Method and system for predicting service life of transformer

Abstract

The invention discloses a transformer life prediction method and system, and relates to the technical field of transformer maintenance, multi-source data of a transformer is collected, macroscopic features and microscopic features of the multi-source data are extracted, the correlation degree of the microscopic features to the macroscopic features is evaluated through mutual information and random forest double indexes, and the transformer life prediction result is obtained. Degradation indexes are constructed in different scenes according to the correlation degree, reliability is ensured through monotonicity, robustness and correlation verification, finally, the residual service life of the transformer is predicted in different scenes through a gating circulation unit or a support vector regression model, and through multi-source data fusion and intelligent algorithm innovation, precision and early stage of residual service life prediction are achieved. Aging symptoms can be recognized in advance, different transformers can be adapted, a basis is provided for preventive maintenance, and the fault risk and the operation and maintenance cost are reduced.

Description

Technical Field

[0001] This invention relates to the field of transformer maintenance technology, and in particular to a method and system for predicting transformer lifespan. Background Technology

[0002] As a crucial energy conversion device in the power system, the power transformer is primarily used to regulate voltage and transform the form of electrical energy transmission. Internally, a power transformer contains complex electrical and mechanical components, including insulation materials, coils, and cooling systems. Over time, these components age due to voltage, thermal, and mechanical stresses, leading to performance degradation and increased risk of failure. Timely maintenance and replacement are essential to ensure stable operation of power transformers. However, improper timing of maintenance and replacement can result in unplanned power outages and reduced equipment efficiency. In more serious cases, it can threaten the safe operation of the entire power grid, further impacting the stability of power supply and the economic benefits of power companies. Therefore, research on power transformer maintenance technology has always been a key focus for power companies.

[0003] Currently, Chinese invention patent application CN118246246B discloses a method and system for predicting transformer lifespan based on multi-source data fusion. By acquiring test data and operating condition data of power transformers, it constructs an operating status assessment system, calculates a comprehensive health index, and, based on the Weibull distribution model and failure rate correction coefficient, achieves lifespan prediction under both maintenance-free and maintenance-free conditions. The limitations of this existing technology are mainly reflected in: insufficient data dimensions, relying primarily on traditional macroscopic parameters and lacking utilization of microscopic data such as high-frequency electromagnetic noise and the dynamic behavior of microbubbles in oil, resulting in limited ability to capture early degradation features and difficulty in achieving true early warning; and limitations in feature extraction capabilities. Limited by its limitations, although weight calculation combines subjective and objective methods, feature extraction still relies on a static indicator system and does not incorporate a deep learning model. Therefore, it cannot adaptively extract deep degradation information from complex data and has weak resistance to noise interference. The model also exhibits poor adaptability; failure rate correction is based on Weibull distribution and empirical formulas, and the correction process depends on historical maintenance data. This results in insufficient generalization ability for transformers operating under varying conditions or new operating modes, and it does not consider the dynamic correlation between microscopic physical changes and macroscopic degradation. Furthermore, the model fails to strike a good balance between real-time performance and accuracy, focusing on post-maintenance status correction. However, the prediction model is sensitive to short-term fluctuations and struggles to maintain high accuracy in data-scarce or noisy environments, thus failing to meet real-time monitoring requirements. Summary of the Invention

[0004] The technical problem solved by this invention is that the existing technology is unable to solve the pain points of traditional prediction methods, such as insensitivity to early aging detection, incomplete feature extraction, insufficient prediction accuracy and poor adaptability.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a transformer life prediction method, comprising the following steps: Step S1: Collect multi-source data of the transformer, including condition monitoring data, static data and micro data; Step S2: Extract degradation features from the state monitoring data, extract static features from the static data, extract micro features from the micro data, use the degradation features and static features as macro features, evaluate the correlation between the micro features and macro features, and obtain the correlation degree between the micro features and macro features. The correlation degree includes a first correlation degree and a second correlation degree. Step S3: When the correlation degree is the first correlation degree, a first degradation index is constructed based on degradation features and static features; when the correlation degree is the second correlation degree, a second degradation index is constructed based on degradation features, static features and micro features. Step S4: When using the first degradation index, the first remaining lifetime is predicted by a pre-trained gated recurrent unit time series model. When using the second degradation index, the second degradation index and micro-features are input into a pre-trained support vector regression model to output the second remaining lifetime.

[0006] Preferably, step S1 includes the following sub-steps: Step S101: Collect multi-source data throughout the entire life cycle of the transformer; The condition monitoring data includes winding leakage current, insulation resistance absorption ratio, trace water content in oil, insulating oil dielectric loss, furfural content in oil, relative total hydrocarbon gas production rate, total hydrocarbon volume fraction, hydrogen gas integral number, and acetylene volume fraction. The static data is extracted from the equipment file, and the static data includes rated capacity, voltage level, ambient temperature, and ambient humidity. The microscopic data includes high-frequency electromagnetic noise and the dynamic behavior of microbubbles in oil. The dynamic behavior of microbubbles in the oil includes microbubble images, bubble concentration, size distribution, rise rate, and coalescence frequency.

[0007] Preferably, step S1 further includes the following sub-steps: The preprocessing of the aforementioned status monitoring data and static data includes: The 3σ criterion was used to remove outliers from the state monitoring data and static data respectively. Linear interpolation was used to fill in the missing values ​​of the state monitoring data and static data respectively. Based on normalization, the state monitoring data and static data were mapped to the [0,1] interval respectively. Redundant variables with Pearson correlation coefficients greater than the first threshold were deleted. The preprocessing of the microscopic data includes: The high-frequency electromagnetic noise is denoised using wavelet thresholding, and the time-domain signal is converted into a time-frequency matrix using short-time Fourier transform. Gaussian filtering was used to remove image noise from the microbubble image, bubble regions were extracted by threshold segmentation, and morphological opening algorithm was used to remove tiny impurities.

[0008] Preferably, step S2 includes the following sub-steps: Step S201: Extract degradation features from the state monitoring data by improving the convolutional neural network; Step S202: Construct a similarity matrix for the preprocessed static data, standardize the Laplacian matrix, cluster the data into k′ clusters using K-means, and extract the cluster centers as static features; The degradation features and static features are considered as macroscopic features; Step S203 involves extracting the micro-features of the micro-data using a graph attention network, specifically including: Construct a graph structure, where the nodes of the graph structure are transformer components, the edges of the graph structure are the physical interaction relationships between components, and the micro data are used as node attributes in the graph structure; A graph attention network is used to calculate the attention weights of nodes and output micro-features. Step S204 involves evaluating the correlation between the microscopic and macroscopic features. The specific logic is as follows: The K-nearest neighbor estimation method is used to calculate the mutual information value between each micro-feature and each macro-feature, and the proportion of micro-features with mutual information values ​​greater than the second threshold is counted. Using the transformer aging stage as a label, and taking both micro and macro features as input, a random forest model is trained, and the average importance ratio of micro features is statistically analyzed. The transformer aging stages include the healthy stage, the early aging stage, and the critical aging stage. Step S205: Obtain the correlation degree between micro-features and macro-features, wherein the correlation degree includes a first correlation degree and a second correlation degree; If the proportion of micro-features with mutual information values ​​greater than the second threshold is less than or equal to the first percentage, and the average importance proportion of micro-features is less than or equal to the second percentage, then the correlation between micro-features and macro-features is the first correlation degree. If the proportion of micro-features with mutual information values ​​greater than the second threshold is greater than the first percentage, or the proportion of the average importance of micro-features is greater than the second percentage, then the correlation between micro-features and macro-features is the second correlation degree.

[0009] Preferably, step S201 includes the following sub-steps: Step S2011: Construct an improved convolutional neural network, the improved convolutional neural network including a dual-channel processing structure, the dual-channel processing structure including a first channel and a second channel; The first channel is a one-dimensional depthwise separable convolutional neural network, and the second channel is a two-dimensional transposed convolutional neural network; Step S2012: Perform channel convolution and point convolution operations through the first channel to process the original time-series signal in the multi-source data. The original time-series signal includes the sequence of winding leakage current changing with time and the time series of trace water content in oil. Output a time-series feature vector. Perform transpose convolution on the reconstructed matrix signal through the second channel to output a spatial feature vector. Step S2013: After merging the temporal feature vector and the spatial feature vector, a spatiotemporal feature vector is obtained. The spatiotemporal feature vector is then input into a deep residual shrinking network, and an identity mapping is performed through the residual blocks of the deep residual shrinking network. The noise in the spatiotemporal feature vector after identity mapping is filtered by a soft thresholding module, and the weights of the denoised spatiotemporal feature vector are assigned by an attention mechanism to output the preliminary degraded features. Step S2014 involves using the whale optimization algorithm to optimize and improve the hyperparameters of the convolutional neural network, outputting degenerate features, specifically including: With the goal of minimizing prediction error, the hyperparameter space is mapped to the position of individual whales. The parameters are adjusted by encirclement and the position is updated by spiral calculation. The time step, the number of neurons in the hidden layer and the learning rate are iteratively optimized. The iteration terminates when the preset maximum number of iterations is reached or the error converges.

[0010] Preferably, step S3 includes the following sub-steps: Step S301: The degradation index includes a first degradation index and a second degradation index. When the correlation degree is the first correlation degree, the first degradation index is constructed based on the degradation feature and the static feature. The formula for calculating the first degradation index is: ; in, As the primary indicator of degradation, A characteristic of degradation, The preset degradation feature weights, It is a static feature. These are the preset static feature weights. Preferably, step S3 further includes the following sub-steps: Step S302: Construct a degradation index based on the correlation degree for different scenarios. When the correlation degree is the second correlation degree, construct a second degradation index based on the degradation features, static features, and micro features. The formula for calculating the second degradation index is: ; in, As the second degradation indicator, The corrected weights for the degenerate features. A characteristic of degradation, It is a static feature. For the preset static feature weights, For the preset micro-feature weights, Microscopic features; Step S303: Perform performance verification on the monotonicity, robustness and correlation of the degradation index.

[0011] Preferably, the specific logic of the performance verification includes: When the monotonicity of the degradation index is less than the third threshold, the robustness is greater than the fourth threshold, and the correlation is less than the fifth threshold, it indicates that the performance of the degradation index is unqualified, and then return to step S204 to readjust the first and second percentages of the correlation determination.

[0012] Preferably, step S4 includes the following sub-steps: Step S401: When using the first degradation index, the sequence of the first degradation index within a unit time in the future is predicted by the pre-trained gated recurrent unit time series model. In the first degradation index sequence, the moment when the first degradation index is first satisfied and is greater than or equal to the preset failure threshold is found and defined as the first failure time. The time difference between the first failure time and the current time is the first remaining service life. Step S402: When using the second degradation index, the second degradation index and micro-features are input into the pre-trained support vector regression model, and the second remaining lifetime is output. If the micro-features fluctuate, the updated micro-features are re-inputted, and the support vector regression model outputs a new second remaining lifetime in real time.

[0013] A transformer life prediction system includes an acquisition module, an extraction module, a construction module, and a prediction module; The acquisition module is used to acquire multi-source data from the transformer; The extraction module is used to extract the macroscopic and microscopic features of the multi-source data, evaluate the correlation between the microscopic and macroscopic features, and obtain the correlation degree between the microscopic and macroscopic features. The construction module is used to construct degradation indicators based on the correlation degree for different scenarios; The prediction module is used to predict the remaining service life of the transformer based on the degradation index.

[0014] The beneficial effects of this invention are as follows: This invention achieves more accurate and earlier prediction of remaining lifetime through multi-source data fusion and intelligent algorithm innovation. Multi-source data acquisition takes into account both macroscopic operating status and microscopic physical signals, breaking through the limitations of traditional methods that rely solely on macroscopic data. It can identify early signs of aging in advance. The feature extraction stage combines improved convolutional neural networks, graph attention networks, and hyperparameter optimization to effectively capture spatiotemporal features and component interaction relationships, improving feature representation capabilities. Degradation indicators are constructed for different scenarios and rigorously verified to ensure that the indicators can adapt to the differences in feature contributions at different aging stages. In the prediction stage, gated recurrent units and support vector regression models are selected for different scenarios, balancing prediction accuracy and computational efficiency, and supporting real-time updates when microscopic features fluctuate. This significantly improves the accuracy, early warning capability, and robustness of remaining lifetime prediction. It can be adapted to transformers of different specifications, making it easy to integrate into existing monitoring systems. It provides a reliable basis for preventive maintenance of power equipment, reduces the risk of sudden failures, extends equipment service life, and saves operation and maintenance costs. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of a transformer life prediction method according to an embodiment of the present invention; Figure 2 This is a basic flowchart of a transformer life prediction system provided in one embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example 1, referring to Figure 1 A method for predicting transformer lifespan includes the following steps: Step S1: Collect multi-source data of the transformer, including condition monitoring data, static data and micro data.

[0018] Step S2: Extract degradation features from the state monitoring data, extract static features from the static data, extract micro features from the micro data, use the degradation features and static features as macro features, evaluate the correlation between the micro features and macro features, and obtain the correlation degree between the micro features and macro features. The correlation degree includes the first correlation degree and the second correlation degree.

[0019] Step S3: When the correlation degree is the first correlation degree, construct the first degradation index based on the degradation characteristics and static characteristics; when the correlation degree is the second correlation degree, construct the second degradation index based on the degradation characteristics, static characteristics and micro characteristics.

[0020] Step S4: When using the first degradation index, the first remaining lifetime is predicted by the pre-trained gated recurrent unit time series model. When using the second degradation index, the second degradation index and micro-features are input into the pre-trained support vector regression model to output the second remaining lifetime.

[0021] This invention achieves more accurate and earlier prediction of remaining lifetime through multi-source data fusion and intelligent algorithm innovation. Multi-source data acquisition considers both macroscopic operating conditions and microscopic physical signals, overcoming the limitations of traditional methods that rely solely on macroscopic data. This allows for early identification of early aging signs. The feature extraction stage combines improved convolutional neural networks, graph attention networks, and hyperparameter optimization to effectively capture spatiotemporal features and component interactions, enhancing feature representation capabilities. Degradation indicators are constructed for different scenarios and rigorously validated to ensure they adapt to the differences in feature contributions at different aging stages. In the prediction stage, gated recurrent units and support vector regression models are selected for different scenarios, balancing prediction accuracy and computational efficiency, and supporting real-time updates for microscopic feature fluctuations. The overall solution significantly improves the accuracy, early warning capability, and robustness of remaining lifetime prediction. It is adaptable to transformers of different specifications, easily integrated into existing monitoring systems, provides a reliable basis for preventative maintenance of power equipment, reduces the risk of sudden failures, extends equipment service life, and saves on operation and maintenance costs.

[0022] Step S1 includes the following sub-steps: Step S101: Collect multi-source data throughout the entire life cycle of the transformer.

[0023] Condition monitoring data is collected using standard industrial sensors. The condition monitoring data includes winding leakage current, insulation resistance absorption ratio, trace water content in oil, insulating oil dielectric loss, furfural content in oil, relative total hydrocarbon generation rate, total hydrocarbon volume fraction, hydrogen gas integral number, and acetylene volume fraction.

[0024] Extract static data from the equipment file. Static data includes rated capacity, voltage level, ambient temperature, and ambient humidity.

[0025] Microscopic data, including high-frequency electromagnetic noise and the dynamic behavior of microbubbles in oil, reveal early aging from a physical microscopic level.

[0026] High-frequency electromagnetic noise is collected by a high-frequency current transformer or radio frequency sensor at a preset sampling rate to capture transient signals.

[0027] The dynamic behavior of microbubbles in oil is acquired by ultrasonic imaging or microfluidic sensors. The dynamic behavior of microbubbles in oil includes microbubble images, bubble concentration, size distribution, rise rate and coalescence frequency, which can reveal the aging of oil paper insulation at an early stage.

[0028] Microscopic data, as a supplement, approaches the problem from a physical microscopic level, enabling early degradation to be detected before macroscopic parameters become abnormal, achieving breakthroughs in details. High-frequency noise can identify precursors to insulation deterioration, and microbubble movement reflects the trend of oil degradation.

[0029] It should be noted that a wideband high-frequency current transformer (HFCT) is used, and its magnetic core should be made of nanocrystalline or ultracrystalline material to ensure high permeability and low loss at high frequencies. The wideband high-frequency current transformer (HFCT) is connected to the transformer core grounding wire, clamp grounding wire, or neutral point grounding wire. The sensor's signal acquisition range is 3MHz to 30MHz (for partial discharge pulses) or higher. The sampling frequency is greater than 100MS / s to satisfy the Nyquist sampling law and fully preserve transient waveform characteristics. An automatic trigger logic is set to start recording when the signal amplitude is greater than three times the root mean square value of the background noise.

[0030] Employing an industrial-grade macro CCD / CMOS camera with a resolution greater than 2048×1536 pixels and a frame rate set between 100fps and 500fps, it captures high-speed moving bubbles. Equipped with a high-brightness LED array light source and a backlit illumination layout, it enhances the contrast of bubble edges. A high-voltage tempered glass viewing window is installed on the side wall of the transformer tank, or a microfluidic observation chamber is connected in parallel in the cooling oil circulation loop. A flow rate limiting valve reduces the oil flow velocity to below 0.1m / s, ensuring clear and stable imaging.

[0031] The equivalent volume and bubble concentration of each microbubble within the observation area were extracted using an image recognition algorithm. The mathematical expression is: ; in, This refers to the bubble concentration. This represents the total number of microbubbles extracted within the observation area. The first image extracted by the image recognition algorithm The equivalent volume of a microbubble This represents the total volume of the observation area.

[0032] The equivalent projected diameter method is used to calculate the pixel area of ​​each microbubble, and the pixel area is then converted into the physical diameter of the microbubble. The mathematical expression for the physical diameter of the microbubble is as follows: ; in, The physical diameter of the transformed microbubbles. For the extracted first The pixel area of ​​a microbubble.

[0033] Based on the calculated physical diameters of each microbubble, the probability density function of the physical diameter is statistically analyzed to obtain the size distribution of the bubbles. .

[0034] The centroid tracking algorithm is used to obtain the displacement vector of the same microbubble in two adjacent frames. The mathematical expression for the rate of rise is: ; in, For the rate of increase, and These are the acquisition times of two adjacent image frames, and satisfy the following conditions: The time difference between two frames. For the same microbubbles in Centroid coordinates in the time-lapse image For the same microbubbles in The centroid coordinates in the time-lapse image.

[0035] The merging of microbubbles is recorded within a set time period. The mathematical expression for the merging frequency is: ; in, The coalescence frequency of microbubbles The set observation time interval, In the observation time interval The number of times two or more microbubbles detected within the body merge into a single larger bubble. This represents the total number of microbubbles within the observation time interval.

[0036] Step S101 involves collecting multi-source data throughout the transformer's entire lifecycle, simultaneously covering both macroscopic and microscopic data. Condition monitoring data reflects real-time operating status, static data embodies inherent equipment attributes, high-frequency electromagnetic noise captures early signs of insulation degradation, and the dynamic behavior of microbubbles in the oil reflects oil degradation trends. This approach breaks through the limitations of traditional methods that rely solely on macroscopic data, enabling early detection of degradation before macroscopic parameters become abnormal, thus providing data support for early warning. Standard industrial sensors and professional microscopic acquisition equipment are used to ensure the professionalism and accuracy of data acquisition.

[0037] Step S102 involves preprocessing the multi-source data, specifically including: Preprocessing of condition monitoring data and static data includes: The 3σ criterion was used to remove outliers from the state monitoring data and static data respectively. Linear interpolation was used to fill in the missing values ​​of the state monitoring data and static data respectively. Based on normalization, the state monitoring data and static data were mapped to the [0,1] interval respectively. Redundant variables with Pearson correlation coefficients greater than the first threshold were deleted. Preprocessing of microscopic data includes: Wavelet thresholding is used to denoise high-frequency electromagnetic noise, and the time-domain signal is converted into a time-frequency matrix through short-time Fourier transform.

[0038] Gaussian filtering was used to remove image noise from the microbubble image, bubble regions were extracted by threshold segmentation, and morphological opening algorithm was used to remove tiny impurities.

[0039] Step S102 involves preprocessing the state monitoring data and static data to reduce interference from invalid data and improve data consistency. Microscopic data is denoised using wavelet thresholding, Gaussian filtering, and morphological operations to effectively filter environmental noise. The time-domain signal is converted into a time-frequency matrix, and the effective region of the bubble is extracted, providing high-quality data for subsequent feature extraction and avoiding feature distortion caused by noise.

[0040] Step S1 achieves comprehensive data coverage by collecting both macroscopic and microscopic data. It includes traditional operating status information as well as microscopic physical signals that reflect early aging. This allows for the capture of signs of degradation before macroscopic parameters become abnormal. The preprocessing stage significantly improves the signal-to-noise ratio of the data through anomaly removal, missing data filling, redundancy reduction, and noise reduction, laying a high-quality data foundation for subsequent feature extraction.

[0041] Step S2 includes the following sub-steps: Step S201: Extract degradation features from state monitoring data by improving the convolutional neural network.

[0042] Step S202: Construct a similarity matrix for the preprocessed static data, standardize the Laplacian matrix, and perform K-means clustering into k′ clusters. Extract the cluster centers as static features. Clustering captures the essential differences in static data, making the indicators adaptable to different transformer specifications. High-capacity transformers and low-capacity transformers are divided into different clusters.

[0043] Step S202: Static data clustering extracts static features. Through similarity matrix construction, Laplace matrix standardization, and K-means clustering, the essential differences in static data (rated capacity and voltage level) are captured (such as high / low capacity transformer clustering). The cluster centers are extracted as static features so that subsequent indicators can be adapted to transformers of different specifications, improve the model's universality, and avoid prediction bias caused by differences in equipment specifications.

[0044] Degradation features and static features are considered as macroscopic features.

[0045] Step S203 involves extracting micro-features from the micro-data using a graph attention network, specifically including: Construct a graph structure where nodes represent transformer components, edges represent physical interactions between components, and micro-data serves as node attributes.

[0046] Transformer components include windings, insulating oil, core, and bushings. The physical interactions between components include heat exchange between the windings and insulating oil, electromagnetic coupling between the windings and core, and sealing interaction between the insulating oil and bushings.

[0047] A graph attention network is used to calculate the attention weights of nodes and output micro-features. The graph attention network processes unstructured data, makes up for the shortcomings of traditional convolutional networks, and the attention weights provide interpretability and highlight the contribution of micro-details.

[0048] It should be noted that constructing the graph structure involves abstracting the internal solid components of the transformer into a topological mapping in a non-Euclidean space, specifically including: Define the set of nodes in a graph structure. ,in, Corresponding windings, Corresponding iron core, Corresponding insulating oil, Corresponding sleeve, The corresponding cooling system uses microscopic data (such as the microbubble feature vector of insulating oil and the high-frequency electromagnetic noise feature vector of the winding region) as the initial attribute matrix input for the corresponding node.

[0049] Define the edge set of the graph structure as ,node With nodes Edge weights between Characterizing the nonlinear correlation strength under multiphysics coupling. This embodiment employs a thermo-electromagnetic dual-field cooperative quantization mechanism, the mathematical expression of which is: ; in, and The first and second field contribution coefficients are preset according to the transformer's rated operating conditions, and satisfy the following conditions: , The normalized heat conduction interaction strength is determined by the transient heat transfer coefficient and effective contact area between the two components, and is used to characterize the thermodynamic evolution boundary of heat conduction from the winding or core to the insulating oil. The normalized electromagnetic coupling strength is quantified based on the leakage magnetic inductance or parasitic capacitance between the two components to characterize the spatial radiation and conduction losses of high-frequency electromagnetic noise between the components.

[0050] If there is no direct physical contact or electromagnetic coupling between the two components (such as the iron core and the cooling system), then the corresponding initial weights are... .

[0051] To capture deep, degenerate implicit features and avoid the oversmoothing effect of graph convolution, graph attention networks include an input layer, The layers are a hidden layer and an output layer, where, The hidden layer employs a multi-head attention mechanism, with the number of attention heads... The number of attention heads is set to 4 or 8, and each attention head independently calculates the attention distribution between nodes. The node feature dimension of the hidden layer is set to 64 dimensions, and LeakyReLU is used as a non-linear activation function (the slope of the negative half axis is set to 0.2) to retain weak aging feature signals. Finally, the output layer concatenates or averages the features of the multi-head attention and outputs a deep microscopic feature vector that integrates the topological relationships of multi-physics interaction.

[0052] Step S203: The graph attention network extracts micro-features by constructing a graph structure with transformer components as nodes and physical interactions as edges. Micro-data is used as node attributes to adapt to unstructured micro-data. Attention weights are used to highlight the contribution of micro-features of key components (such as insulating oil), enhance feature interpretability, make up for the inability of traditional convolutional networks to model component interactions, and accurately extract degradation information at the micro level.

[0053] Step S204 involves a dual-indicator correlation assessment of micro and macro characteristics. The specific logic is as follows: The K-nearest neighbor estimation method is used to calculate the mutual information value between each micro-feature and each macro-feature, and the proportion of micro-features with mutual information values ​​greater than the second threshold is counted.

[0054] Using the transformer aging stage as a label, a random forest model is trained by taking both micro and macro features as inputs and calculating the average importance ratio of micro features. The transformer aging stages include the healthy stage, the early aging stage, and the critical aging stage.

[0055] Step S204: The dual-indicator correlation assessment uses mutual information values ​​to quantify the nonlinear correlation between micro and macro features. The random forest model statistically analyzes the importance of micro features in the classification of aging stages. The combination of dual indicators avoids the one-sidedness of a single assessment, accurately identifies the actual contribution of micro features to macro features, provides a scientific basis for subsequent correlation judgment, and avoids redundant and ineffective micro features.

[0056] Step S205: Obtain the correlation between micro-features and macro-features. The correlation includes the first correlation and the second correlation.

[0057] If the proportion of micro-features with mutual information values ​​greater than the second threshold is less than or equal to the first percentage, and the average importance proportion of micro-features is less than or equal to the second percentage, then the correlation between micro-features and macro-features is the first correlation degree.

[0058] If the proportion of micro-features with mutual information values ​​greater than the second threshold is greater than the first percentage, or the proportion of the average importance of micro-features is greater than the second percentage, then the correlation between micro-features and macro-features is the second correlation degree.

[0059] Step S205: Determine the correlation between micro-features and macro-features by dividing the correlation into first and second degree of correlation based on the proportion of mutual information and the proportion of importance. Clarify the degree of influence of micro-features and provide a basis for constructing degradation indicators for subsequent scenarios. Avoid indicator redundancy or information loss caused by one-size-fits-all fusion (e.g., if the micro-feature has a small impact, there is no need to force fusion to reduce the amount of computation).

[0060] It should be noted that the correlation determination rule in step S205 is essentially a quantification of the phase transition boundary of the mapping from micro-evolution to macro-degeneration in the multi-source feature space.

[0061] The first percentage is a critical threshold characterizing the nonlinear coupling depth of mutual information entropy between microscopic and macroscopic features. In the actual transformer aging dynamics model, when the effective proportion of mutual information between high-frequency electromagnetic noise and microbubble features crosses this critical point, it indicates that local microscopic defects have begun to trigger systematic degradation of the global insulation medium. In this embodiment, the value range of the first percentage is set to 30% to 45%, preferably 35%. The determination of the first percentage is based on the statistical law of information entropy increment in the early stage of transformer insulating oil paper aging.

[0062] The second percentage represents the average contribution boundary of micro-feature nodes to the decrease in Gini impurity in a random forest ensemble manifold. Its physical significance lies in filtering topological noise in the context of high-frequency measurements. In this embodiment, the value range of the second percentage is set to 15% to 25%, preferably 20%. When the feature importance percentage exceeds the second percentage, it indicates that the micro-feature has achieved statistical significance in independently defining the aging stage.

[0063] The third threshold is the lower bound of monotonicity. This is used to constrain the directional consistency of the degradation index sequence on the time scale. To ensure the convergence of the gradient descent of the gated recurrent unit, the index is required to have an irreversible aging evolution trend. In this embodiment, the third threshold is set to 0.65 to 0.80, preferably 0.75. When the calculated monotonicity is less than... This indicates that the sequence exhibits violent oscillations and inversions, violating the laws of irreversible thermodynamic evolution.

[0064] The fourth threshold is the upper bound of robustness. This is used to define the maximum tolerance of the degenerate characteristic manifold to random walk noise. In this embodiment, the fourth threshold is set to 0.15 to 0.30, preferably 0.20. If the calculated ratio of the random component to the stationary component (robustness calculation value) is greater than... This indicates that the current indicator has been overwhelmed by transient high-frequency disturbances in the micro data, and its anti-interference ability has failed.

[0065] The fifth threshold is the lower bound of correlation. This is used to measure the degree of homomorphic mapping between the extracted degradation indicators and the actual physical service time of the equipment. In this embodiment, the fifth threshold is set to 0.70 to 0.85, preferably 0.80. A value less than the fifth threshold means that the constructed indicators are outside the actual aging timeline of the equipment and lack a physical benchmark for lifespan prediction.

[0066] Furthermore, when the degradation index fails to meet the aforementioned threshold constraint (i.e., performance is unqualified), the system will trigger an adaptive adjustment strategy and return to step S204. The specific operator logic for the adjustment is as follows: Based on the gradient descent principle, dynamic step size adjustments are applied to the first and second percentages. If the indicator shows insufficient correlation or monotonicity, the system will adjust the step size accordingly. The algorithm adaptively increases the set values ​​of the first and second percentages to more strictly eliminate weakly correlated micro-noise features. If the robustness exceeds the limit, the algorithm weights are adjusted until the topology of the degenerate index manifold is stabilized again within the set threshold boundary.

[0067] Step S201 includes the following sub-steps: Step S2011: Construct an improved convolutional neural network. The improved convolutional neural network includes a dual-channel processing structure, which includes a first channel and a second channel.

[0068] The first channel is a one-dimensional depthwise separable convolutional neural network, and the second channel is a two-dimensional transposed convolutional neural network.

[0069] Step S2011: Construct an improved convolutional neural network. The first channel is adapted to temporal signals, and the second channel is adapted to spatial signals. The dual-channel design takes into account both temporal and spatial feature extraction, avoiding the loss of key information in a single channel. The lightweight structure of depthwise separable convolution reduces the amount of computation, and transposed convolution can reconstruct spatial correlations, laying the network foundation for efficient extraction of multi-dimensional features.

[0070] Step S2012: Perform channel convolution and point convolution operations through the first channel to process the original time-series signals in the multi-source data. The original time-series signals include the sequence of winding leakage current changing with time and the time series of trace water content in oil. Output the time-series feature vector. Through the lightweight structure of channel convolution and point convolution, the feature information of the time-series dimension is preserved while reducing the amount of computation.

[0071] The second channel performs transpose convolution on the reconstructed matrix signal to process the original spatial signal in the multi-source data and output spatial feature vectors. First, the original time series signal of the multi-source data is reconstructed into a two-dimensional matrix according to preset rules. Then, the single time series is transformed into a multi-column matrix by combining the dimensions of the time window and the monitoring indicators. Finally, the spatial correlation features of the multi-source data are extracted through transpose convolution.

[0072] Step S2012: Dual-channel processing of timing and spatial signals. The first channel processes the winding leakage current timing signal through channel convolution and point convolution, compressing the computational load while fully preserving the timing dynamic trend. The second channel reconstructs the timing signal into a two-dimensional matrix and then performs transposed convolution to extract spatial correlation features between multiple monitoring indicators, achieving collaborative extraction of timing-spatial features and improving feature comprehensiveness.

[0073] Step S2013: After merging the temporal feature vector and the spatial feature vector, a spatiotemporal feature vector is obtained. The spatiotemporal feature vector is then input into a deep residual shrinking network, and an identity mapping is performed through the residual blocks of the deep residual shrinking network.

[0074] The noise in the spatiotemporal feature vector after identity mapping is filtered by a soft thresholding module, and the weights of the denoised spatiotemporal feature vector are assigned by an attention mechanism to output the preliminary degraded features.

[0075] Step S2013: The deep residual shrinking network optimizes the feature residual blocks and performs identity mapping to avoid gradient vanishing in the deep network. The soft thresholding module filters residual noise in the spatiotemporal features. The attention mechanism assigns weights to the denoised features (such as highlighting the key feature of acetylene content in oil), effectively improving feature purity and discriminability and reducing the interference of invalid features on subsequent modeling.

[0076] Step S2014 involves using the whale optimization algorithm to optimize and improve the hyperparameters of the convolutional neural network, outputting the final degenerate features, specifically including: With the goal of minimizing prediction error, the hyperparameter space is mapped to the position of individual whales. The parameters are adjusted by encirclement and the position is updated by spiral calculation. The time step, the number of neurons in the hidden layer and the learning rate are iteratively optimized. The iteration terminates when the preset maximum number of iterations is reached or the error converges.

[0077] Step S2014: The whale optimization algorithm optimizes hyperparameters with the goal of minimizing prediction error. It maps hyperparameters (time step, number of neurons, and learning rate) to individual whale positions and achieves iterative optimization through encirclement and spiral updates. This avoids the subjectivity of setting hyperparameters based on human experience, quickly finds the optimal parameter configuration, and improves the accuracy of network feature extraction and training efficiency.

[0078] It should be noted that, in a preferred embodiment of the present invention, the improved convolutional neural network is used to extract degradation features from the original one-dimensional device vibration signal. To enable those skilled in the art to accurately implement the present invention, the specific structural parameters, connection relationships, and data flow of each sub-network (including a one-dimensional depthwise separable convolutional network, an attention mechanism module, a depthwise residual shrinkage network, and a two-dimensional transposed convolutional neural network) in the improved convolutional neural network are described in detail below: The improved convolutional neural network connects the various sub-networks in series. The original one-dimensional device vibration signal is first input to a one-dimensional depthwise separable convolutional network for preliminary feature extraction and dimensionality reduction. The preliminary feature map output by the one-dimensional depthwise separable convolutional network is directly input to the attention mechanism module for feature channel weight redistribution. The feature map after weighting by the attention mechanism module is then input to a depth residual shrinking network (with an embedded soft thresholding module) for deep nonlinear feature extraction and adaptive noise reduction. The one-dimensional depth feature vector output by the depth residual shrinking network and dimensionality reduction is reshaped and input to a two-dimensional transposed convolutional neural network for upsampling and feature mapping. Finally, the two-dimensional transposed convolutional neural network outputs a two-dimensional degenerate feature matrix.

[0079] A one-dimensional depthwise separable convolutional network receives a single-channel raw one-dimensional device vibration signal with a length of 1024 (signal input size is 1 channel, length 1024). The one-dimensional depthwise separable convolutional network consists of a channel-wise convolutional layer, a pointwise convolutional layer, and a ReLU activation layer.

[0080] The kernel size of the channel-wise convolutional layer is set to 31, the stride to 2, the padding to 15, the number of input channels to 1, and the number of output channels to 16.

[0081] The kernel size of the pointwise convolutional layer is set to 1, the stride is set to 1, the number of input channels is 16, and the number of output channels is expanded to 32.

[0082] After processing by a one-dimensional depthwise separable convolutional network, a preliminary feature map with 32 channels and a length of 512 is output.

[0083] The attention mechanism module employs a one-dimensional SE structure. It receives a preliminary feature map with 32 channels and a length of 512, output from the one-dimensional depthwise separable convolutional network. A global average pooling layer is used to compress the preliminary feature map with 32 channels and a length of 512 in spatial dimension, generating a channel descriptor with 32 channels and a length of 1. A first fully connected layer reduces the number of channels in the channel descriptor from 32 to 8 and connects it to a ReLU activation function. A second fully connected layer restores the number of channels to 32 and connects it to a Sigmoid activation function, generating weight coefficients for each channel. The values ​​of these weight coefficients range from 0 to 1. The generated weight coefficients are then multiplied channel-wise with the preliminary feature map received by the attention mechanism module, outputting a weighted feature map with the same size (32 channels, 512 length).

[0084] The deep residual shrinking network and soft thresholding module: The deep residual shrinking network receives weighted feature maps output by the attention mechanism module. The deep residual shrinking network consists of three consecutively cascaded residual shrinking blocks. The backbone of each residual shrinking block contains a first 1D standard convolutional layer and a second 1D standard convolutional layer. The kernel size of the first 1D standard convolutional layer is 3, and the number of channels is increased from 32 to 64. Only in the first residual shrinking block is the stride of the first 1D standard convolutional layer set to 2 for downsampling; in the second and third residual shrinking blocks, the stride is set to 1. The kernel size of the second 1D standard convolutional layer is 3, the stride is 1, and the number of channels remains at 64. Both 1D standard convolutional layers are followed by a batch normalization layer and a ReLU activation function. This is embedded after the second 1D standard convolutional layer of each residual shrinking block. The soft thresholding module first performs global absolute pooling on the input features, then calculates the scaling factor α using a two-layer fully connected network (first dimensionality reduction, then dimensionality increase) and a sigmoid activation function. The dynamic threshold τ is obtained by multiplying the scaling factor α by the average absolute value of the feature maps in the current layer. Features are then processed by the soft thresholding function. After processing, the feature size is added to and fused with the shortcut branch of the corresponding residual shrinking block. After processing through the above three residual shrinking blocks, the feature size is reduced to 64 channels and 256 in length. Then, a global average pooling layer is passed to output a one-dimensional deep degradation feature vector with 64 channels and a length of 1.

[0085] A 2D transposed convolutional neural network is used to map the aforementioned one-dimensional deep degradation feature vector into an interpretable 2D degradation feature map in a high-dimensional space. First, the aforementioned one-dimensional deep degradation feature vector with 64 channels and a length of 1 is reshaped into a 2D feature map. The reshaped 2D feature map has 1 channel, 8 heights, and 8 widths (represented as 1×8×8).

[0086] The first transposed convolutional layer receives the reshaped two-dimensional feature map. The kernel size of the first transposed convolutional layer is 3×3, the stride is 2, the padding is 1, and the number of output channels is 16. It is connected to the batch normalization layer and the ReLU activation function. The output feature map size is enlarged to 16 channels, 16 heights, and 16 widths (i.e., 16×16×16).

[0087] The second transposed convolutional layer has a kernel size of 3×3, a stride of 2, padding of 1, and 8 output channels. It connects to a batch normalization layer and a ReLU activation function, and the output feature map is enlarged to 8 channels, 32 in height, and 32 in width (i.e., 8×32×32).

[0088] The output convolutional layer uses a standard two-dimensional convolution with a kernel size of 1×1, a stride of 1, and 1 output channel. It is connected to a Sigmoid activation function to normalize the output value to the (0,1) interval.

[0089] The two-dimensional transposed convolutional neural network outputs a two-dimensional degenerate feature matrix with 1 channel, a height of 32, and a width of 32 (i.e., 1×32×32).

[0090] Through the above network structure design, hyperparameter configuration, and data flow between layers, strong background noise in equipment vibration signals can be effectively filtered out, and two-dimensional degradation features that can accurately reflect the trend of equipment performance degradation can be extracted.

[0091] Step S2 employs an improved convolutional neural network, using a dual-channel structure and a deep residual shrinking network to extract macroscopic degradation features. This preserves the dynamic trends of time-series signals and captures the spatial correlations of multi-source data. The whale optimization algorithm is combined to optimize hyperparameters, improving the accuracy and efficiency of feature extraction. Graph attention networks are used to process microscopic data, modeling the physical interactions between transformer components, highlighting the contributions of key microscopic features, and enhancing feature interpretability. The correlation between microscopic features and macroscopic features is quantified through a dual-index evaluation of mutual information and random forest importance, providing a scientific basis for constructing degradation indices for different scenarios and avoiding interference from invalid features.

[0092] Step S3 includes the following sub-steps: Step S301: The degradation index includes a first degradation index and a second degradation index. When the correlation degree is the first correlation degree, the first degradation index is constructed based on the degradation characteristics and static characteristics.

[0093] The formula for calculating the first degradation index is: ; in, As the primary indicator of degradation, A characteristic of degradation, The preset degradation feature weights, It is a static feature. These are the preset static feature weights.

[0094] Step S301: Construct the first degradation index by integrating degradation features and static features to avoid redundant and ineffective features when the micro-level impact is small. This simplifies the calculation while ensuring that the index can reflect the macro-level degradation trend. By presetting weights, the degradation feature weights highlight the impact of the operating status, while the static feature weights reflect the differences in equipment specifications. This allows the index to be adapted to scenarios with small micro-level contributions, ensuring the accuracy of basic predictions.

[0095] Step S3 also includes the following sub-steps: Step S302: When the correlation degree is the second correlation degree, construct the second degradation index based on the degradation characteristics, static characteristics and micro characteristics.

[0096] The formula for calculating the second degradation index is: ; in, As the second degradation indicator, The corrected weights for the degenerate features. A characteristic of degradation, It is a static feature. For the preset static feature weights, For the preset micro-feature weights, These are microscopic features.

[0097] Step S302: Construct a second degradation index that incorporates micro-features on the basis of macro-features. By modifying the weights of degradation features and adding weights of micro-features, the contribution of micro-features to degradation is quantified (e.g., increasing the weight of micro-features when the concentration of microbubbles is high). This enables the index to fully reflect the synergistic degradation process of macro and micro-features and improves the index's characterization ability in complex aging scenarios.

[0098] Step S303: Perform performance verification on the monotonicity, robustness and correlation of the degradation index.

[0099] The specific logic for performance verification includes: The degradation index series is analyzed using the moving average method. Decomposed into stationary components With random components The formula for calculating the monotonicity of the degradation index is:

[0100] in, For the monotonicity of degradation indicators, For unit out-of-order functions, For the first The stationary components of the degradation index sequence at each monitoring time point For the first The stable component of the degradation index at each monitoring time. Let be the length of the degradation index sequence.

[0101] Monotonicity is used to assess the consistency of the trend direction of a sequence; the closer the monotonicity is to 1, the more stable the degradation trend.

[0102] The formula for calculating the robustness of the degradation index is: ; in, For the robustness of degradation indicators, Let be the length of the degradation index sequence. For the vth monitoring time, For the first The stationary components of the degradation index sequence at each monitoring time point For the random components of the degradation index sequence.

[0103] Robustness is used to measure the ability to withstand fluctuations by analyzing the magnitude of changes in robustness under noise. Higher robustness indicates stronger resistance to interference.

[0104] The formula for calculating the correlation of degradation indicators is: ; in, The correlation of degradation indicators, For the vth monitoring time, For the first The stationary components of the degradation index sequence at each monitoring time point For the first The stable component of the degradation index at each monitoring time. Let be the length of the degradation index sequence.

[0105] When the monotonicity of the degradation index is less than the third threshold, the robustness is greater than the fourth threshold, and the correlation is less than the fifth threshold, it indicates that the performance of the degradation index is unqualified. Then, return to step S204 to readjust the first and second percentages of the correlation determination.

[0106] Step S303: The performance verification of degradation indicators is carried out through three-dimensional verification of monotonicity (assessing the stability of degradation trend), robustness (assessing the ability to resist interference), and correlation (assessing the correlation with aging time) to ensure that the indicators have a stable degradation trend, strong resistance to fluctuations, and high correlation with aging. If the indicators are not qualified, the correlation judgment rules are adjusted in reverse to form a closed loop of indicator verification and rule optimization, further improving the reliability of the indicators and avoiding prediction bias caused by unqualified indicators.

[0107] Step S3: Based on the correlation of micro-features, construct degradation indicators for different scenarios. The first indicator includes macro-features and static features, while the second indicator additionally incorporates micro-features. This allows the indicators to adapt to the differences in feature contributions at different aging stages. Through monotonicity, robustness, and correlation verification, ensure that the degradation indicators have a stable degradation trend, strong anti-interference ability, and high correlation with equipment aging. Eliminate unqualified indicators and optimize the correlation judgment rules in reverse to further improve the reliability of the indicators.

[0108] Step S4 includes the following sub-steps: Step S401: When using the first degradation index, the sequence of the first degradation index within a unit time in the future is predicted by the pre-trained gated recurrent unit time series model. In the first degradation index sequence, the moment when the first degradation index is first satisfied and is greater than or equal to the preset failure threshold is found and defined as the first failure time. The time difference between the first failure time and the current time is the first remaining service life.

[0109] Step S401: The first remaining useful life is predicted by the gated loop unit. The gated loop unit is adapted to the temporal characteristics of the first degradation index and can accurately capture the degradation trend of the index over time. By predicting the future index sequence and locating the time when the failure threshold is first reached, the remaining useful life is directly output. The calculation logic is clear and adapted to the temporal data, ensuring the prediction accuracy in the first correlation scenario.

[0110] Step S402: When using the second degradation index, input the second degradation index and micro-features into the pre-trained support vector regression model and output the second remaining lifetime. If the micro-features fluctuate, simply re-input the updated micro-features, and the support vector regression model can output the new second remaining lifetime in real time.

[0111] Step S402: Predict the second remaining lifetime using a support vector regression model. The support vector regression model is good at handling the nonlinear correlation between macroscopic and microscopic features, adapts to the complex feature structure of the second degradation index, and can output the new lifetime in real time by simply re-inputting updated data when microscopic features fluctuate, without the need to retrain the model. It responds quickly and meets the real-time prediction requirements in dynamic aging scenarios.

[0112] Step S4: For the first degradation index, a gated cyclic unit model is adopted to adapt to the degradation trend prediction of time series data and accurately locate the failure time. For the second degradation index, a support vector regression model is adopted to adapt to the nonlinear correlation between macro and micro features. The prediction results can be updated in real time when micro features fluctuate, and the response is rapid. Adaptive models are selected according to different scenarios to simplify calculations while ensuring prediction accuracy and facilitate engineering implementation.

[0113] Example 2, refer to Figure 2 A transformer life prediction system is provided, including a data acquisition module, an extraction module, a construction module, and a prediction module.

[0114] The acquisition module is used to acquire multi-source data from the transformer.

[0115] The extraction module is used to extract macro and micro features from multi-source data, evaluate the correlation between micro and macro features, and obtain the correlation degree between micro and macro features.

[0116] The module is used to build degradation indicators based on the degree of correlation in different scenarios.

[0117] The prediction module is used to predict the remaining service life of the transformer based on degradation indicators.

[0118] This invention utilizes a multi-source data acquisition system, encompassing both macroscopic and microscopic data, to capture early degradation signals before macroscopic parameter anomalies occur, providing advance warning of aging. The feature extraction stage employs an improved convolutional neural network and graph attention network, combined with a whale optimization algorithm to optimize hyperparameters. This efficiently extracts spatiotemporal features from macroscopic data while accurately capturing microscopic features arising from transformer component interactions, enhancing feature representation capabilities. Degradation indicators are constructed for different scenarios, and their reliability is ensured through monotonicity, robustness, and correlation verification, avoiding interference from invalid features. In the prediction stage, gated recurrent units and support vector regression models are used for different scenarios. These models are simple to implement and respond quickly, updating prediction results in real time when microscopic features fluctuate, significantly improving the accuracy of remaining life prediction. This provides maintenance personnel with precise maintenance guidance, effectively reducing the risk of sudden transformer failures, extending equipment service life, and saving maintenance costs.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the lifespan of a transformer, characterized in that, Includes the following steps: Step S1: Collect multi-source data of the transformer, including condition monitoring data, static data and micro data; Step S1 includes the following sub-steps: Step S101: Collect multi-source data throughout the entire life cycle of the transformer; The condition monitoring data includes winding leakage current, insulation resistance absorption ratio, trace water content in oil, insulating oil dielectric loss, furfural content in oil, relative total hydrocarbon gas production rate, total hydrocarbon volume fraction, hydrogen gas integral number, and acetylene volume fraction. The static data is extracted from the equipment file, and the static data includes rated capacity, voltage level, ambient temperature, and ambient humidity. The microscopic data includes high-frequency electromagnetic noise and the dynamic behavior of microbubbles in oil. The dynamic behavior of microbubbles in the oil includes microbubble images, bubble concentration, size distribution, rise rate, and coalescence frequency; Step S2: Extract degradation features from the state monitoring data, extract static features from the static data, extract micro features from the micro data, use the degradation features and static features as macro features, evaluate the correlation between the micro features and macro features, and obtain the correlation degree between the micro features and macro features. The correlation degree includes a first correlation degree and a second correlation degree. Step S2 includes the following sub-steps: Step S201: Extract degradation features from the state monitoring data by improving the convolutional neural network; Step S202: Construct a similarity matrix for the preprocessed static data, standardize the Laplacian matrix, cluster the data into k′ clusters using K-means, and extract the cluster centers as static features; The degradation features and static features are considered as macroscopic features; Step S203 involves extracting the micro-features of the micro-data using a graph attention network, specifically including: Construct a graph structure, where the nodes of the graph structure are transformer components, the edges of the graph structure are the physical interaction relationships between components, and the micro data are used as node attributes in the graph structure; A graph attention network is used to calculate the attention weights of nodes and output micro-features. Step S204 involves evaluating the correlation between the microscopic and macroscopic features. The specific logic is as follows: The K-nearest neighbor estimation method is used to calculate the mutual information value between each micro-feature and each macro-feature, and the proportion of micro-features with mutual information values ​​greater than the second threshold is counted. Using the transformer aging stage as a label, and taking both micro and macro features as input, a random forest model is trained, and the average importance ratio of micro features is statistically analyzed. The transformer aging stages include the healthy stage, the early aging stage, and the critical aging stage. Step S205: Obtain the correlation degree between micro-features and macro-features, wherein the correlation degree includes a first correlation degree and a second correlation degree; If the proportion of micro-features with mutual information values ​​greater than the second threshold is less than or equal to the first percentage, and the average importance proportion of micro-features is less than or equal to the second percentage, then the correlation between micro-features and macro-features is the first correlation degree. If the proportion of micro-features with mutual information values ​​greater than the second threshold is greater than the first percentage, or the proportion of the average importance of micro-features is greater than the second percentage, then the correlation between micro-features and macro-features is the second correlation degree. Step S3: When the correlation degree is the first correlation degree, a first degradation index is constructed based on degradation features and static features; when the correlation degree is the second correlation degree, a second degradation index is constructed based on degradation features, static features and micro features. Step S4: When using the first degradation index, the first remaining lifetime is predicted by a pre-trained gated recurrent unit time series model; when using the second degradation index, the second degradation index and micro-features are input into a pre-trained support vector regression model to output the second remaining lifetime. Step S4 includes the following sub-steps: Step S401: When using the first degradation index, the sequence of the first degradation index within a unit time in the future is predicted by the pre-trained gated recurrent unit time series model. In the first degradation index sequence, the moment when the first degradation index is first satisfied and is greater than or equal to the preset failure threshold is found and defined as the first failure time. The time difference between the first failure time and the current time is the first remaining service life. Step S402: When using the second degradation index, the second degradation index and micro-features are input into the pre-trained support vector regression model, and the second remaining lifetime is output. If the micro-features fluctuate, the updated micro-features are re-inputted, and the support vector regression model outputs a new second remaining lifetime in real time.

2. The transformer life prediction method as described in claim 1, characterized in that, Step S1 further includes the following sub-steps: The preprocessing of the aforementioned status monitoring data and static data includes: The 3σ criterion was used to remove outliers from the state monitoring data and static data respectively. Linear interpolation was used to fill in the missing values ​​of the state monitoring data and static data respectively. Based on normalization, the state monitoring data and static data were mapped to the [0,1] interval respectively. Redundant variables with Pearson correlation coefficients greater than the first threshold were deleted. The preprocessing of the microscopic data includes: The high-frequency electromagnetic noise is denoised using wavelet thresholding, and the time-domain signal is converted into a time-frequency matrix using short-time Fourier transform. Gaussian filtering was used to remove image noise from the microbubble image, bubble regions were extracted by threshold segmentation, and morphological opening algorithm was used to remove tiny impurities.

3. The transformer life prediction method as described in claim 2, characterized in that, Step S201 includes the following sub-steps: Step S2011: Construct an improved convolutional neural network, the improved convolutional neural network including a dual-channel processing structure, the dual-channel processing structure including a first channel and a second channel; The first channel is a one-dimensional depthwise separable convolutional neural network, and the second channel is a two-dimensional transposed convolutional neural network; Step S2012: Perform channel convolution and point convolution operations through the first channel to process the original time-series signal in the multi-source data. The original time-series signal includes the sequence of winding leakage current changing with time and the time series of trace water content in oil. Output a time-series feature vector. Perform transpose convolution on the reconstructed matrix signal through the second channel to output a spatial feature vector. Step S2013: After merging the temporal feature vector and the spatial feature vector, a spatiotemporal feature vector is obtained. The spatiotemporal feature vector is then input into a deep residual shrinking network, and an identity mapping is performed through the residual blocks of the deep residual shrinking network. The noise in the spatiotemporal feature vector after identity mapping is filtered by a soft thresholding module, and the weights of the denoised spatiotemporal feature vector are assigned by an attention mechanism to output the preliminary degraded features. Step S2014 involves using the whale optimization algorithm to optimize and improve the hyperparameters of the convolutional neural network, outputting degenerate features, specifically including: With the goal of minimizing prediction error, the hyperparameter space is mapped to the position of individual whales. The parameters are adjusted by encirclement and the position is updated by spiral calculation. The time step, the number of neurons in the hidden layer and the learning rate are iteratively optimized. The iteration terminates when the preset maximum number of iterations is reached or the error converges.

4. The transformer life prediction method as described in claim 3, characterized in that, Step S3 includes the following sub-steps: Step S301: Construct a degradation index based on the correlation degree for different scenarios. When the correlation degree is the first correlation degree, construct a first degradation index based on the degradation features and static features. The formula for calculating the first degradation index is: ； in, As the primary indicator of degradation, A characteristic of degradation, The preset degradation feature weights, It is a static feature. These are the preset static feature weights.

5. The transformer life prediction method as described in claim 4, characterized in that, Step S3 further includes the following sub-steps: Step S302: When the correlation degree is the second correlation degree, a second degradation index is constructed based on the degradation feature, static feature and micro feature; The formula for calculating the second degradation index is: ； in, As the second degradation indicator, The corrected weights for the degenerate features. A characteristic of degradation, It is a static feature. For the preset static feature weights, For the preset micro-feature weights, Microscopic features; Step S303: Perform performance verification on the monotonicity, robustness and correlation of the degradation index.

6. The transformer life prediction method as described in claim 5, characterized in that, The specific logic for the performance verification includes: When the monotonicity of the degradation index is less than the third threshold, the robustness is greater than the fourth threshold, and the correlation is less than the fifth threshold, it indicates that the performance of the degradation index is unqualified, and then return to step S204 to readjust the first and second percentages of the correlation determination.

7. A transformer life prediction system, applied in a transformer life prediction method as described in any one of claims 1-6, characterized in that, It includes a data acquisition module, an extraction module, a construction module, and a prediction module; The acquisition module is used to acquire multi-source data from the transformer; The extraction module is used to extract the macroscopic and microscopic features of the multi-source data, evaluate the correlation between the microscopic and macroscopic features, and obtain the correlation degree between the microscopic and macroscopic features. The construction module is used to construct degradation indicators based on the correlation degree for different scenarios; The prediction module is used to predict the remaining service life of the transformer based on the degradation index.

Images