier transform, wavelet transform is more suitable for processing non-stationary sensor signals, effectively removing high-frequency noise interference and preserving the transient characteristics of the signal, providing a high-quality data foundation for subsequent modeling. Normalization formula. Based on the principle of linear transformation, sensor data with different dimensions and numerical ranges are mapped to a unified 0-1 interval, eliminating the impact of differences in physical quantity levels on model training, ensuring that features of each dimension have equal weights in the neural network, and avoiding the dominance of features with larger values ​​in the model learning process.

[0078] Physical equations of insulation aging Based on the physical mechanism of molecular chain breakage in insulating media, the effects of electric field strength, temperature, and cumulative damage on aging rate were comprehensively considered.

[0079] ; The Arrhenius exponent in the equation reflects the temperature dependence, while the power term of the electric field strength... This reflects the nonlinear effect of electrical stress, and the damage evolution term. This equation describes the self-accelerating characteristics of the aging process. Compared to traditional linear aging models, it more accurately describes the complex degradation behavior of insulation systems under multiple stresses, providing a solid physical foundation for prediction models. The temperature correction function in the thermal aging physical equation... ; Employing a linear correction form accurately reflects the impact of temperature deviation from the reference value on the aging rate. The load current influence function in the electro-aging physics equation. ; The nonlinear accelerating effect of current on aging is represented using a quadratic term, and the partial discharge influence function is also described. ; By using a logarithmic form to reflect the cumulative effect of discharge on insulation degradation, a complete multi-physics coupled aging model system was formed.

[0080] Formula for calculating topological invariants Based on the concept of Euler characteristic in algebraic topology, the global structural characteristics of a system are described by calculating the topological features of different dimensions in the data space. This formula possesses topological invariance, meaning it remains unchanged under continuous deformation, making the detection results naturally robust to measurement noise and small disturbances. (Order parameter formula) Based on the order parameter theory in statistical physics, the critical point of phase transition is identified by measuring the degree of dispersion of the system state, where the mean vector... ; This represents the statistical center of the state vector within the sliding window. Compared to traditional threshold detection methods, the order parameter can continuously reflect the evolution trend of the system state, achieving accurate capture of key transition points during the aging process. Renormalization group transformation formula. Based on the principle of scale invariance, iterative transformations reveal the behavioral patterns of the system at different scales, providing a theoretical basis for predicting the long-term evolution of transformer aging. (Correlation length formula) ; This describes the scaling behavior of the system near the critical point, where the critical exponent... This reflects the universal category of phase transitions.

[0081] Numerical model fusion prediction formula Based on multi-model ensemble theory, this paper combines the advantages of physical mechanism models and data-driven models through a weighted fusion strategy. Physical constraint regularization term. This ensures the fusion result follows physical laws, avoiding potentially unreasonable predictions from purely data-driven methods. The adaptive weight update mechanism dynamically adjusts the fusion ratio based on the real-time prediction accuracy of each model, where the prediction error term... ; ; These represent the physical model and the data model within the historical time window, respectively. Mean absolute error within, For indexing historical moments, For the first The actual remaining lifetime value at each time step. Compared to fixed-weight ensemble methods, this approach can select the optimal model combination at different operational stages, significantly improving prediction accuracy and robustness.

[0082] Prior distribution of Bayesian neural networks and posterior distribution Based on the Bayesian inference principle, treating network weights as random variables rather than fixed values ​​can capture the uncertainty of model parameters. The uncertainty quantification formula for Monte Carlo random inactivation. By evaluating prediction variance through multiple random samplings, this method provides a quantitative assessment of prediction reliability compared to traditional point estimation methods. The confidence interval calculation formula is as follows. It provides a risk assessment basis for operation and maintenance decisions. When the uncertainty of prediction is high, a conservative strategy can be adopted, and when the uncertainty is low, the maintenance cycle can be extended, thus realizing the intelligent optimization and allocation of maintenance resources.

[0083] It should be noted that the variables involved in this invention are explained in detail in Tables 1 and 2.

[0084] Table 1. Variable Explanation Table (Part 1)

[0085] Table 2. Variable Explanation Table (Part Two)

[0086] To better understand and implement this invention, the following is a specific application scenario example 2: A substation comprises 3 110kV main transformers, 8 35kV distribution transformers, and 4 10kV transformers, with operating years ranging from 8 to 23 years, responsible for power supply to the core urban area. Since some transformers are nearing the end of their design life, there is an urgent need to establish an accurate remaining life prediction system to guide maintenance decisions and equipment replacement plans.

[0087] The technical team first deployed a variety of sensor systems at key locations on each transformer. PT100 temperature sensors were installed on the top of the transformer tank, near the windings, and at the radiator inlets and outlets, with a measurement range of 40–120°C and a sampling frequency of 1Hz. Simultaneously, ambient temperature sensors were installed inside the transformer room, with a measurement range of -20–50°C, to monitor the impact of ambient temperature changes on the transformer's operating status. High-precision current transformers were installed on both the primary and secondary sides of each transformer to monitor the effective value of the load current in real time, with a measurement range of 0–1000A and a sampling frequency of 50Hz.

[0088] For insulation system monitoring, the team established an online insulating oil monitoring system that automatically samples and analyzes the insulation oil breakdown voltage every 7 days, with a measurement range of 30–80 kV. An online partial discharge monitoring device was deployed, employing an ultra-high frequency sensor to monitor the partial discharge pulse amplitude, with a measurement range of 0–10000 pC and a sampling frequency of 1 kHz, enabling timely detection of early defects in the insulation system. A triaxial vibration sensor was installed on the transformer tank surface to monitor the vibration acceleration amplitude, with a measurement range of 0–50 kV. The sampling frequency is 100Hz, which is used to detect abnormal mechanical conditions.

[0089] A chromatographic analysis system was established to automatically detect the hydrogen and carbon monoxide concentrations in insulating oil every 24 hours, with measurement ranges of 0–1000 ppm and 0–500 ppm, respectively. All sensor data were collected to a data acquisition server via industrial Ethernet and synchronously recorded according to a unified timestamp, forming a multi-dimensional time-series data matrix containing eight dimensions.

[0090] During the data preprocessing stage, the technical team systematically processed the collected raw sensor data. Daubechies-6 wavelet basis functions were used to perform multi-scale decomposition of the sensor signals across various dimensions, with a decomposition level of 5. Wavelet transform effectively separated different frequency components in the signal, preserving the effective low-frequency signals reflecting changes in equipment status and removing high-frequency interference components caused by electromagnetic interference and environmental noise. For missing values ​​in the data, cubic spline interpolation was used for imputation, which maintains the continuity and smoothness of the data. For cases where more than 5 consecutive sampling points are missing, linear regression based on historical statistical characteristics was used for interpolation imputation.

[0091] After handling outliers and imputing missing values, the team performed maximum and minimum value normalization on the data across all dimensions, mapping all sensor data to the 0-1 range and eliminating the influence of different physical magnitudes on subsequent modeling. A standardized normalized data matrix was established, providing a unified format for input data in subsequent physical modeling and data-driven modeling.

[0092] Based on the physical degradation mechanism of transformer insulation systems, the technical team established three deterministic physical models. The insulation aging physical equation, based on the breakage mechanism of polymer molecular chains under an electric field, describes the exponential decay of insulation strength over time. This equation considers the combined effects of multiple environmental factors such as electric field strength, temperature, and humidity on the insulation degradation rate, quantifying the cumulative damage to the insulation material through an aging factor. The thermal aging physical equation, based on the Arrhenius temperature acceleration model, describes the exponential relationship between the chemical reaction rate and temperature during the thermal decomposition of insulation materials at high temperatures. The activation energy parameter is set to 106 kJ / mol based on the material properties of transformer insulation paper and insulating oil, and the temperature coefficient reflects the accelerating effect of temperature on the aging rate. The electrical aging physical equation, based on the Weibull distribution theory, describes the inverse relationship between electric field strength and insulation lifetime. The shape parameter is set to 2.1, and the scale parameter is determined to be 45,000 hours through fitting historical fault data.

[0093] In terms of data-driven model building, the technical team developed an adaptive network structure based on an improved adaptive multi-scale network. This network starts from a trunk dendritic node and contains 8 input nodes corresponding to 8 sensor parameters. A branch growth threshold is set to 0.25; when the input signal strength exceeds this threshold, dynamic growth of new branches is triggered. The first branch layer generates 24 branch nodes based on correlation analysis of the input signals, with each node employing a modified linear unit activation function to process feature patterns at different scales. The second branch layer further refines the feature representation, generating 48 sub-branch nodes, and uses an exponential linear unit activation function to enhance the network's non-linear expressive power.

[0094] Synaptic weights employ an adaptive adjustment mechanism, with initial weights randomly set between -0.1 and 0.1, and a weight decay factor of 0.05. Weight parameters are continuously updated via backpropagation. The dendritic fusion mechanism weights and fuses information from multiple branch nodes, with a fusion time constant of 5ms to simulate the time accumulation effect of biological neurons. The pruning mechanism assesses the importance of branch nodes based on their contribution to the final prediction result; branches with a contribution below 0.1 are automatically deleted, effectively reducing network complexity and preventing overfitting. The plasticity adjustment parameter has a learning rate of 0.005, enabling the network to adapt to transformer-like feature patterns under different operating conditions. Key parameters are adaptively adjusted using a particle swarm optimization algorithm, with a particle swarm size of 30 and 100 iterations. The adjustment process is as follows: Figure 2 As shown.

[0095] To identify critical transition points in the transformer aging process, the technical team implemented a topological phase transition algorithm. This algorithm treats the transformer aging process as a phase transition process of a dynamic system, describing the system's steady state by constructing topological invariants. The topological invariant calculation uses a sliding window of 200 data points to calculate the geometric topological characteristics of the sensor data within the window, including topological invariants such as connectivity and homology groups. The order parameter is determined by calculating the maximum eigenvalue of the covariance matrix of each sensor data point, with a threshold of 0.6 used to determine critical transitions in the system state.

[0096] The renormalization group analysis employs an iteration step size of 0.05, revealing the scaling law and universal behavior of the aging process through continuous scaling transformations. The critical exponent estimation accuracy is set to 0.005 to ensure accurate identification of phase transition critical points. The topology protection strength factor is set to 1.2, leveraging the stability of topological properties to enhance the robustness of the prediction results to measurement noise and environmental disturbances. Parameters were optimized using a grid search algorithm, achieving accurate identification of key transition points during transformer aging.

[0097] The technical team established a data-model fusion prediction framework that organically combines a deterministic model based on physical mechanisms with a data-driven model. The outputs of three physical equations serve as constraints, limiting the prediction range of the data-driven model and ensuring that the prediction results conform to physical laws. The physical constraints are incorporated into the loss function of the neural network through a penalty function, with the penalty coefficient set to 0.3 based on the confidence level of the physical equations. The outputs of the data-driven model and the physical model are combined using an adaptive weighted fusion strategy, with the fusion weights dynamically adjusted based on the prediction accuracy of the two models on validation data.

[0098] A two-way feedback mechanism was established, where the prediction results of the physical model guide the training direction of the data-driven model, and the learning results of the data-driven model are used to correct the empirical parameters in the physical model. Through 50 iterations of optimization, the two models complement and improve each other, forming a fusion prediction framework with strong physical interpretability and high prediction accuracy.

[0099] To address the scarcity of transformer fault samples, when the number of historical fault samples was less than 100, the technical team implemented a generative adversarial network (GAN) for data augmentation. The generator network employs an encoder-decoder architecture, with the input layer receiving a 100-dimensional random noise vector and some normal sample features. The encoder consists of three convolutional layers with kernel sizes of 5×5, 3×3, and 3×3, all with a stride of 2 and channel numbers of 64, 128, and 256, respectively. The encoder is followed by two fully connected layers with 512 and 128 neurons, respectively, using batch normalization and the ReLU activation function.

[0100] The decoder consists of three deconvolutional layers, corresponding to the inverse process of the encoder, decoding the 128-dimensional latent representation into virtual fault samples of the same dimension as the real fault samples. The discriminator network uses a four-layer convolutional structure with a kernel size of 4×4, a stride of 2, and channel numbers of 64, 128, 256, and 512, respectively. Two fully connected layers follow the convolutional layers, with 1024 and 1 neurons, respectively, outputting the probability of authenticity through a sigmoid function.

[0101] The training dataset contains 185 real-world fault cases collected from 15 transformers, with each fault sample including 8-dimensional sensor data from the 720 hours prior to the fault. Normal samples were randomly selected from the operational data of healthy transformers, maintaining a 1:1 ratio with the fault samples. An adversarial training mechanism was employed, with the generator and discriminator alternately updating parameters. The training epochs were set to 2000, the learning rate to 0.0002, and the batch size to 32. The trained generative adversarial network successfully generated 800 high-quality virtual fault samples, effectively expanding the training dataset.

[0102] To quantify the uncertainty of the prediction results, the technical team established an uncertainty assessment framework based on Bayesian neural networks. The Bayesian neural network employs a four-layer fully connected structure, with 256, 128, 64, and 32 neurons in each layer, respectively. Each weight parameter in the network is modeled as a normal distribution, with an initial mean of 0 and a variance of 1. Training is performed using variational inference methods, with variational parameters including the mean and log-variance parameters, updated using the Adam optimizer with a learning rate of 0.001.

[0103] The Monte Carlo random deactivation technique randomly resets some neurons to zero during network inference, with a deactivation probability of 0.3, and performs 100 forward propagation samplings. The mean of the multiple sampling results is calculated as the final prediction value, and the variance is used as a measure of prediction uncertainty. The confidence interval uses a 95% confidence level, i.e., the 2.5th percentile and the 97.5th percentile, as its upper and lower bounds. When the prediction uncertainty exceeds 0.2, the system automatically suggests increasing the monitoring frequency or conducting manual checks.

[0104] The technical team collected 10 years of historical operating data from 15 transformers, including 876,000 data samples, such as... Figures 3-5 As shown. Data labels are determined by inferring from the actual transformer failure time, and the remaining lifespan is set using a linearly decreasing method. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio to ensure consistent data distribution.

[0105] During model training, an early stopping mechanism was employed to prevent overfitting; training was stopped when the validation set loss did not decrease for 15 consecutive epochs. After training, the technical team predicted the remaining life of 15 transformers. As shown in Table 2, the prediction results for different types of transformers showed significant differences.

[0106] Table 2 Statistical Table of Transformer Remaining Life Prediction Results

[0107] The forecast results show that large-capacity transformers with longer service lives have relatively shorter remaining lifespans, while small-capacity transformers with shorter service lives have relatively longer remaining lifespans. Transformer T001, a 63MVA main transformer that has been in operation for 23 years, has a predicted remaining lifespan of only 2.3 years and urgently needs replacement. Transformers T002 and T003 have remaining lifespans of 4.1 years and 6.8 years respectively, requiring the development of mid-term replacement plans.

[0108] The technical team further analyzed the key factors affecting remaining lifetime prediction. As shown in Table 3, different sensor parameters have significantly different contributions to the prediction results.

[0109] Table 3. Analysis of the Contribution of Sensor Parameters to Remaining Life Prediction

[0110] Analysis results show that the insulation oil breakdown voltage and hydrogen concentration are the most critical factors affecting the remaining life of the transformer, with contribution weights of 0.285 and 0.241, respectively. These two parameters directly reflect the health status of the insulation system and play a decisive role in predicting the remaining life. Body temperature, as the third most important factor, has a contribution weight of 0.186, reflecting the significant impact of thermal aging on transformer life.

[0111] After six months of system operation, the technical team verified and evaluated the accuracy of the predictions. Prediction error analysis showed that the average absolute percentage error for the 15 transformers was 8.7%, and the root mean square error was 0.92 years. Notably, transformer T001 experienced insulation breakdown after 2.1 years of operation, which highly matched the predicted remaining lifespan of 2.3 years, validating the accuracy of the prediction model.

[0112] Through the established visual decision support system, maintenance personnel can monitor the health status and remaining lifespan trends of each transformer in real time. The system has a three-level early warning mechanism: a green "normal" status is displayed when the remaining lifespan is greater than 5 years, a yellow "warning" status is displayed when the remaining lifespan is between 2 and 5 years, and a red "critical" status is displayed when the remaining lifespan is less than 2 years. When the degradation rate exceeds the set threshold of 0.15 years / month, the system automatically triggers an emergency warning.

[0113] Based on the forecast results, the technical team developed an equipment replacement and maintenance plan for the substation. Transformer T001 was scheduled for replacement during the next maintenance window, transformers T002 and T003 were included in the mid-term replacement plan, and the remaining transformers were maintained according to their normal maintenance cycles. Through early warning and planned replacement, power outages caused by sudden faults were avoided, ensuring the safe and stable operation of the power grid.

[0114] This invention represents a significant technological advancement over traditional transformer life prediction methods. Traditional methods primarily rely on single methods such as insulating oil chromatography or partial discharge detection, which can only reflect one aspect of the transformer's degradation state and are insufficient to comprehensively assess the overall health level of the equipment. This invention, through multi-dimensional sensor data fusion, can simultaneously monitor multiple degradation mechanisms, including insulation, thermal aging, and mechanical vibration, providing a more comprehensive and accurate health status assessment.

[0115] Traditional data-driven methods lack physical constraints, and prediction results may violate the physical degradation laws of equipment, leading to low prediction accuracy and a lack of interpretability. The mathematical-model fusion framework established in this invention combines the deterministic constraints of the physical model with the learning capabilities of the data-driven model. This ensures that the prediction results conform to physical laws while capturing complex nonlinear degradation characteristics, significantly improving prediction accuracy and reliability.

[0116] Traditional neural networks have a fixed structure, making it difficult to adapt to the characteristic differences of different types of transformers. The improved adaptive multi-scale network used in this invention can adaptively adjust the network structure according to the characteristics of the input data. Through branching and pruning mechanisms, it achieves automatic feature extraction and selection, improving the model's adaptability to different operating conditions and different types of equipment.

[0117] Traditional methods cannot quantify the uncertainty of prediction results, making it difficult to provide a credibility assessment for decision-making. This invention establishes a complete uncertainty quantification framework through Bayesian neural networks and Monte Carlo random deactivation techniques, providing a confidence interval for each prediction result, enabling maintenance personnel to formulate corresponding maintenance strategies based on the prediction credibility.

[0118] Traditional phase transition detection methods are mainly based on statistical analysis, which makes it difficult to accurately identify critical transition points in equipment conditions. The topological phase transition algorithm introduced in this invention is based on condensed matter physics theory. By constructing topological invariants and order parameters, it can accurately identify key transition points in the aging process, enabling early warning of fault conditions and providing a scientific basis for preventive maintenance.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A digital-analog fusion-based power transformer residual life prediction method, characterized in that, The method comprises the following steps: Collecting the body temperature, ambient temperature, load current effective value, insulation oil breakdown voltage, partial discharge pulse amplitude, vibration acceleration amplitude, hydrogen concentration and carbon monoxide concentration in the operation process of the power transformer to establish a multi-dimensional sensor data time sequence; preprocessing the multi-dimensional sensor data time sequence to establish a normalized data matrix; constructing a physical equation according to the normalized data matrix to establish a deterministic physical model; constructing an improved adaptive multi-scale network based on the normalized data matrix to establish a data-driven model; and establishing a topological phase transition algorithm module; Fusing the deterministic physical model and the data-driven model to establish a numerical model fusion prediction framework; when the number of fault samples is less than a preset threshold, using a generative adversarial network for data enhancement; establishing an uncertainty quantification framework to output a residual life prediction value and a confidence interval thereof.

2. The digital-analog fusion based power transformer residual life prediction method of claim 1, wherein, The preprocessing step of the multi-dimensional sensor data time sequence is specifically filtering high-frequency noise through wavelet transform, filling in missing data points by using an interpolation method, and establishing a normalized data matrix.

3. The digital-analog fusion based power transformer residual life prediction method of claim 2, wherein, the multi-dimensional sensor data time series, in particular comprising a body temperature measurement value T ∈ [40, 120] °C, an ambient temperature measurement value ∈ [-20, 50] °C, a load current effective value ∈ [0, 1000] A, an insulating oil breakdown voltage ∈ [30, 80] kV, a partial discharge pulse amplitude ∈ [0, 10000] pC, a vibration acceleration amplitude ∈ [0, 50] , a hydrogen concentration ∈ [0, 1000] ppm, a carbon monoxide concentration ∈ [0, 500] ppm.

4. The digital-analog fusion based power transformer residual life prediction method of claim 3, wherein, The step of constructing a physical equation is specifically constructing insulation aging physical equations, thermal aging physical equations and electrical aging physical equations, and establishing a deterministic physical model based on transformer insulation medium decomposition dynamics, Arrhenius temperature acceleration model and electric field strength degradation model.

5. The digital-analog fusion based power transformer residual life prediction method of claim 4, wherein, The insulation aging physical equation is established based on the molecular chain rupture mechanism of the insulation medium under the action of an electric field, and describes the degradation law of insulation strength over time; the thermal aging physical equation is established based on the Arrhenius temperature acceleration model, and describes the influence of high temperature on the aging rate of insulation material; and the electrical aging physical equation is established based on the inverse relationship between electric field strength and insulation life, and describes the acceleration effect of electric stress on insulation degradation.

6. The digital-analog fusion based power transformer residual life prediction method of claim 5, wherein, The step of constructing an improved adaptive multi-scale network is specifically taking a main dendrite as a starting node, dynamically growing branch nodes according to the intensity of input signals, and realizing adaptive extraction of multi-scale features through synaptic weight adjustment.

7. The digital-analog fusion based power transformer residual life prediction method of claim 6, wherein, The step of establishing a topological phase transition algorithm module is specifically constructing a topological invariant to describe the stable state of the transformer, detecting key transition points in the aging process through order parameter changes, and identifying phase transition critical points.

8. The digital-analog fusion based power transformer residual life prediction method of claim 7, wherein, The step of establishing a numerical model fusion prediction framework is specifically taking the outputs of the insulation aging physical equation, the thermal aging physical equation and the electrical aging physical equation as constraint conditions of the data-driven model.

9. The digital-analog fusion based power transformer residual life prediction method of claim 8, wherein, The step of using a generative adversarial network for data enhancement is specifically generating virtual fault samples to supplement a training data set, and improving the model generalization ability through a transfer learning technique.

10. The digital-analog fusion based power transformer residual life prediction method of claim 9, wherein, The step of establishing an uncertainty quantification framework is specifically capturing cognitive uncertainty by using a Bayesian neural network, and evaluating prediction confidence by using a Monte Carlo random inactivation technique.

Images