Power grid equipment residual life dynamic prediction and calibration method based on multi-source sensing data fusion

Abstract

This invention discloses a method for dynamic prediction and calibration of the remaining lifetime of power grid equipment based on multi-source sensor data fusion, belonging to the field of smart grid equipment condition monitoring technology. The method includes: acquiring multi-source sensor data from power grid equipment to construct a time-series feature vector; using a continuous cohomology algorithm for topology filtering to extract persistent features; calculating the instantaneous transfer entropy to determine the strength of directed causal associations, and synthesizing a dynamic health index accordingly; inputting this index into a physical information neural network to predict the probability distribution of the remaining lifetime; calculating the Wasserstein centroid distance between this distribution and the real-time observed distribution, and performing real-time projection correction on the hidden layer parameters of the network to obtain the calibration result. This invention addresses the problems of multi-source signal noise aliasing, feature fusion distortion, and the lack of underlying physical constraints and dynamic adaptive calibration capabilities in traditional models under complex environments.

Description

Technical Field

[0001] This invention relates to the field of smart grid equipment condition monitoring technology, and more specifically, to a method for dynamic prediction and calibration of the remaining life of power grid equipment based on multi-source sensor data fusion. Background Technology

[0002] Accurately predicting the remaining lifespan of equipment is not only the core basis for formulating condition-based maintenance strategies, but also an important line of defense against cascading power grid failures. Currently, power grid equipment in substations and other scenarios is generally equipped with multi-dimensional condition monitoring sensor systems covering thermal, electrical, chemical, and mechanical aspects, aiming to comprehensively capture multi-physical field degradation information of the equipment. This lays the hardware foundation for a life assessment system based on multi-source data fusion.

[0003] Currently, mainstream life expectancy prediction schemes in the industry are generally divided into two stages: feature fusion and time-series extrapolation. In the feature extraction and fusion stage, existing technologies mostly use principal component analysis or static weight allocation methods based on data synchronous fluctuations to reduce the dimensionality of multi-source sensor data and synthesize it into a one-dimensional health index. In the time-series extrapolation prediction stage, purely data-driven models such as long short-term memory networks are generally used, or integer-order physical degradation models based on the Markov assumption of no aftereffect are used to estimate and calculate the future trajectory of the synthesized health index.

[0004] However, existing technologies have significant limitations under complex real-world conditions: First, due to the lack of in-depth analysis of the topological characteristics of time-series signals and accurate quantification of asymmetric causal relationships, strong non-stationary electromagnetic noise is easily superimposed with weak degradation precursors, and static fusion weights cannot reflect the true driving logic of multi-physics field evolution, resulting in severe distortion of the extracted health index; Second, conventional prediction networks lack physical mechanism characterization of long-term cumulative damage to polymer insulating materials (i.e., aging memory effect), and when faced with small-sample cold starts or long-period extrapolation, the predicted trajectory is prone to diverging from physical laws; Third, existing calibration mechanisms are mostly based on the assumption of an ideal stationary Gaussian environment. When faced with external abrupt changes or non-Gaussian pulse disturbances, they lack in-depth geometric measurement of the difference between the predicted distribution and the actual observed distribution, and cannot achieve real-time adaptive centroid projection correction of model parameters, thus making it difficult to ensure the safe operation of power grid equipment under complex conditions. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a dynamic prediction and calibration method for the remaining lifetime of power grid equipment based on multi-source sensor data fusion. This method extracts the topological features of multi-source signals using a continuous coherence algorithm and instantaneous transfer entropy, synthesizes a causally driven dynamic health index, and then combines it with a physical information neural network for prediction. Furthermore, it uses Wasserstein centroid distance to perform real-time projection correction of network parameters. This addresses the problems of noise aliasing and feature fusion distortion in multi-source sensor signals under complex power grid environments, as well as the inaccurate remaining lifetime prediction caused by the lack of underlying physical mechanism constraints and dynamic adaptive calibration capabilities in traditional models.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for dynamic prediction and calibration of the remaining lifetime of power grid equipment based on multi-source sensor data fusion includes the following steps: acquiring multi-source sensor time-series data of power grid equipment under multiple operating conditions; performing topological data analysis on the multi-source sensor time-series data to extract persistent topological features characterizing the equipment degradation process; calculating the instantaneous transfer entropy between persistent topological features to quantify the nonlinear causal driving relationship between multiple physical quantities; fusing multi-source data based on the causal driving relationship to construct a dynamic health index; predicting the remaining lifetime probability distribution by using a physical information neural network with an embedded material aging physical model through the dynamic health index; calculating the Wasserstein distance between the remaining lifetime probability distribution and the real-time state observation distribution of the equipment, and performing online adaptive calibration of the network parameters of the physical information neural network with the goal of minimizing the Wasserstein distance, and outputting the corrected remaining lifetime prediction result.

[0007] In a preferred embodiment, the topological filtering of multidimensional temporal feature vectors using a persistent cohomology algorithm includes: constructing a simple complex sequence based on multi-source sensor temporal data and performing persistent cohomology analysis as the scale parameter changes; recording the birth and death times of the topological features of each cohomology group to generate a persistence histogram; and filtering and extracting feature points in the persistence histogram whose lifespan exceeds a preset threshold as persistent topological features characterizing precursors to device degradation.

[0008] In a preferred embodiment, calculating the instantaneous transfer entropy of the topological persistent histogram features includes: based on transfer entropy theory, calculating the conditional probability deviation value of the degree of reduction in uncertainty of the future value of the target feature under the condition of known historical information of the source features, and using this deviation value as the instantaneous transfer entropy from the source feature to the target feature to quantify the strength of the directed causal association. In a preferred embodiment, the physical information neural network integrates a Caputo fractional derivative operator; the construction process of the physical information neural network includes: introducing a fractional differential operator into the output layer of the neural network. Where t is the time variable, It is the aging memory coefficient, and The fractional partial differential equations reflecting material degradation are embedded as physical constraint terms into the loss function of the neural network.

[0009] In a preferred embodiment, calculating the Wasserstein centroid distance between the remaining lifetime probability distribution and the real-time state observation distribution includes: constructing an optimal transmission cost function between the remaining lifetime probability distribution and the real-time state observation distribution; and solving the transmission cost function to obtain the Wasserstein centroid distance.

[0010] In a preferred embodiment, real-time projection correction of the hidden layer parameters of the physical information neural network includes: calculating the gradient of the Wasserstein distance between the remaining lifetime probability distribution and the real-time state observation distribution with respect to the hidden layer parameters of the network; and, based on the gradient and a preset observation noise covariance matrix, projecting the predicted distribution output by the network onto the optimal transmission centroid direction of the real-time state observation distribution to update the network parameters.

[0011] In a preferred embodiment, the method further includes: real-time monitoring of the monotonicity rate of change of the dynamic health index; if the monotonicity rate of change exceeds a preset threshold, triggering the real-time projection correction to eliminate the non-Gaussian bias of the degradation trajectory.

[0012] In a preferred embodiment, before inputting the dynamic health index into the physical information neural network, the method further includes: using accelerated aging data from the source domain and real-world operating data from the target domain to calculate the maximum mean difference loss in the latent space; and minimizing this loss through domain adversarial training to reduce the feature distribution difference between the source domain and the target domain.

[0013] In a preferred embodiment, after obtaining the calibrated remaining lifetime prediction result, the method further includes: acquiring measured insulation strength data after the power grid equipment has been overhauled; and dynamically updating the aging memory coefficient in the physical information neural network based on the measured insulation strength data using a Bayesian back-calculation algorithm.

[0014] In a preferred embodiment, after obtaining the calibrated remaining lifetime prediction result, the method further includes: if the failure risk corresponding to the calibrated remaining lifetime prediction result is higher than a preset level, then, based on the load importance of the power grid equipment in the topology, a maintenance scheduling instruction is generated using a deep reinforcement learning algorithm.

[0015] This invention provides a dynamic prediction and calibration system for the remaining lifetime of power grid equipment based on multi-source sensor data fusion, comprising: a data acquisition module for acquiring multi-source sensor time-series data of power grid equipment under multiple operating conditions; a feature extraction module for performing topological data analysis on the multi-source sensor time-series data to extract persistent topological features characterizing the equipment degradation process; a causal modeling module for calculating the instantaneous transfer entropy between persistent topological features and quantifying the nonlinear causal driving relationship between multiple physical quantities; a health index synthesis module for fusing multi-source data based on causal driving relationships to construct a dynamic health index; a probability prediction module for predicting the remaining lifetime probability distribution by using a physical information neural network with an embedded material aging physical model; and a remaining lifetime correction module for calculating the Wasserstein distance between the remaining lifetime probability distribution and the real-time state observation distribution of the equipment, and performing online adaptive calibration of the network parameters of the physical information neural network with the goal of minimizing the Wasserstein distance, and outputting the corrected remaining lifetime prediction result.

[0016] The technical effects and advantages of this invention, which is a dynamic prediction and calibration method for the remaining life of power grid equipment based on multi-source sensor data fusion, are as follows: This invention constructs a multi-dimensional temporal feature vector by acquiring multi-source sensor data and uses a continuous cohomology algorithm for topology filtering, effectively filtering out complex environmental noise and extracting robust topological persistence features. Furthermore, by calculating instantaneous transfer entropy, it accurately quantifies the strength of directed causal relationships in the multi-physics evolution process, thus scientifically synthesizing a high-fidelity dynamic health index. Based on this, the index is input into a physical information neural network, achieving accurate prediction of the remaining lifetime probability distribution that conforms to the underlying degradation mechanism. Finally, it innovatively utilizes the Wasserstein centroid distance metric to measure the underlying geometric difference between the predicted distribution and the real-time observed distribution, and accordingly performs real-time projection correction on the network's hidden layer parameters, completely establishing a closed loop for the model's adaptive evolution with the actual operating state. This significantly improves the accuracy, anti-interference capability, and physical interpretability of dynamic prediction of the remaining lifetime of power grid equipment under complex non-stationary operating conditions. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process for dynamic prediction and calibration of the remaining life of power grid equipment based on multi-source sensor data fusion, provided in an embodiment of the present invention.

[0018] Figure 2 A certain accelerated aging test provided in an embodiment of the present invention has the following scale parameters. A schematic diagram of the topological persistence scatter distribution simulation under evolution.

[0019] Figure 3This is a schematic diagram illustrating the real-time monitoring of dynamic health index and the simulation of monotonic change rate trend provided in an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of the comparison curves for predicting the remaining life of a transformer, provided in an embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram of a dynamic prediction and calibration system for the remaining life of power grid equipment based on multi-source sensor data fusion, provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1, Figure 1 This invention presents a method for dynamic prediction and calibration of the remaining life of power grid equipment based on multi-source sensor data fusion, comprising the following steps: S1, acquire multi-source sensor time-series data of power grid equipment under multiple operating conditions.

[0024] In this embodiment, taking a power transformer as the target power grid equipment as an example, the multi-source sensing data covers heterogeneous monitoring signals reflecting the thermodynamic, chemical, and mechanical states of the equipment. The system acquires the temporal changes in winding hot spot temperature in real time using a fiber optic grating temperature sensor array (e.g., an FBG sensor with a measurement accuracy of up to 0.1 degrees Celsius) pre-deployed inside the transformer. Simultaneously, it continuously acquires the concentration sequence of dissolved gases in the oil using an online transformer oil chromatography monitoring device, focusing on extracting the concentrations of hydrogen (H2) and acetylene (C2H2) characteristic gases that characterize partial discharge and arc faults. Furthermore, it uses a broadband piezoelectric accelerometer installed on the transformer tank surface to collect high-frequency mechanical vibration signals characterizing mechanical loosening of the core and windings. To address the time-scale mismatch between high-frequency vibration signals and other slowly changing state signals, the system first extracts features from the high-frequency vibration signals, such as calculating the effective value RMS, peak factor, or specific frequency band energy within each 10-minute window, thereby converting it into a low-frequency degradation feature sequence consistent with the temperature and gas change rhythm.

[0025] To address the issue of diverse physical dimensions and inconsistent sampling rates in the collected multi-source heterogeneous sensor data, the system employs resampling technology to unify the temporal resolution. A specific engineering implementation example is as follows: A unified temporal resolution of 1 hour is set. The vibration characteristics and temperature data after frequency reduction are downsampled and averaged, while spline interpolation is performed on the low-sampling-rate oil chromatography gas data. Subsequently, the Z-score method is used to eliminate dimensional differences between features. Next, a sliding time window of length L (e.g., L=24 steps, representing a 24-hour time span) is used to truncate and flatten the feature matrix. Finally, the multi-source time series sequences are concatenated along the feature dimensions to construct a high-dimensional continuous temporal feature matrix for a single sample, i.e., the multi-dimensional temporal feature vector.

[0026] This step effectively eliminates the differences in dimensions and sampling rates of thermodynamic, chemical, and mechanical vibration data by performing targeted time-frequency domain feature extraction, resampling, and sliding window construction on multi-source heterogeneous data. It truly restores the data preprocessing logic of the industrial site and provides high-quality standardized data input for subsequent extraction of topological features using the continuous homology algorithm.

[0027] S2 performs topological data analysis on multi-source sensor time-series data to extract persistent topological features that characterize the device degradation process.

[0028] In this embodiment, the system constructs a simple complex sequence based on multi-source sensor temporal data and performs continuous homology analysis as the scale parameter changes. In actual computation, the system treats each multi-dimensional temporal feature vector in the high-dimensional state space as an independent vertex in the point cloud data and employs the Vietoris-Rips complex construction method. The system then uses a preset scale parameter (denoted as the spatial connectivity radius) to perform continuous homology analysis. It expands continuously, uniformly, and monotonically from zero, when the Euclidean distance between any two eigenvector vertices is less than or equal to the current scale parameter. When the distance between any two vertices is less than or equal to 1, an edge is drawn between them; when the distance between any two pairs of any n+1 vertices is less than or equal to 1, an edge is drawn between them. When this happens, an n-dimensional simplex is formed. Considering the computational complexity and memory consumption of the algorithm, this embodiment limits the maximum cohomology dimension to be calculated to . That is, mainly extracting the zero-dimensional connected components ( ), one-dimensional holes ( ) and two-dimensional voids ( Topological features of ). Scale parameters The physical significance of the gradual expansion lies in the fact that it simulates the evolution of the data observation perspective from microscopic local isolated points to macroscopic global connected structures, so that discrete time-series feature data are gradually connected into simple complex sequences with rich topological structures such as connected branches, one-dimensional holes and high-dimensional holes.

[0029] Subsequently, the system records the birth and death times of the topological features of homology groups in each dimension, generating persistent histograms. During the topological filtering evolution of simple complex sequences, when the scale parameter... When the scale parameter increases to a certain critical value, a new topological feature is formed for the first time. The corresponding scale parameter value at this point is defined as the birth time of this topological feature on the topological graph; as... As the scale increases further, these topological features will eventually be filled with higher-dimensional simplexes, resulting in closure and vanishing. The scale parameter value corresponding to this vanishing is defined as the death moment of the topological feature on the topological graph. The formula for calculating the durability lifetime is as follows: (1) In the formula, This represents the persistent lifetime of the topological features of the nth-dimensional homology group; This represents the scale parameter corresponding to the time when the topological feature of the nth homology group is filled and disappears during the topological filtering process, i.e., the death time. This represents the scale parameter corresponding to the first formation of the topological feature of the nth homology group during the topological filtering process, i.e., the birth time.

[0030] Furthermore, the system filters and extracts feature points in the persistence graph whose lifetime exceeds a preset threshold as persistent topological features characterizing precursors to device degradation. This includes: calculating the median and absolute deviation of the median of the persistent lifetime set of the target homology group, and constructing an adaptive preset threshold based on the median and absolute deviation; filtering out feature points whose persistence lifetime is lower than the adaptive preset threshold to retain topological feature points that reflect the evolution of the device's true health status; and converting the retained topological feature points into continuous Betti number curves in a time sliding window order as input feature sequences for calculating instantaneous transfer entropy.

[0031] Specifically, in the actual operation and monitoring environment of power grid equipment, random fluctuation noise such as high-frequency electromagnetic interference and sensor transient sampling errors typically exhibits topological characteristics with extremely short lifecycles. Conversely, nonlinear dynamic evolution processes reflecting real physical degradation such as equipment insulation deterioration and increased internal partial discharge will manifest as stable voids or loop structures with long lifecycles in the topological space. Therefore, the system introduces a median absolute deviation (MAD) algorithm based on the statistical distribution of the persistence lifetime of global topological features to set an adaptive preset threshold. Specifically, let Median be the median of the persistence lifetime set of a certain homology group, and MAD be the median absolute deviation. Then, the formula for calculating the adaptive preset threshold T is: (2) The empirical coefficient k typically ranges from 2 to 3. Using this threshold, the system filters out short-lived topological points with a lifespan below the preset threshold, thus accurately preserving topological feature points that reflect the true evolution of the device's health status. Furthermore, the system transforms the preserved long-lifetime topological feature points into continuous curves of the Betti number over time, following the original time sliding window order. This maps the disordered topological scatter points back to a low-dimensional topological feature time series consistent with the original sampling rhythm, serving as the input feature sequence for subsequent calculation of instantaneous transfer entropy.

[0032] Specifically, for each time sliding window The system extracts and retains the first... A set of dimensional topological features. For a given scale parameter Betti curve The value is strictly defined as the total number of features that still survive at this scale, and its transformation calculation formula is: (3) In the formula, Indicates within the time window The first one retained by threshold filtering The total number of topological feature points in dimensionality; As an indicator function, when the evaluation scale Located in the Lifecycle range of each feature When the time frame is met, the indicator function is set to 1; otherwise, it is set to 0. Subsequently, the system obtains the sliding window by calculating the Betti values ​​of the one-dimensional and two-dimensional homology groups at specific feature scales, or by directly integrating the Betti curve across the entire scale space to obtain the area under the curve. The scalar eigenvalues ​​are then mapped back to low-dimensional feature time series that are completely consistent with the original sensor sampling beat as the time window continues to slide.

[0033] To visually verify the effectiveness of this topology filtering mechanism, Figure 2 This demonstrates the scale parameters of an accelerated aging test. Simulation diagram of topological persistence scatter distribution under evolution. (e.g.) Figure 2 As shown in the figure, the horizontal axis represents the birth time of the simple complex extension, and the vertical axis represents the death time. A large number of these are densely distributed near the diagonal (i.e., lifespan). The extremely short scattering points correspond to environmental noise such as injected broadband electromagnetic interference; while the star-shaped feature points that are significantly off the diagonal (whose durability exceeds the adaptive threshold of 0.45) precisely correspond to the actual microscopic pore evolution of internal insulation degradation. Figure 2Simulation results demonstrate that the proposed method can clearly extract subtle degradation precursors hidden under strong non-stationary noise in the topological space.

[0034] This step maps multi-source heterogeneous time-series signals to a high-dimensional topological space and utilizes the persistence characteristics of topological invariants to achieve the physical separation of non-stationary strong noise from weak precursors of equipment degradation. This significantly improves the robustness of feature extraction and provides standardized time-series feature inputs for subsequent causal network construction.

[0035] S3 calculates the instantaneous transfer entropy between persistent topological features, quantifying the nonlinear causal driving relationship between multiple physical quantities.

[0036] In this embodiment, the calculation process includes: based on the transfer entropy theory, calculating the conditional probability deviation value of the degree of reduction in uncertainty of the future value of the target feature under the condition of known historical information of the source feature; using this deviation value as the instantaneous transfer entropy from the source feature to the target feature to quantify the strength of the directed causal association. Specifically, taking the physical process of "sudden change in winding temperature" causing "change in gas production rate in oil" inside a transformer as an example, the traditional Pearson correlation coefficient can only measure the degree of linear synchronous fluctuation between the two on the time axis, and cannot characterize the strict sequential driving mechanism of thermal anomalies leading to subsequent chemical insulation degradation. Therefore, the system uses instantaneous transfer entropy to quantify this asymmetric information flow.

[0037] It should be noted that, in order to preserve the independence of the physical subsystems, when performing the topology filtering described in the preceding steps, the system extracts persistent topological graphs for both the source terminal system (such as the thermodynamic monitoring matrix) and the target terminal system (such as the chemical monitoring matrix), and transforms them into corresponding one-dimensional continuous topological feature sequences (i.e., as described below). and Based on this, the system first extracts the source feature historical state sequence characterizing temperature change and the target feature historical state sequence characterizing gas production rate change, and constructs their joint probability distribution. Since the extracted topological state values ​​are continuous floating-point sequences, the system employs the kernel density estimation (KDE) method or a k-nearest neighbor-based method. The KSG estimation algorithm approximates the probability density of a continuous sequence to calculate the discretized joint probability and marginal probability distribution. Then, by calculating the marginal probability distribution, it assesses the extent to which introducing historical information from source features can reduce the uncertainty in predicting the future state of the target feature, given the known historical evolution trajectory of the target feature. This degree of uncertainty reduction is reflected in the conditional probability bias value. Subsequently, the system determines the strength of the directed causal association based on the conditional probability bias value. The instantaneous transfer entropy is calculated as follows: (4) In the formula, Indicates source features To target features The instantaneous transfer entropy, i.e., the quantified strength of the directed causal relationship; This represents the topological state value of the source feature at time t, which is the integral value of the Betti curve feature of the source subsystem at time window t in the aforementioned steps. This represents the topological state value of the target feature at time t, which is the integral value of the Betti curve feature of the target subsystem at time window t. Indicates the target feature after a time interval The future evolution state value, where the time interval is... The setting needs to be based on the physical hysteresis time of the thermo-chemical coupling reaction of the power grid equipment. In this embodiment, Preset The sampling step size corresponding to each hour; This represents the joint probability distribution among the current state of the source feature, the current state of the target feature, and its future state. It represents the conditional probability of the target feature evolving into a future state, given the current states of the target feature and the source feature; This represents the marginal conditional probability that a target feature will evolve into its future state only under the marginal condition that its current state is known. The logarithmic term of the ratio of the two conditional probabilities constitutes the conditional probability deviation value.

[0038] To further quantify and verify the ability of information transmission entropy to characterize asymmetric physical causality, this embodiment extracts three key monitoring features before and after the occurrence of partial discharge faults. The calculation results of their conditional probability deviation and directed causal correlation strength are shown in Table 1.

[0039] Table 1

[0040] The statistical data in Table 1 clearly shows that "localized overheating of the winding" points to "oil instability". The instantaneous transfer entropy of the "gas production rate" is as high as 0.872, while its reverse information transfer entropy is only 0.105. This numerical difference of up to 8 times strictly conforms to the unidirectional physical causal law of thermal stress-driven chemical reactions.

[0041] This step eliminates the interference of spurious correlation features between multi-source sensor signals by quantifying the asymmetric causal driving relationship in the multi-physics evolution process, laying a reliable physical and logical foundation for the subsequent extraction of the real device degradation trajectory.

[0042] S4, based on causal relationships, integrates multi-source data to construct a dynamic health index.

[0043] In this embodiment, step S4 includes: calculating the sum of all instantaneous transfer entropies of all inputs when each topological feature is used as a target feature, to obtain the dynamic causal weight of each topological feature; setting a health baseline value and a failure threshold value, using a negative exponential mapping function to unify the polarity of the original observation values ​​of each topological feature, and constraining them to a preset health state range; and weighting and summing the state values ​​of each feature with their corresponding dynamic causal weights to obtain the dynamic health index.

[0044] Specifically, the system uses the strength of the directed causal relationship between the features calculated above (i.e., instantaneous transfer entropy) as a dynamic weighting coefficient to perform real-time weighted summation of the extracted and retained topological features. Through this fusion method based on causal-driven logic, the multidimensional nonlinear degradation features representing different physical fields (such as heat, electricity, and mechanics) can be mapped and transformed into a continuous health index with values ​​strictly between 0 and 1. The dynamic health index is represented by a curve, where 1 indicates that the power grid equipment is in a completely healthy state, and 0 indicates that the equipment performance has completely degraded to the point of functional failure. The formula for calculating the dynamic health index is as follows: (5) In the formula, represents the dynamic health index synthesized at time t; K represents the total number of pre-deterioration features retained through topological filtering; This represents the state value of the k-th topological feature at time t after polarity unification and normalization. This represents the dynamic causal weight of the k-th topological feature calculated based on the directed causal association strength at time t, and the sum of the weights of all features satisfies .

[0045] In this embodiment, the system uses the sum of the instantaneous propagation entropies of all inputs when the k-th feature is used as the target feature as its basic weight factor, specifically calculated as follows: (6) Formula (5) is used to assign higher weights to key degenerative features that evolve under the coupling of multiple physical factors. The system sets health baseline values ​​for acquisition. With failure threshold If the characteristics increase with aging (e.g., gas production rate or partial discharge amount), then polarity unification and normalization are performed using a negative exponential mapping function, the formula of which is as follows: (7) in, This is the shape adjustment coefficient. The original observation values ​​of the topological features at time t are used to ensure that each input feature is strictly constrained within the healthy state interval of [0,1].

[0046] Furthermore, the system monitors the monotonic rate of change of the dynamic health index in real time; if the monotonic rate of change exceeds a preset threshold, the real-time projection correction is triggered to eliminate the non-Gaussian deviation of the degradation trajectory. From the physical logic analysis of material fatigue and insulation aging, in the normal and gradual physical aging process experienced by power grid equipment, its health index curve should exhibit a slow and monotonically decreasing trend. However, when the system detects a violent non-monotonic oscillation in the health index, it usually indicates that the power grid equipment has encountered a sudden shock from adverse external conditions (such as electrodynamic shock from short-circuit current or lightning overvoltage damage), or that the underlying sensing and monitoring system has been subjected to strong non-Gaussian pulse electromagnetic interference. Therefore, the system calculates the first-order time derivative of the health index within adjacent time windows. Due to the strict decrease in HI during normal aging (i.e., Once detected (Manifested as an abnormally healthy rebound and oscillation that does not follow physical laws), or When the negative absolute value of the signal rises sharply and exceeds the preset lower limit threshold based on historical stable period statistics (manifested as a non-Gaussian pulse impact), the system will determine that the current degradation trajectory deviates from the steady-state Gaussian distribution assumption, and then automatically generate a trigger signal and hand it over to the subsequent calibration module to start the real-time parameter projection correction mechanism.

[0047] like Figure 3 The simulated curves of the dynamic health index real-time monitoring and monotonic change rate trend show that, within the normal aging range of 0 to 500 hours, the synthesized HI curve exhibits a smooth, monotonically decreasing trend, with its change rate stabilizing around -0.002. At the 520th hour, due to the simulated strong electromagnetic non-Gaussian pulse disturbance caused by the injection of external lightning overvoltage, the HI curve oscillated violently, and the monotonic change rate instantly surged to 0.085, successfully triggering the set safety warning threshold (0.05). These simulation results intuitively demonstrate the system's accurate capture and adaptive triggering capabilities for unsteady, sudden changes in operating conditions.

[0048] This step, through the scientific synthesis of dynamic causal weights and real-time monitoring of the monotonicity of physical evolution, provides a comprehensive characterization of the actual degradation state of equipment under complex working conditions with adaptive physical sensitivity, effectively preventing misjudgment of aging state caused by sudden pulse interference.

[0049] S5 uses a physical information neural network embedded with a material aging physical model to predict the probability distribution of remaining lifespan based on the dynamic health index.

[0050] In this embodiment, before inputting the dynamic health index into the physical information neural network, the method further includes: calculating the maximum mean difference loss in the latent space using accelerated aging data from the source domain and actual operating data from the target domain; minimizing this loss through domain adversarial training to reduce the feature distribution difference between the source and target domains. To address the problem of missing fault samples in the initial stage of newly commissioned equipment, this embodiment introduces a domain adaptation mechanism. In specific operation, the system uses accelerated aging test data obtained under extreme conditions such as increased thermal and electrical stress in a laboratory environment, along with their corresponding complete lifespan labels, as the source domain. Simultaneously, it uses actual operating data collected during substation operation that lacks failure labels as the target domain. The system utilizes a feature extraction network to simultaneously map the multidimensional temporal features of the source and target domains into a translation-invariant regenerative kernel Hilbert space (RKHS), constructing a maximum mean difference (MMD) loss function to quantify the statistical distance between the feature distributions of the two domains. Furthermore, through a minimax game between the feature extractor and the domain discriminator in adversarial training, the feature edge distributions of both domains are aligned. The formula for calculating the maximum mean difference loss is as follows: (8) In the formula, This represents the maximum mean difference loss between the accelerated aging data in the source domain and the actual operational data in the target domain within the latent space. and These represent the total number of samples in the source domain dataset and the target domain dataset, respectively. and Let i and j represent the i-th source domain sample and the j-th target domain sample, respectively. This represents a high-dimensional nonlinear kernel mapping function that maps the original input features to the reproducing kernel Hilbert space; This represents the norm distance calculated in the regenerated kernel Hilbert space.

[0051] After completing domain feature alignment, the system constructs a deep prediction model, inputting the dynamic health index into a physical information neural network to predict the probability distribution of remaining lifespan. In terms of network structure design, the output layer of the physical information neural network contains two parallel branches, each predicting the mean of the Gaussian distribution followed by the remaining lifespan. with standard deviation RUL stands for Remaining Useful Life. This represents the expected remaining lifespan predicted at time t. This represents the standard deviation of the uncertainty of the predicted value. At this point, the data-driven loss term... The network is optimized using negative log-likelihood loss (NLLLoss), which enables the network to not only output a point estimate of the remaining lifetime, but also extrapolate the remaining lifetime probability distribution that includes the uncertainty boundary.

[0052] It should be noted that the physical information neural network integrates a Caputo fractional derivative operator; the construction process of the physical information neural network includes: introducing a fractional differential operator into the output layer of the neural network. ,in It is the aging memory coefficient, and The fractional partial differential equations reflecting material degradation are embedded as physical constraints into the loss function of the neural network. Specifically, this embodiment uses fractional relaxation kinetics equations to describe the insulation degradation process, and its physical governing equations are expressed as follows: (9) in The degradation rate constant is The steady-state recovery coefficient is... This indicates that the starting point is 0 and the order is defined using Caputo. The fractional derivative operator, i.e. Therefore, the physical constraint loss term is defined as the mean square residual of the equation at each placement point: (10) In the formula, M is the total number of configuration points, that is, the number of time nodes involved in the physical loss calculation; This represents the m-th specific time point used to calculate the physical residual; It represents the square of the L2 norm, used to quantify the degree of deviation between the left and right sides of a physical equation.

[0053] At the underlying physical logic level, conventional integer-order differential degradation models are typically based on the Markov assumption of no aftereffect, focusing only on the instantaneous rate of change of the current state. They cannot effectively remember and accumulate the alternating electrothermal stress fatigue effects experienced by equipment under long-term complex operating conditions. Conversely, Caputo fractional calculus, due to its global integration characteristics, can perfectly describe the complex sub-diffusional nonlinear degradation process of polymeric insulating materials such as transformer oil-paper insulation. Among these, the aging memory coefficient... Physically, this characterizes the dependence of material performance degradation trajectories on historical cumulative damage; a smaller value indicates a more significant long-term memory effect. During network training, the system combines the data-driven loss term derived from actual multi-source sensor observation data (i.e., the aforementioned NLLLoss) with the fractional partial differential equation residuals derived from Arrhenius lifetime degradation theory (i.e., the aforementioned physical constraint term). The data-driven and physical constraint terms are combined to form a composite loss function. This mechanism allows the data-driven and physical constraint terms to jointly guide the gradient descent of the model parameters, thereby limiting the solution space of the neural network and ensuring that the final output remaining lifetime probability distribution strictly conforms to the objective physical laws of insulation aging. The total loss function and fractional-order operator expressions of the neural network are as follows: (11) (12) In the formula, This represents the total loss function of a physical information neural network during the training process; This represents the negative log-likelihood data-driven loss term used to optimize the mean and standard deviation of the remaining lifetime probability distribution; This represents the physical constraint loss term, which consists of the residuals of the fractional-order relaxation dynamics equations reflecting material degradation. An adaptive penalty weighting factor to balance data-driven and physical constraints; The standard gamma function; For the dynamic health index and its historical integral time variable The first ordinary derivative; t is the current prediction time point.

[0054] To verify the superiority of physical information neural networks in long-period extrapolation prediction by incorporating the Caputo fractional derivative operator. Figure 4 A comparative curve of remaining lifetime (RUL) predictions using a conventional pure data-driven LSTM network and the PINN model of this invention on the same target domain transformer dataset is presented. Figure 4 As shown, in the later stages of the test (i.e., the section where the actual lifespan of the device exceeds 80%), the conventional LSTM model, lacking the physical memory of long-term cumulative electrothermal damage, has a predicted mean that deviates significantly from the actual degradation trajectory, with a root mean square error (RMSE) as high as 12.4%. In contrast, the PINN model, which integrates fractional-order aging memory physical constraints, consistently has a 95% confidence interval for its predicted probability distribution that closely encloses the actual lifespan evolution curve, with the RMSE dropping significantly to 3.1%. Figure 4 The comparative experiments fully demonstrate that the model architecture designed in this step not only improves the prediction accuracy, but also ensures the physical rationality of the extrapolation results.

[0055] S6. Calculate the Wasserstein distance between the remaining lifetime probability distribution and the real-time status observation distribution of the equipment, and perform online adaptive calibration of the network parameters of the physical information neural network with the goal of minimizing the Wasserstein distance, and output the corrected remaining lifetime prediction result.

[0056] In this embodiment, the specific calculation process is as follows: The system first defines the optimal transmission cost function between the remaining lifetime probability distribution and the real-time state observation distribution; then, it solves for the joint distribution scheme that minimizes the total transmission cost, obtaining the Wasserstein centroid distance. From the perspective of underlying metric geophysics, the gradient of the traditional Kullback-Leibler (KL) divergence tends to zero or infinity when the predicted distribution and the actual observed distribution do not overlap at all (i.e., a sudden device state shift occurs), thus preventing network parameters from being updated. However, the Wasserstein distance, based on optimal transmission theory, treats the probability distribution as a mass stack in the topological space, and its physical meaning is the minimum geometric work required to transport the predicted state mass to the actual observed state mass. Therefore, even when the two distribution support sets have no intersection, the Wasserstein distance can still provide smooth and effective directional gradient guidance. The formula for calculating the Wasserstein centroid distance is as follows: (13) In the formula, This represents the second-order Wasserstein centroid distance between the remaining lifetime probability distribution and the real-time state observation distribution; This represents the remaining lifetime probability distribution output by the physical information neural network; This represents the distribution of real-time state observations inferred from the latest sensor data; Indicates and It is the set of all possible joint distribution schemes of the absolute marginal distribution; This represents a specific joint distribution scheme in the set; x and y are the prediction state spaces, respectively. and observation state space The state random variable in; The squared Euclidean distance, defined in Euclidean space, is the optimal transmission cost function, which is the distance from state x to state y. This represents the extremum operator that solves for the expected total transmission cost under this joint distribution scheme, achieving the infimum. In actual computation, the real-time state observation distribution... The scalar health index is extracted and integrated at the current moment. An empirical Gaussian distribution is constructed with mean as the mean and variance as the pre-defined joint noise covariance of the sensors. To achieve end-to-end backpropagation differentiation in the neural network, the system calculates... An entropy regularization term is introduced for distance calculation, and the Sinkhorn-Knopp iterative algorithm is used to approximate the infimum of the optimal transmission cost under discrete empirical distribution, thereby obtaining a smooth and differentiable loss gradient.

[0057] After obtaining the geometric distance metric between the aforementioned high-dimensional distributions, the system performs dynamic correction of network parameters: calculating the gradient of the Wasserstein distance between the remaining lifetime probability distribution and the real-time state observation distribution with respect to the hidden layer parameters; based on the gradient and a preset observation noise covariance matrix, projecting the predicted distribution output by the network onto the optimal propagation centroid direction of the real-time state observation distribution, and updating the network parameters. This process draws on the physical idea of ​​state gain update in Kalman filtering. The system uses the backpropagation algorithm to obtain the Jacobian gradient of the Wasserstein distance with respect to the current network weights, and introduces the preset sensor observation noise covariance matrix (reflecting the uncertainty and reliability of the observation data) as a regularization amplitude modulation coefficient matrix into the gradient descent step size. When the observation noise is small, the projection step size increases, prompting the hidden layer weights of the neural network to update rapidly along the direction of partial derivatives, thereby forcing the predicted distribution at the next moment to accelerate manifold projection towards the centroid of high-quality real-time observation data, and finally outputting a calibrated expected remaining lifetime and confidence interval with high fidelity. The calculation formula for the real-time projection correction of the hidden layer parameters is as follows: (14) In the formula, This represents the hidden layer weight parameter matrix of the neural network, updated at time t after real-time projection correction; This represents the hidden layer weight parameter matrix of the network used in the previous prediction at time t-1; This represents the adaptive gain coefficient matrix calculated based on a preset observation noise covariance matrix; This represents the Wasserstein barycenter distance versus the hidden layer weight parameter matrix at the previous time step. The partial derivatives (i.e., the gradient vector). Specifically, the adaptive gain coefficient matrix. It can be dynamically calculated from the observation noise covariance matrix R and the prediction state error covariance matrix P, for example. This ensures that the projection correction step size is automatically reduced when the observation data is degraded by strong electromagnetic interference, thus preventing the network parameters from overfitting to noise.

[0058] In practical power engineering applications, after obtaining the calibrated remaining life prediction result, the method further includes: acquiring the measured insulation strength data of the power grid equipment after maintenance; and based on the measured insulation strength data, using a Bayesian back-calculation algorithm to calculate the aging memory coefficient in the physical information neural network. Dynamic updates are performed. After planned physical overhauls or offline high-voltage tests, maintenance personnel can extract absolute physical benchmark indicators reflecting the true aging of materials (such as high-precision polymerization degree DP values ​​or insulation resistance measurements like absorption ratio / polarization index of transformer insulation paper). The system uses these low-frequency but extremely high-precision offline hard indicators as the true observation basis for the likelihood function. Employing Bayesian inference frameworks such as Markov Chain Monte Carlo (MCMC), it updates the prior-given aging memory coefficients to a posterior probability distribution containing the true degradation characteristics of individual equipment. This empowers the physical information neural network to adaptively evolve parameters from a "general model" to a "one-machine-one-policy" approach as the equipment's lifespan extends. The Bayesian back-inference update formula is as follows: (15) In the formula, This represents the aging memory coefficient updated after obtaining actual overhaul data. The posterior probability distribution; This refers to the measured insulation strength data obtained after the power grid equipment has been repaired; This represents the likelihood function of the current measured insulation data observed under a given aging memory coefficient; This represents the prior probability distribution of the aging memory coefficient before this update; the integral term in the denominator is the marginal likelihood constant to ensure the normalization of the posterior probability. In calculating the likelihood function... At that time, the system pre-establishes a system that reflects the health index. An empirical transformation equation (e.g., analytical aging model) maps the network to absolute physical benchmarks (such as the cohesion DP value). This equation represents the network's current state. The predicted state of the output is transformed into a response to... The expected value of the observation is obtained, and then the probability density of the measured data under this expectation is calculated, so as to realize the reverse inference of cross-modal parameters.

[0059] Ultimately, to substantially transform lifetime prediction data into operational productivity, after obtaining the calibrated remaining lifetime prediction result, the system further includes: if the failure risk corresponding to the calibrated remaining lifetime prediction result is higher than a preset level, then, in conjunction with the load importance of the power grid equipment in the topology, a maintenance scheduling instruction is generated using a deep reinforcement learning algorithm. In specific implementation, the system continuously assesses the probability that the lower limit of the remaining lifetime probability distribution will reach the failure threshold (i.e., failure risk). Once the preset safety warning line is exceeded, the system will trigger the scheduling decision module. This module constructs a Markov decision process aimed at maximizing long-term economic benefits. Its state space includes the calibrated remaining lifetime distribution and the node betweenness centrality (i.e., load importance) of the target equipment in the complex power grid network graph; its reward function is the penalty for cascading power outages caused by unplanned power outages of the equipment, and the fixed asset depreciation cost caused by premature equipment reset. Through long-term offline training and online optimization using reinforcement learning algorithms such as Deep Q-Network (DQN) or Proximal Policy Optimization (PPO), the agent ultimately outputs a specific action strategy with optimal spatiotemporal benefits.

[0060] This step achieves closed-loop self-healing calibration of the predicted distribution to the real-time observed manifold through optimal transport metric, combines Bayesian backpropagation to individualize the fractional-order physical parameters, and utilizes reinforcement learning to optimize operating parameters under multi-objective constraints. This mechanism technically eliminates the accumulation of non-Gaussian bias in long-term degradation prediction, significantly improves the dynamic fidelity of equipment remaining life prediction, and effectively slows down the nonlinear degradation rate of power insulation materials by actively intervening in technical parameters, ensuring the operational boundary safety of power grid equipment under complex operating conditions.

[0061] Example 2, Figure 5 A dynamic prediction and calibration system for the remaining lifetime of power grid equipment based on multi-source sensor data fusion is presented, including: The data acquisition module is used to acquire multi-source sensor time-series data of power grid equipment under various operating conditions; The feature extraction module is used to perform topological data analysis on multi-source sensor time-series data and extract persistent topological features that characterize the device degradation process. The causal modeling module is used to calculate the instantaneous transfer entropy between persistent topological features and quantify the nonlinear causal driving relationship between multiple physical quantities. The health index synthesis module is used to construct a dynamic health index by fusing multi-source data based on causal driving relationships. The probability prediction module is used to predict the probability distribution of remaining lifespan by using the physical information neural network of the dynamic health index embedded in the material aging physical model. The remaining lifetime correction module is used to calculate the Wasserstein distance between the remaining lifetime probability distribution and the real-time status observation distribution of the equipment, and to perform online adaptive calibration of the network parameters of the physical information neural network with the goal of minimizing the Wasserstein distance, and output the corrected remaining lifetime prediction result.

[0062] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0063] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0064] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0065] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0067] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power grid equipment residual life dynamic prediction and calibration method based on multi-source sensor data fusion, characterized in that, Includes the following steps: Acquire multi-source sensor time-series data of power grid equipment under various operating conditions; Topological data analysis is performed on multi-source sensor time-series data to extract persistent topological features that characterize the device degradation process; Calculate the instantaneous transfer entropy between persistent topological features, quantify the nonlinear causal driving relationship between multiple physical quantities, and construct a dynamic health index based on the causal driving relationship by fusing multi-source data. The dynamic health index is used to predict the probability distribution of remaining lifespan through a physical information neural network with an embedded material aging physical model. The Wasserstein distance between the remaining lifetime probability distribution and the real-time status observation distribution of the equipment is calculated, and the network parameters of the physical information neural network are adaptively calibrated online with the goal of minimizing the Wasserstein distance, and the corrected remaining lifetime prediction result is output.

2. The method for dynamic prediction and calibration of remaining life of power grid equipment based on multi-source sensor data fusion according to claim 1, characterized in that, The step of performing topological data analysis on multi-source sensor time-series data to extract persistent topological features characterizing the device degradation process includes: Simple complex sequences are constructed based on multi-source sensor time-series data, and continuous homology analysis is performed as the scale parameter changes. Record the birth and death times of the topological features of homology groups in each dimension, and generate a persistent histogram; Feature points in the persistence graph whose lifecycle exceeds a preset threshold are selected and extracted as persistent topological features that characterize precursors to device degradation.

3. The method for dynamic prediction and calibration of remaining life of power grid equipment based on multi-source sensor data fusion according to claim 1, characterized in that, The calculation of the instantaneous transfer entropy between persistent topological features includes: Based on the theory of transfer entropy, the conditional probability deviation value of the degree of reduction in uncertainty of the future value of the target feature is calculated under the condition of known historical information of the source feature. This deviation value is used as the instantaneous transfer entropy from the source feature to the target feature to quantify the strength of the directed causal relationship.

4. The method for dynamic prediction and calibration of remaining life of power grid equipment based on multi-source sensor data fusion according to claim 1, characterized in that, The physical information neural network integrates a Caputo fractional derivative operator; the construction process of the physical information neural network includes: Introducing fractional differential operators into the output layer of neural networks Where t is the time variable, It is the aging memory coefficient, and ; The fractional partial differential equations reflecting material degradation are embedded as physical constraints in the loss function of the neural network.

5. The method for dynamic prediction and calibration of remaining life of power grid equipment based on multi-source sensor data fusion according to claim 2, characterized in that, The process of filtering and extracting feature points in the persistence graph whose lifecycle exceeds a preset threshold, as persistent topological features characterizing precursors to device degradation, includes: Calculate the median and absolute deviation of the median of the target homology group persistence lifetime set, and construct an adaptive preset threshold based on the median and absolute deviation of the median; Feature points with a durability lifespan lower than the adaptive preset threshold are filtered out to retain topological feature points that reflect the evolution of the device's true health status. The retained topological feature points are transformed into continuous Betti number curves in a time sliding window order, which serve as the input feature sequence for calculating the instantaneous transfer entropy.

6. The method for dynamic prediction and calibration of remaining lifespan of power grid equipment based on multi-source sensor data fusion according to claim 1, characterized in that, Online adaptive calibration of the network parameters of the physical information neural network includes: Calculate the gradient of the Wasserstein distance between the remaining lifetime probability distribution and the real-time state observation distribution with respect to the hidden layer parameters of the network; Based on the gradient and the preset observation noise covariance matrix, the predicted distribution output by the network is projected onto the optimal transmission centroid direction of the real-time state observation distribution, and the network parameters are updated.

7. The method for dynamic prediction and calibration of remaining life of power grid equipment based on multi-source sensor data fusion according to claim 1, characterized in that, The construction of a dynamic health index based on the fusion of multi-source data driven by causal relationships includes: The sum of the instantaneous transfer entropy of all inputs when each topological feature is used as the target feature is calculated to obtain the dynamic causal weight of each topological feature. Set a health baseline value and a failure threshold value, and use a negative exponential mapping function to unify the polarity of the original observation values ​​of each topological feature and constrain them to a preset health state range. The dynamic health index is obtained by weighting and summing the state values ​​of each feature with their corresponding dynamic causal weights.

8. The method for dynamic prediction and calibration of remaining life of power grid equipment based on multi-source sensor data fusion according to claim 1, characterized in that, Before applying the dynamic health index through a physical information neural network embedded with a material aging physics model, the following steps are also included: The maximum mean difference loss is calculated in the latent space by using accelerated aging data from the source domain and field operation data from the target domain. This loss is minimized through domain adversarial training to reduce the difference in feature distribution between the source and target domains.

9. The method for dynamic prediction and calibration of remaining life of power grid equipment based on multi-source sensor data fusion according to claim 4, characterized in that, After outputting the corrected remaining lifetime prediction results, the following is also included: Obtain measured insulation strength data of power grid equipment after maintenance; Based on measured insulation strength data, a Bayesian back-inference algorithm is used to dynamically update the aging memory coefficient in the physical information neural network.

Images