An online life prediction method considering adaptive fault failure threshold differentiation of transformer in consideration of missing of deterioration information

Abstract

The application discloses a transformer adaptive fault failure threshold differentiation online life prediction method considering missing degradation information, belongs to the field of power equipment degradation evaluation and operation and maintenance, and comprises the following steps: a prediction framework including degradation data analysis, comprehensive degradation data composition, intelligent smoothing fitting, degradation model construction and adaptive dynamic prediction is constructed. Through a fuzzy logic and TIME-LLM hybrid model, multi-index weighted fusion is realized, intelligent adaptive smoothing fitting is adopted to process comprehensive degradation data with different noise characteristics, a TFD-Hformer model is used to capture time-frequency domain nonlinear coupling characteristics, and a variational Bayesian neural network is used to generate a differentiated fault threshold correction factor by combining stochastic variation inference. The application can realize reliable prediction in the historical data missing scene, adaptively adapt to the individual differences of different transformers, and support efficient prediction of large-scale clusters.

Description

Technical Field

[0001] This invention belongs to the field of power equipment degradation assessment and operation and maintenance, and particularly relates to an online life prediction method for transformer adaptive fault failure threshold differentiation that takes into account the lack of degradation information. Background Technology

[0002] As core equipment in the power grid, the service life of power transformers directly affects the reliability of power supply. Predicting the remaining life of transformers aims to accurately assess their health status and estimate their end-of-life, thus providing a basis for condition-based maintenance and operation and maintenance decisions. Existing technologies for remaining life prediction largely rely on continuous and complete degradation trajectory data of the transformer throughout its entire lifecycle, fitting degradation trends by constructing physical or data-driven models. Commonly used methods include degradation models based on single or multi-indicator fusions such as dissolved gas analysis in oil, furfural concentration, and degree of polymerization, as well as time-series prediction methods based on deep learning such as Long Short-Term Memory Networks and Transformers. These methods typically assume that the monitoring data is relatively complete and reflects the complete degradation process from health to failure, and establish specific prediction models for transformers of specific types or operating conditions.

[0003] However, existing technologies still have the following problems in practical engineering applications: First, due to untimely sensor deployment, communication channel blockage, low monitoring frequency, or maintenance operations (such as oil filtration and oil change), historical monitoring data of transformers often suffers from large-scale and long-term gaps, resulting in discontinuous or even completely lost deterioration trajectories. Existing methods struggle to build effective deterioration models and achieve reliable predictions when deterioration information is severely lacking or incomplete. Second, different transformers exhibit significant differences in manufacturing processes, structural types, operating environments (such as temperature, humidity, air pressure, and salt spray concentration), and operating conditions (such as load fluctuations and vibration noise). Existing prediction models are often built separately for a specific type of equipment or specific operating condition, lacking a generalized adaptive framework. When applied to large-scale, multi-type transformer clusters, repeated modeling is required for each piece of equipment or each scenario, leading to a sharp increase in time costs and computational resource consumption, making it difficult to achieve cost-effective and efficient large-scale deployment. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an online lifetime prediction method for transformer adaptive fault failure threshold differentiation that considers the lack of degradation information, comprising: Based on the transformer type, operating environment, operating conditions, operating time and historical data of the target transformer, a prediction framework is constructed that includes deterioration data analysis, comprehensive deterioration data composition, deterioration data smoothing fitting, deterioration model construction and adaptive dynamic prediction. Based on the prediction framework and the differences between the target transformer and similar transformers, the historical data of the target transformer are supplemented and corrected to obtain a complete input dataset. Based on the input dataset, a hybrid model combining fuzzy logic and TIME-LLM model is used for weight analysis and environmental correction to generate comprehensive degradation data. Based on the temporal characteristics of the comprehensive degradation data, an intelligent adaptive smoothing fitting method is used to process the data and obtain the fitted degradation trajectory. Based on the fitted degradation trajectory and the covariates of the operating environment and operating conditions, a degradation model is constructed using a hybrid time-frequency domain enhancement decomposition model based on an improved Transformer architecture. Based on real-time monitoring data, the parameters of the degradation model are dynamically adjusted by combining variational Bayesian neural network with stochastic variational inference algorithm, and a fault threshold correction factor for the target transformer is generated to output the remaining life prediction result.

[0005] Optionally, it also includes: performing sensitivity analysis based on the historical data of the target transformer and its differential factors to obtain the variation range of deterioration data under the influence of different factors.

[0006] Optionally, the generation of comprehensive degradation data specifically includes: Based on the input dataset, each degradation index is mapped to a degradation level and quantified into a single index degradation score using fuzzy logic. Based on the single-indicator degradation score, the weights of indicators at each time step are dynamically learned and long-term dependencies are captured by using a TIME-LLM model incorporating an attention mechanism, and a weighted fusion is performed to obtain a preliminary comprehensive degradation index. The preliminary comprehensive degradation index is corrected for deviation based on historical fault data, and a correction coefficient is introduced based on the operating environment to form the comprehensive degradation data.

[0007] Optionally, the intelligent adaptation smooth fitting method includes: When the overall deterioration data fluctuates greatly in the short term but remains stable overall, a Gaussian weighted moving average filter is used for processing. When the comprehensive deterioration data exhibits a nonlinear trend and key inflection points need to be preserved, adaptive smoothing based on quantum neural networks is used for processing. When the overall degraded data is high noise, quantum adaptive transformation is used to decompose and reconstruct multi-scale frequency components.

[0008] Optionally, the construction of the degradation model using the hybrid time-frequency domain enhancement decomposition model of the improved Transformer architecture specifically includes: Based on the fitted degraded trajectory and covariates, multi-scale temporal dependencies are captured through the multi-head self-attention mechanism of the hybrid time-frequency domain augmented decomposition model of the improved Transformer architecture. The frequency domain enhancement module maps the degraded time-domain data and covariates to the time-frequency joint space to learn nonlinear coupling characteristics. The dynamic weight allocation mechanism of the decomposed trend learning module is used to enhance the data influence at key time points. The layered residual learning module is used to identify degradation inflection points, and the output layer is designed to classify three stages: slow degradation, accelerated degradation, and critical state.

[0009] Optionally, the step of dynamically adjusting the parameters of the degradation model using a variational Bayesian neural network combined with a stochastic variational inference algorithm, and generating a fault threshold correction factor for the target transformer, specifically includes: Based on the real-time incoming monitoring data, the posterior distribution of the network weights is dynamically adjusted using the variational Bayesian neural network and the stochastic variational inference algorithm. By using a latent variable hierarchical mechanism, the fault threshold correction factor is generated based on the historical stability, operating condition fluctuation intensity, and environmental impact of the target transformer, so as to adaptively adjust the early warning boundary.

[0010] Optionally, the difference factors include: The target transformer's type, operating environment, operating conditions, operating time, and historical data trajectory; The transformer type is used to guide the screening of historical data; the operating environment and operating conditions are used to provide environmental correction coefficients and serve as covariates; the operating time is used to assist in judging data trends and identifying deterioration stages; and the historical data serves as the original basis for generating single-index deterioration scores.

[0011] Optionally, the historical data of the target transformer is supplemented and corrected to obtain a complete input dataset, specifically including: Based on the type, operating time, operating environment and operating conditions of the target transformer, historical data of transformers of the same type and similar environment and operating conditions accumulated in the prediction framework are called; Based on the degradation patterns of similar service times matched with the operating time, the missing degradation index data is supplemented; By combining the current monitoring values ​​with the environmental-operating condition association rules of similar equipment, the unreliable supplementary data is corrected to form the complete input dataset.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: First, addressing the issue of missing historical degradation data, this invention integrates a fuzzy logic enhancement mechanism with a data migration compensation strategy for similar equipment. Based on the target transformer's type, operating environment, and operating conditions, it intelligently matches historical degradation patterns of similar equipment under similar conditions, migrating, completing, and correcting missing data. This allows for the construction of an effective comprehensive degradation index even in scenarios with incomplete information, enabling reliable prediction of transformer remaining lifespan and overcoming the reliance on a complete full-cycle degradation trajectory. Second, addressing the significant differences between different transformers and the high cost of large-scale deployment, this invention incorporates a differentiated configuration system based on multi-dimensional parameters such as transformer type, operating environment, and operating conditions. This system automatically adapts to the degradation characteristics of different equipment. Furthermore, it employs a variational Bayesian neural network combined with a stochastic variational inference algorithm to achieve adaptive differential adjustment of fault thresholds, significantly reducing the time and cost of individual modeling for different equipment. This efficiently supports batch remaining life prediction for large-scale transformer clusters. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a specific understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 A schematic diagram showing how degradation data can be generated with a single click after considering various factors in a transformer embodiment of the present invention. Figure 3 This is a flowchart illustrating the dynamic trend prediction process using VNBBs combined with SVI in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides an online lifetime prediction method for transformer adaptive fault failure threshold differentiation that considers the lack of degradation information, including: Based on the transformer type, operating environment, operating conditions, operating time and historical data of the target transformer, a prediction framework is constructed that includes deterioration data analysis, comprehensive deterioration data composition, deterioration data smoothing fitting, deterioration model construction and adaptive dynamic prediction. Based on the prediction framework and the differences between the target transformer and similar transformers, the historical data of the target transformer are supplemented and corrected to obtain a complete input dataset. Based on the input dataset, a hybrid model combining fuzzy logic and TIME-LLM model is used for weight analysis and environmental correction to generate comprehensive degradation data. Based on the temporal characteristics of the comprehensive degradation data, an intelligent adaptive smoothing fitting method is used to process the data and obtain the fitted degradation trajectory. Based on the fitted degradation trajectory and the covariates of the operating environment and operating conditions, a degradation model is constructed using a hybrid time-frequency domain enhancement decomposition model based on an improved Transformer architecture. Based on real-time monitoring data, the parameters of the degradation model are dynamically adjusted by combining variational Bayesian neural network with stochastic variational inference algorithm, and a fault threshold correction factor for the target transformer is generated to output the remaining life prediction result.

[0019] In one specific embodiment of the present invention, the prediction framework is as follows: Figure 1 As shown.

[0020] In practice, the process begins by acquiring monitoring data on various degradation indicators of different transformers under diverse environmental conditions using sensors. Then, a sensitivity analysis is performed on the differences between these data to determine the magnitude of degradation data changes caused by variations in different influencing factors and their degrees. This provides a reliable basis for predicting the remaining lifespan of transformers when historical data is unreliable or missing.

[0021] Based on the obtained monitoring data of various degradation indicators, this example uses a hybrid model combining fuzzy logic and TIME-LLM to perform weight analysis on the data of each indicator, thereby obtaining the comprehensive degradation data of different transformers under different environments. The specific operation is as follows: First, the data for each degradation indicator are standardized, with unified dimensions and outliers removed. Second, using a fuzzy logic module, based on industry standards and expert rules, the standardized indicator data are mapped to different degradation levels, and each degradation level is quantified into a single indicator degradation score. Then, these scores are input into a TIME-LLM model that incorporates an attention mechanism. This model dynamically learns the weights of each indicator at different times and captures long-term dependencies between indicators. After weighted fusion, a preliminary comprehensive degradation index is obtained. Finally, historical fault data is used for bias correction, and correction coefficients are introduced for special environments to form the final comprehensive degradation data.

[0022] This invention employs a TIME-LLM model that incorporates an attention mechanism, the specific structure of which is as follows.

[0023] I. Overall structure of the model; The TIME-LLM model consists of the following four modules: Input embedding module: Standardizes multi-source degradation indicators, encodes time location and transformer type; Attention encoder: Employs a multi-layer Transformer encoder, with a multi-head self-attention mechanism at its core; Weight generation module: Aggregates attention weights and dynamically outputs the fusion coefficients of various degradation indicators; Comprehensive degradation output module: After weighted fusion, the comprehensive degradation index is output through the regression layer and bias correction is performed; II. Integration methods of attention mechanisms; The combination of attention mechanisms and the TIME-LLM model is reflected in the following three aspects: (1) Self-attention fusion in the time step dimension; Let the length of the input sequence be... The number of degradation indicators is The input matrix after embedding is ,in For the embedding dimension. The calculation method for multi-head self-attention is as follows: ; ; in , , These are query, key, and value matrices, respectively. Scaling factor The number of attention heads. This mechanism enables the model to capture degradation dependencies between any two time steps, such as the effect of slight early degradation on accelerated degradation later.

[0024] (2) Interactive attention fusion based on feature dimensions; To differentiate the importance of different degradation metrics, this invention introduces feature interaction attention at each time step to dynamically weight the feature dimensions: In the formula For time step The hidden state, and For learnable parameters, For the first Each degradation indicator at time step Attention weights. This mechanism enables the model to automatically adjust index weights based on environment and operating conditions, such as automatically increasing the weight of the water content index in oil under high humidity conditions.

[0025] (3) Temporal-feature joint attention encoding; This invention further employs a joint attention encoding strategy, mapping time steps and feature dimensions to a unified sequence, enabling the attention mechanism to simultaneously learn "when" and "which features" are more important for degradation trends. The specific implementation is as follows: Input matrix Flattened to a length of The joint sequence; Incorporate learnable temporal location encoding and feature type encoding; The input is fed into a Transformer encoder with shared parameters, and the output is a joint attention-enhanced representation. ; Restored to its original state through aggregation operations The temporal representation.

[0026] III. Calculation of the overall deterioration index; The attention-enhanced feature vector is input into the linear regression layer, and the overall degradation index at the current time step is output: ; In the formula For time step Joint attention enhancement features, and These are the parameters of the regression layer.

[0027] Subsequently, deviation corrections were performed by incorporating historical fault data: ; in This is a deviation correction term learned from historical fault data of similar equipment.

[0028] Finally, a correction factor is introduced for special environments. : ; This is the final comprehensive degradation data, which is used for subsequent degradation modeling and lifetime prediction.

[0029] To address the differences in the temporal continuity of comprehensive degradation data, this embodiment establishes a technique for intelligently adapting and smoothing the fit based on the degradation trajectory. Specifically: When the data fluctuates significantly in the short term but the overall trend is stable, a Gaussian weighted moving average filter is used. This method dynamically allocates weights based on a Gaussian distribution and performs weighted calculations on local window data, giving higher weights to data near the center. This allows it to retain local trend characteristics more accurately than the traditional moving average while weakening high-frequency noise. When data exhibits a nonlinear trend and it is necessary to retain key inflection points, an adaptive smoothing based on quantum neural networks is adopted. This model utilizes the properties of quantum superposition and entanglement to construct an adaptive model. It does not require a pre-set fixed trend factor and can dynamically learn the nonlinear laws of the data. While tracking long-term trends, it accurately retains key features such as inflection points and effectively suppresses random fluctuations. For high-noise scenarios, quantum adaptive transformation is combined, and quantum parallelism is used to decompose the data into multi-scale frequency components. After filtering out the noise components through adaptive thresholding, the signal is reconstructed. Its dynamically adjusted transformation characteristics can preserve the original trend features to the greatest extent while performing deep denoising, making it particularly suitable for mixed noise scenarios that are difficult to handle by traditional methods.

[0030] Finally, the fitting effect was verified by residual analysis: the sum of squared residuals between the fitted curve and the original data was calculated. If the residuals showed a random distribution and the amplitude was small, it indicated that the smoothing effect was good. If there was a systematic bias, the smoothing parameters were adjusted until the fitted curve could effectively eliminate noise and accurately reflect the true evolution trend of the overall deterioration data.

[0031] In this embodiment, comprehensive degradation data is used as the core input, and covariates such as environment and operating conditions are also included to construct a hybrid time-frequency domain enhancement decomposition model (i.e., TFD-Hformer) based on an improved Transformer architecture as the degradation model.

[0032] The TFD-Hformer model described in this invention consists of the following four core modules: (1) Timing partitioning and input embedding module; The comprehensive degradation data is divided into sequence samples at a fixed time step, and each sample incorporates environmental parameters and operating condition parameters as multi-dimensional covariates. Linear mapping and positional encoding are performed on each sequence to form an input feature vector with temporal semantics.

[0033] (2) Frequency domain enhancement module (F2A-unit); This module is used to synchronously map temporal degradation data and covariates to a time-frequency joint space. Specifically, it performs a short-time Fourier transform on the input sequence to extract the energy distribution features of the degradation trend across different frequency bands; then, through a learnable frequency domain attention mechanism, it filters out frequency band components closely related to degradation acceleration and the occurrence of inflection points. The output of this module is an enhanced temporal representation that incorporates frequency domain features.

[0034] (3) Bullish Self-Attention and Trend Decomposition Module (DeCom-unit); Building upon frequency domain enhancement, the model employs a multi-head self-attention mechanism to capture multi-scale temporal dependencies. Simultaneously, a trend decomposition module is introduced, which uses a dynamic weight allocation mechanism to decompose the degradation sequence into slow trend components and fluctuating components. This module assigns differentiated attention weights to different degradation stages, strengthening the data influence of key time nodes such as abrupt changes in operating conditions and dramatic shifts in environmental parameters.

[0035] (4) Hierarchical residual learning module (SHRUL-unit); To improve the model's accuracy in identifying degradation inflection points, this invention designs a hierarchical residual learning module. This module introduces cross-layer residual connections within each layer of the Transformer and performs weighted fusion of local temporal features extracted from the lower layers with global semantic features extracted from the higher layers, enabling the model to retain key details of the original degradation trajectory even in deep networks.

[0036] Based on the above structure, the specific construction process of the TFD-Hformer model is as follows: The specific construction process is as follows: First, the comprehensive degradation data is divided into time series to form sequence samples with time steps; then, the multi-head self-attention mechanism of TFD-Hformer is used to capture multi-scale temporal dependencies, and the time-domain degradation data and multi-dimensional covariates are mapped to the time-frequency joint space through the frequency domain enhancement module (F2A-unit) to learn the nonlinear coupling characteristics between the comprehensive degradation value and environmental and operating condition variables; next, the output layer is designed to classify three stages: slow degradation, accelerated degradation, and critical state, and the dynamic weight allocation mechanism of the decomposition trend learning module (DeCom-unit) is used to strengthen the data influence of key time nodes such as sudden changes in operating conditions and drastic changes in environmental parameters; at the same time, the hierarchical residual learning (SHRUL-unit) is combined to improve the model's accuracy and robustness in identifying degradation inflection points.

[0037] like Figure 3As shown, to achieve online updating of model parameters and quantification of uncertainty in prediction results, this example employs a variational Bayesian neural network (VBNNs) combined with a stochastic variational inference (SVI) algorithm. This module can dynamically adjust the posterior distribution of network weights based on real-time incoming monitoring data. More importantly, through a latent variable hierarchical mechanism, a unique fault threshold correction factor is generated for each device. This factor can automatically adjust the warning boundary based on the device's historical stability, operating condition fluctuation intensity, and environmental influences, thereby truly achieving adaptive differentiation of fault failure thresholds.

[0038] The combination of Variational Bayesian Neural Networks (VBNNs) and Stochastic Variational Inference (SVI) algorithm, and its combination process and model structure are as follows: I. Model Structure; VBNNs, based on standard neural networks, replace deterministic values ​​with probability distributions for the weight parameters of each layer. Let the network's... The weights of the layers follow a Gaussian distribution: ; in and These are the mean and standard deviation of the weights for this layer, respectively, both of which are trainable parameters. The network output is no longer a fixed value, but a predicted distribution: ; In the formula Input features (comprehensive degradation data, environmental parameters, operating condition parameters). This is the predicted remaining lifespan. To improve the accuracy of noise observation.

[0039] II. The process of combining variational inference with SVI; VBNNs use variational inference to approximate the true posterior distribution. Introducing variational distribution By minimizing The KL divergence between the true posterior and the actual posterior is optimized. The optimization objective is the lower bound of evidence (ELBO): ; The first term is the likelihood expectation (fit accuracy), and the second term is the variational distribution and prior. The KL divergence (regularization term) between them.

[0040] To achieve large-scale online updates, this invention introduces the Stochastic Variational Inference (SVI) algorithm. SVI decomposes ELBO into the sum of contributions from each data point, and each iteration only randomly samples a mini-batch of data. The variational parameters are updated using natural gradient descent: ; ; in For the first The learning rate is updated every time. SVI enables the model to quickly update the posterior distribution of weights based on the current mini-batch of samples, without retraining on the full dataset, as the model monitors the inflow of data in real time.

[0041] III. Adaptive Fault Threshold Differentiation Correction To achieve personalized early warning for different transformers, this invention introduces a hidden variable hierarchical mechanism in VBNNs. Let the first... Fault threshold correction factor for transformer By latent variables generate: ; Among them, latent variables Comply with the equipment's historical stability index Prior distribution conditioned on (including historical degradation fluctuations, frequency of sudden changes in operating conditions, and number of extreme environmental events): ; The latent variable is updated synchronously with the network weights through variational inference and SVI updates. Finally, the... The adaptive fault failure threshold for the transformer is: ; In the formula This serves as the baseline fault threshold. When the equipment has poor historical stability, experiences high fluctuations in operating conditions, or is significantly affected by environmental factors... Automatically decrease the warning threshold to tighten it; conversely, increase it to avoid frequent false alarms.

[0042] IV. Online update process; 1. Real-time incoming monitoring data (comprehensive degradation index, environmental parameters, and operating condition parameters) form a small batch sample; 2. Calculate the ELBO gradient under the current mini-batch using the SVI algorithm, and update the mean and standard deviation of all weight distributions in VBNNs; 3. Synchronously update hidden variables on each device. Recalculate the correction factor ; 4. Output the predicted remaining lifetime (the mean of the predicted distribution) and the 95% confidence interval; 5. When the predicted value is lower than the adaptive threshold An alert is triggered at any time.

[0043] The process of intelligently matching different factors between transformers includes: In this embodiment, different transformers exhibit differences in factors such as type, operating environment, operating conditions, operating time, and historical data. Specifically: Transformer types include, but are not limited to, oil-immersed transformers and dry-type transformers; The operating environment includes specific parameters such as temperature, humidity, air pressure, rainfall frequency, solar radiation intensity, and concentration of corrosive substances in the surrounding environment; Operating conditions include load size, vibration and noise, voltage, current, etc. Operating time refers to the service time of a transformer from commissioning to the present moment; Historical data includes the trajectory trends of indicators that can reflect and influence the degree of transformer deterioration, such as the content of various dissolved gases in the oil, the concentration of furfural in the oil, the concentration of methanol in the oil, and the amount of partial discharge.

[0044] In this embodiment, the intelligent combination structure of the prediction framework and transformer factors is as follows: Figure 2 As shown, this prediction framework serves as the algorithm selection basis for predicting the remaining life of transformers. It can quickly and effectively determine the remaining life of a specific transformer based on the specific parameters of its four sub-modules. Each sub-module acts as a "parameter input source" for the prediction framework, providing targeted data support for the core components. (1) The transformer type is used to guide the targeted screening and interpretation of historical data in the “deterioration data analysis” stage, so as to ensure that the single indicator deterioration score extracted from the historical data can accurately reflect the true deterioration degree of the transformer of this type, thereby providing a guarantee for the accuracy of subsequent comprehensive deterioration data.

[0045] (2) The operating environment and operating conditions provide the basis for calculating the environmental correction coefficient in the "Comprehensive Deterioration Data Composition" stage. At the same time, they are included as covariates in the "Constructing Deterioration Model" stage to help the TFD-Hformer model learn the impact of changes in environment and operating conditions on the remaining life, such as the deterioration acceleration effect caused by a sudden increase in load.

[0046] (3) Historical data serves as the basic data source for the “deterioration data analysis” step, providing the original basis for generating single-index deterioration scores using fuzzy logic and the TIME-LLM model. Its completeness and reliability directly affect the accuracy of the comprehensive deterioration data.

[0047] (4) Running time serves as a key reference for time series characteristics, assisting in judging the data trend type in the "smoothing fit" stage and strengthening the identification of equipment degradation stages in the "constructing degradation model" stage.

[0048] Meanwhile, the prediction framework achieves adaptive adjustment by dynamically calling submodule parameters. Specifically, in the "dynamic trend prediction" stage, when the transformer's operating environment or operating conditions change significantly (e.g., a sudden increase in temperature and humidity or continuous overload), the VBNNs model is triggered to update online first. The network parameters are quickly adjusted through the SVI algorithm to adapt the prediction results to the new operating conditions. If the historical data is missing or unreliable, the weight of the expert rules in the fuzzy logic module is automatically strengthened, and compensation and correction are performed in combination with the operating environment and operating condition data of the same period, thereby ensuring the consistency of the comprehensive degradation data.

[0049] Ultimately, through real-time data interaction and parameter linkage between sub-modules and the prediction framework, the entire chain of "input parameters - processing strategy - prediction results" is adapted. This not only accurately captures the degradation characteristics of specific transformers under their unique environment and operating conditions, but also continuously optimizes the prediction logic through dynamic feedback, forming a customized remaining life prediction scheme for individual equipment.

[0050] In this embodiment, the prediction method specifically includes the following processes.

[0051] First, data collection and training are conducted. Specifically, complete data on various types of transformers is collected, including transformer type, operating environment, operating conditions, operating time, and historical data. The collected data is then used for learning, training, and sensitivity analysis.

[0052] Secondly, data supplementation and integration are performed. Specifically, the type, operating time, and environmental and operating condition data factors of the target transformer are determined. Then, historical data of transformers of the same type and with similar operating environments and conditions as the target transformer, accumulated in the "deterioration data analysis" step of the prediction framework, are used to supplement the missing deterioration index data based on the deterioration patterns of similar service times matching the operating time. For unreliable parts in the supplemented data, corrections are made by combining the current monitoring values ​​with the environmental-operating condition correlation rules of similar equipment, thereby forming a complete input dataset.

[0053] Finally, adaptability analysis and prediction are performed. Specifically, during the construction of comprehensive degradation data, the expert rule weights of the fuzzy logic module are strengthened to reduce the impact of unreliable data. Single-indicator degradation scores are generated using the supplemented dataset, and after weighted fusion and environmental correction using the TIME-LLM model, comprehensive degradation data is obtained. An intelligent adaptive smoothing fitting method is used to smoothly fit the comprehensive degradation data, focusing on preserving the trend characteristics reflected by the current monitoring values. A TFD-Hformer degradation model is constructed based on the supplemented data, strengthening the correlation weight between the current monitoring values ​​and operating time through an attention mechanism (e.g., increasing the weight of insulation degradation indicators for equipment that has been in service for more than 10 years). The online update mechanism of VBNNs is activated, using the current monitoring data as new samples to optimize model parameters, outputting remaining lifetime prediction results adapted to scenarios with missing or unreliable data, and continuously dynamically adjusting as new monitoring data is input.

[0054] Example 2 This embodiment provides an online lifetime prediction method for transformer adaptive fault failure threshold differentiation that considers the lack of degradation information, including: In this embodiment, 100 transformers of different types in a regional power grid with missing historical data are used as examples to verify the adaptability of the present invention in large-scale, multi-scenario applications. The 100 transformers include 50 oil-immersed transformers, 30 dry-type transformers, and 20 special transformers, covering four types of environments: coastal high humidity, plateau low air pressure, inland aridity, and industrial pollution. All transformers have been in operation for 5 to 8 years.

[0055] I. The modular parameter configuration process includes: Based on transformer type, environmental characteristics, and data integrity, differentiated parameter templates are loaded in batches through intelligent linkage between inter-transformer differences and the prediction framework. The specific configuration process is as follows: (a) Configure core parameters according to transformer type.

[0056] For 50 oil-immersed transformers, the "oil quality-dominated degradation template" is applied, with the following basic weight allocation: dissolved gas in oil accounts for 0.35, water in oil accounts for 0.25, partial discharge accounts for 0.20, and furfural concentration accounts for 0.20.

[0057] For 30 dry-type transformers, the "solid insulation dominant template" is loaded, and its basic weight allocation is as follows: insulation paper polymerization degree accounts for 0.40, winding temperature accounts for 0.25, partial discharge accounts for 0.20, and vibration noise accounts for 0.15.

[0058] For 20 special transformers (such as SF6 insulated transformers), the "gas insulation template" is loaded, and its basic weight allocation is as follows: SF6 decomposition products account for 0.40, gas pressure accounts for 0.30, and partial discharge accounts for 0.30.

[0059] (ii) Configure correction parameters according to environmental characteristics.

[0060] For 15 transformers located in coastal high-humidity environments (humidity 80%–90%), the "humidity correction factor 1.2" is automatically activated, and the dynamic weight of the oil moisture index is increased from 0.25 to 0.30. For coastal transformers with salt spray concentration of 0.02–0.05 mg / m³, an additional "salt spray corrosion factor" is added, which increases the dielectric loss weight of bushing insulation deterioration by 10%.

[0061] For 20 transformers located in a high-altitude, low-pressure environment (altitude 2000–3000m, air pressure 60–80kPa), a "low-pressure correction factor of 1.15" is applied to reduce the partial discharge allowable value from 10pC to 8pC to lower the insulation strength threshold, and to increase the weight of the influence of winding temperature on degradation from 0.25 to 0.30.

[0062] For 10 transformers in an industrial pollution environment (surrounding SO2 concentration 0.1–0.3 mg / m³), a "chemical corrosion template" was applied to increase the weight of the acid value index in the oil from 0.15 to 0.25 and correlate it with the rainfall frequency (if the monthly rainfall days are ≥15 days, the acid value deterioration rate is multiplied by a coefficient of 1.1).

[0063] For 55 transformers located in inland arid environments (humidity 30%–50%), a "dry environment correction factor of 0.9" is applied to reduce the weight of moisture in the oil from 0.25 to 0.20 and to enhance the impact of load fluctuations on deterioration.

[0064] (iii) Configure dynamic parameters according to load characteristics.

[0065] For 30 high-load transformers (daily average load rate 80%–100%, including impact load), a "load-sensitive template" is applied so that the contribution weight of winding hot spot temperature deterioration increases by 5% for every 10% increase in load rate over the rated value.

[0066] For 70 low-load transformers (daily average load rate 30%–60%), a "stable load template" is adopted, which reduces the weight of load-related parameters by 10% by default and instead strengthens the long-term impact of ambient temperature (for example, for every 5°C increase in the annual average temperature, the weight of the insulation degradation index increases by 8%).

[0067] (iv) Configure compensation parameters according to data integrity. For transformers with missing historical data, increase the weight of expert rules to 50% and call the "environment-operating condition-deterioration correlation library" of similar transformers to supplement the data.

[0068] II. The batch forecasting process includes: By linking the differences between transformers with the parameters of the prediction framework, the remaining life of the aforementioned 100 transformers is predicted. The specific process is as follows: First, determine the real-time environment, operating conditions, running time, and available historical data for each transformer, and automatically match the parameter templates from the aforementioned modular parameter configuration stage.

[0069] Next, for the supplemented data samples, a comprehensive degradation calculation based on fuzzy logic and TIME-LLM is directly performed; for data samples with missing data, a compensation formula is activated to generate pseudo data, which is then integrated with the current monitoring values ​​to output a comprehensive degradation index.

[0070] Then, a unified dynamic smoothing strategy is adopted: for example, wavelet transform is automatically applied to high-noise data, triple exponential smoothing is applied to nonlinear trends, and degradation stage classification of "slow degradation / accelerated degradation / critical state" is output in batches through LSTM or GRU models.

[0071] Finally, VBNNs combined with the SVI algorithm are used to batch adjust the model parameters according to the streaming data update frequency of the regional power grid, ensuring that the decommissioning threshold can be adaptively adjusted, so that the prediction results can adapt to sudden changes in the environment or operating conditions in real time.

[0072] In summary, by considering the differences between transformers and the dynamic adaptation mechanism, this invention can efficiently cover the prediction of the remaining life of transformers under different types, different environmental conditions, and different historical data integrity conditions, while taking into account both prediction accuracy and computational efficiency.

[0073] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0074] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0075] The above are merely preferred embodiments 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.

Claims

1. An online lifetime prediction method for transformers that considers the differential nature of adaptive fault failure thresholds due to missing degradation information, characterized in that, include: Based on the transformer type, operating environment, operating conditions, operating time and historical data of the target transformer, a prediction framework is constructed that includes deterioration data analysis, comprehensive deterioration data composition, deterioration data smoothing fitting, deterioration model construction and adaptive dynamic prediction. Based on the prediction framework and the differences between the target transformer and similar transformers, the historical data of the target transformer are supplemented and corrected to obtain a complete input dataset. Based on the input dataset, a hybrid model combining fuzzy logic and TIME-LLM model is used for weight analysis and environmental correction to generate comprehensive degradation data. Based on the temporal characteristics of the comprehensive degradation data, an intelligent adaptive smoothing fitting method is used to process the data and obtain the fitted degradation trajectory. Based on the fitted degradation trajectory and the covariates of the operating environment and operating conditions, a degradation model is constructed using a hybrid time-frequency domain enhancement decomposition model based on an improved Transformer architecture. Based on real-time monitoring data, the parameters of the degradation model are dynamically adjusted by combining variational Bayesian neural network with stochastic variational inference algorithm, and a fault threshold correction factor for the target transformer is generated to output the remaining life prediction result.

2. The method according to claim 1, characterized in that, Also includes: Based on the historical data of the target transformer and its differential factors, a sensitivity analysis was conducted to obtain the variation range of the degradation data under the influence of different factors.

3. The method according to claim 1, characterized in that, The generation of comprehensive degradation data specifically includes: Based on the input dataset, each degradation index is mapped to a degradation level and quantified into a single index degradation score using fuzzy logic. Based on the single-indicator degradation score, the weights of indicators at each time step are dynamically learned and long-term dependencies are captured by using a TIME-LLM model incorporating an attention mechanism, and a weighted fusion is performed to obtain a preliminary comprehensive degradation index. The preliminary comprehensive degradation index is corrected for deviation based on historical fault data, and a correction coefficient is introduced based on the operating environment to form the comprehensive degradation data.

4. The method according to claim 1, characterized in that, The intelligent adaptation and smooth fitting method includes: When the overall deterioration data fluctuates greatly in the short term but remains stable overall, a Gaussian weighted moving average filter is used for processing. When the comprehensive deterioration data exhibits a nonlinear trend and key inflection points need to be preserved, adaptive smoothing based on quantum neural networks is used for processing. When the overall degraded data is high noise, quantum adaptive transformation is used to decompose and reconstruct multi-scale frequency components.

5. The method according to claim 1, characterized in that, The construction of the degradation model using the hybrid time-frequency domain enhancement decomposition model of the improved Transformer architecture specifically includes: Based on the fitted degraded trajectory and covariates, multi-scale temporal dependencies are captured through the multi-head self-attention mechanism of the hybrid time-frequency domain augmented decomposition model of the improved Transformer architecture. The frequency domain enhancement module maps the degraded time-domain data and covariates to the time-frequency joint space to learn nonlinear coupling characteristics. The dynamic weight allocation mechanism of the decomposed trend learning module is used to enhance the data influence at key time points. The layered residual learning module is used to identify degradation inflection points, and the output layer is designed to classify three stages: slow degradation, accelerated degradation, and critical state.

6. The method according to claim 1, characterized in that, The step of dynamically adjusting the parameters of the degradation model using a variational Bayesian neural network combined with a stochastic variational inference algorithm, and generating a fault threshold correction factor for the target transformer, specifically includes: Based on the real-time incoming monitoring data, the posterior distribution of the network weights is dynamically adjusted using the variational Bayesian neural network and the stochastic variational inference algorithm. By using a latent variable hierarchical mechanism, the fault threshold correction factor is generated based on the historical stability, operating condition fluctuation intensity, and environmental impact of the target transformer, so as to adaptively adjust the early warning boundary.

7. The method according to claim 1, characterized in that, The factors that cause the difference include: The target transformer's type, operating environment, operating conditions, operating time, and historical data trajectory; The transformer type is used to guide the screening of historical data; the operating environment and operating conditions are used to provide environmental correction coefficients and serve as covariates; the operating time is used to assist in judging data trends and identifying deterioration stages; and the historical data serves as the original basis for generating single-index deterioration scores.

8. The method according to claim 1, characterized in that, The historical data of the target transformer is supplemented and corrected to obtain a complete input dataset, which specifically includes: Based on the type, operating time, operating environment and operating conditions of the target transformer, historical data of transformers of the same type and similar environment and operating conditions accumulated in the prediction framework are called; Based on the degradation patterns of similar service times matched with the operating time, the missing degradation index data is supplemented; By combining the current monitoring values ​​with the environmental-operating condition association rules of similar equipment, the unreliable supplementary data is corrected to form the complete input dataset.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images