Transformer sample library extension method and device based on digital twinning and storage medium

Abstract

This application discloses a method, apparatus, and storage medium for expanding a transformer sample library based on digital twins. The method includes: acquiring multiple sets of target operating data for a target transformer; determining multiple initial features based on the multiple sets of target operating data, wherein the multiple initial features include initial first features of the actual operating data and initial second features of the virtual operating data, with each set of target operating data corresponding one-to-one with the multiple initial features; determining a first data distribution difference between the actual operating data and the virtual operating data based on the multiple initial features; if the first data distribution difference is greater than a preset difference threshold, filtering the multiple initial features to obtain multiple filtered features; performing drift detection based on the multiple filtered features to obtain multiple target features; and expanding the sample library based on the multiple target features. This application solves the technical problems of sample scarcity and insufficient sample coverage in transformer sample libraries existing in related technologies.

Description

Technical Field

[0001] This application relates to the field of digital data processing, and more specifically, to a method, apparatus, and storage medium for expanding a transformer sample library based on digital twins. Background Technology

[0002] With the advancement of smart grid construction, data-driven digital twins and state-aware models are increasingly being used in intelligent operation and maintenance of transformers. A high-quality, large-scale sample library with good coverage is the core infrastructure supporting model training, evaluation and continuous evolution.

[0003] The relevant technologies mainly employ manual field testing and expand the sample library using traditional physical simulation software. However, these methods have significant shortcomings: First, manual field testing is costly, time-consuming, and carries safety risks; second, traditional physical simulations are often based on idealized assumptions, ignoring noise interference, sensor errors, and complex environmental coupling factors in actual operation, resulting in unreasonable virtual operating data. Directly incorporating this low-fidelity virtual operating data into the sample library not only fails to effectively supplement scarce samples but may also introduce noise interference, leading to biases in the state-aware model training and even reducing the diagnostic accuracy of the state-aware model. Therefore, the relevant technologies suffer from the technical problems of scarce transformer sample libraries and insufficient sample coverage.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a method, apparatus, and storage medium for expanding a transformer sample library based on digital twins, in order to at least solve the technical problems of sample scarcity and insufficient sample coverage in related technologies.

[0006] According to one aspect of the embodiments of this application, a method for expanding a transformer sample library based on digital twins is provided, comprising: acquiring multiple sets of target operating data of a target transformer, wherein the multiple sets of target operating data include real operating data and virtual operating data, the real operating data being obtained by data collection from the target transformer, and the virtual operating data being obtained based on a digital twin of the target transformer; determining multiple initial features based on the multiple sets of target operating data, wherein the multiple initial features include an initial first feature of the real operating data and an initial second feature of the virtual operating data, and the multiple sets of target operating data correspond one-to-one with the multiple initial features; determining a first data distribution difference between the real operating data and the virtual operating data based on the multiple initial features; if the first data distribution difference is greater than a preset difference threshold, filtering the multiple initial features to obtain multiple filtered features; performing drift detection based on the multiple filtered features to obtain multiple target features; and expanding the sample library based on the multiple target features.

[0007] Optionally, based on multiple sets of target operation data, multiple initial features are determined, including: for any set of target operation data, where the set includes vibration data, acoustic data, infrared thermographic data, and oil and gas chromatography data, a one-dimensional depthwise separable convolutional module of the feature extraction model is used to extract time-domain and frequency-domain features from the vibration and acoustic data to obtain initial vibration features and initial acoustic features; a two-dimensional depthwise separable convolutional module of the feature extraction model is used to extract spatial features from the infrared thermographic data to obtain initial infrared features; a fully connected layer of the feature extraction model is used to encode the oil and gas chromatography data to obtain initial oil and gas features; a multi-head cross-attention module of the feature extraction model is used to fuse the initial vibration features, initial acoustic features, initial infrared features, and initial oil and gas features to obtain the initial features of any set of target operation data; and multiple initial features are obtained by determining the initial features of any set of target operation data.

[0008] Optionally, a first data distribution difference between real operating data and virtual operating data is determined based on multiple initial features, including: determining the real data distribution based on an initial first feature; determining the virtual data distribution based on an initial second feature; and determining the first data distribution difference based on the real data distribution and the virtual data distribution.

[0009] Optionally, multiple initial features are filtered to obtain multiple selected features, including: determining the uncertainty index values ​​corresponding to each of the multiple initial features; determining the spatial representativeness index values ​​corresponding to each of the multiple initial features; determining the comprehensive score of the initial first feature based on the uncertainty index values ​​of the initial first feature among the uncertainty index values ​​corresponding to each of the multiple initial features, and the spatial representativeness index value of the initial first feature among the spatial representativeness index values ​​corresponding to each of the multiple initial features; determining the comprehensive score of the initial second feature based on the uncertainty index values ​​of the initial second feature among the uncertainty index values ​​corresponding to each of the multiple initial features, the spatial representativeness index value of the initial second feature among the spatial representativeness index values ​​corresponding to each of the multiple initial features, and the transferability weight of the initial second feature, wherein the transferability weight is used to characterize the marginal contribution of the virtual operating data corresponding to the initial second feature to improving the performance of the state-aware model, and the state-aware model is used to determine the operating state of the target transformer; selecting a preset number of initial features based on the comprehensive scores of the initial first feature and the initial second feature in descending order to obtain multiple candidate features; and labeling the multiple candidate features to obtain multiple selected features.

[0010] Optionally, drift detection is performed based on multiple screening features to obtain multiple target features, including: obtaining a comprehensive drift degree based on multiple screening features; determining that no data drift has occurred if the comprehensive drift degree is less than or equal to a preset drift threshold, and identifying the multiple screening features as multiple target features; or, determining that data drift has occurred if the comprehensive drift degree is greater than the preset drift threshold, and correcting the multiple screening features respectively to obtain multiple target features.

[0011] Optionally, the comprehensive drift degree is obtained based on multiple screening features, including: determining a first drift degree based on the model performance data of the state-aware model in the first time period; dividing the multiple screening features into multiple first screening features, multiple second screening features, and multiple third screening features using a sliding window method, wherein the first time window corresponding to the multiple first screening features is smaller than the second time window corresponding to the multiple second screening features, and the second time window is smaller than the third time window corresponding to the multiple third screening features; determining a second drift degree based on the multiple first screening features, multiple second screening features, multiple third screening features, and multiple historical features in the sample library; dividing the multiple screening features into multiple screening feature subsets according to a preset time window; determining the outlier density corresponding to each of the multiple screening feature subsets, wherein the outlier density is used to characterize the degree of data distribution anomaly of the corresponding screening feature subset; determining a third drift degree based on the outlier density corresponding to each of the multiple screening feature subsets; and determining the comprehensive drift degree based on the first drift degree, the second drift degree, and the third drift degree.

[0012] Optionally, the second degree of drift is determined based on multiple first screening features, multiple second screening features, multiple third screening features, and multiple historical features in the sample library, including: determining the first screening data distribution of multiple first screening features, the second screening data distribution of multiple second screening features, the third screening data distribution of multiple third screening features, and the historical data distribution of multiple historical features; determining the second data distribution difference based on the first screening data distribution and the historical data distribution; determining the third data distribution difference based on the second screening data distribution and the historical data distribution; determining the fourth data distribution difference based on the third screening data distribution and the historical data distribution; and determining the second degree of drift based on the second data distribution difference, the third data distribution difference, and the fourth data distribution difference.

[0013] Optionally, multiple screening features are modified to obtain multiple target features, including: determining the target source causing data drift based on the first drift degree, the second drift degree, and the third drift degree; and modifying multiple screening features based on the comprehensive drift degree and the target source to obtain multiple target features.

[0014] According to another aspect of the embodiments of this application, a transformer sample library expansion device based on digital twins is provided, comprising: a data acquisition module, configured to acquire multiple sets of target operating data of a target transformer, wherein the multiple sets of target operating data include real operating data and virtual operating data, the real operating data being obtained by data collection from the target transformer, and the virtual operating data being obtained based on a digital twin of the target transformer; a first determination module, configured to determine multiple initial features based on the multiple sets of target operating data, wherein the multiple initial features include an initial first feature of the real operating data and an initial second feature of the virtual operating data, and the multiple sets of target operating data correspond one-to-one with the multiple initial features; a second determination module, configured to determine a first data distribution difference between the real operating data and the virtual operating data based on the multiple initial features; a filtering module, configured to filter the multiple initial features to obtain multiple filtered features when the first data distribution difference is greater than a preset difference threshold; a drift detection module, configured to perform drift detection based on the multiple filtered features to obtain multiple target features; and an expansion module, configured to expand the sample library based on the multiple target features.

[0015] Optionally, the first determining module includes: a first feature extraction unit, used to extract time-domain and frequency-domain features from any set of target operation data in a set of multiple sets of target operation data, where the set of target operation data includes vibration data, acoustic data, infrared thermal imaging data, and oil and gas chromatography data, using a one-dimensional depthwise separable convolutional module of the feature extraction model, to obtain initial vibration features and initial acoustic features; a second feature extraction unit, used to extract spatial features from the infrared thermal imaging data using a two-dimensional depthwise separable convolutional module of the feature extraction model, to obtain initial infrared features; an encoding unit, used to encode the oil and gas chromatography data using a fully connected layer of the feature extraction model, to obtain initial oil and gas features; a feature fusion unit, used to fuse the initial vibration features, initial acoustic features, initial infrared features, and initial oil and gas features using a multi-head cross-attention module of the feature extraction model, to obtain initial features of any set of target operation data; and a multiple initial feature determining unit, used to obtain multiple initial features by determining the initial features of any set of target operation data.

[0016] Optionally, the second determining module includes: a first determining unit, configured to determine the real data distribution based on an initial first feature; a second determining unit, configured to determine the virtual data distribution based on an initial second feature; and a third determining unit, configured to determine the first data distribution difference based on the real data distribution and the virtual data distribution.

[0017] Optionally, the filtering module includes: a fourth determining unit for determining uncertainty index values ​​corresponding to multiple initial features; a fifth determining unit for determining spatial representativeness index values ​​corresponding to multiple initial features; a sixth determining unit for determining the comprehensive score of the initial first feature based on the uncertainty index values ​​of the initial first feature among the uncertainty index values ​​corresponding to multiple initial features, and the spatial representativeness index value of the initial first feature among the spatial representativeness index values ​​corresponding to multiple initial features; a seventh determining unit for determining the comprehensive score of the initial second feature based on the uncertainty index values ​​of the initial second feature among the uncertainty index values ​​corresponding to multiple initial features, the spatial representativeness index value of the initial second feature among the spatial representativeness index values ​​corresponding to multiple initial features, and the transferability weight of the initial second feature, wherein the transferability weight is used to characterize the marginal contribution of the virtual operating data corresponding to the initial second feature to improving the performance of the state-aware model, and the state-aware model is used to determine the operating state of the target transformer; a filtering unit for filtering a preset number of initial features in descending order based on the comprehensive scores of the initial first feature and the initial second feature to obtain multiple candidate features; and a labeling unit for labeling the multiple candidate features to obtain multiple filtered features.

[0018] Optionally, the drift detection module includes: a comprehensive drift degree determination unit, used to obtain a comprehensive drift degree based on multiple screening features; an eighth determination unit, used to determine that no data drift has occurred when the comprehensive drift degree is less than or equal to a preset drift threshold, and to determine the multiple screening features as multiple target features; or, a ninth determination unit, used to determine that data drift has occurred when the comprehensive drift degree is greater than the preset drift threshold, and to correct the multiple screening features respectively to obtain multiple target features.

[0019] Optionally, the comprehensive drift degree determination unit includes: a tenth determination unit, used to determine a first drift degree based on the model performance data of the state-aware model in the first time period; a first partitioning unit, used to partition multiple screening features into multiple first screening features, multiple second screening features, and multiple third screening features in a sliding window manner, wherein the first time window corresponding to the multiple first screening features is smaller than the second time window corresponding to the multiple second screening features, and the second time window is smaller than the third time window corresponding to the multiple third screening features; an eleventh determination unit, used to determine a second drift degree based on the multiple first screening features, multiple second screening features, multiple third screening features, and multiple historical features in the sample library; a second partitioning unit, used to partition the multiple screening features into multiple screening feature subsets according to a preset time window; a twelfth determination unit, used to determine the outlier density corresponding to each of the multiple screening feature subsets, wherein the outlier density is used to characterize the degree of data distribution anomaly of the corresponding screening feature subset; a thirteenth determination unit, used to determine a third drift degree based on the outlier density corresponding to each of the multiple screening feature subsets; and a fourteenth determination unit, used to determine the comprehensive drift degree based on the first drift degree, the second drift degree, and the third drift degree.

[0020] Optionally, the eleventh determining unit includes: a fifteenth determining unit, used to determine the first screening data distribution of multiple first screening features, the second screening data distribution of multiple second screening features, the third screening data distribution of multiple third screening features, and the historical data distribution of multiple historical features; a sixteenth determining unit, used to determine a second data distribution difference based on the first screening data distribution and the historical data distribution; a seventeenth determining unit, used to determine a third data distribution difference based on the second screening data distribution and the historical data distribution; an eighteenth determining unit, used to determine a fourth data distribution difference based on the third screening data distribution and the historical data distribution; and a nineteenth determining unit, used to determine a second drift degree based on the second data distribution difference, the third data distribution difference, and the fourth data distribution difference.

[0021] Optionally, the ninth determining unit includes: a twentieth determining unit, used to determine the target source causing data drift based on the first drift degree, the second drift degree, and the third drift degree; and a correction unit, used to correct multiple screening features based on the comprehensive drift degree and the target source to obtain multiple target features.

[0022] According to another aspect of the embodiments of this application, a non-volatile storage medium is provided, which stores multiple instructions adapted for a digital twin-based transformer sample library expansion method to be loaded and executed by a processor.

[0023] According to another aspect of the embodiments of this application, an electronic device is provided, including: one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any one of the following methods for expanding a transformer sample library based on digital twins.

[0024] According to another aspect of the embodiments of this application, a computer program product is provided that, when executed on a data processing device, is adapted to perform the steps of a method for expanding a transformer sample library based on digital twins.

[0025] In this embodiment, multiple sets of target operating data of the target transformer are acquired. These multiple sets of target operating data include real operating data and virtual operating data. The real operating data is obtained by data collection from the target transformer, and the virtual operating data is obtained based on a digital twin of the target transformer. Based on the multiple sets of target operating data, multiple initial features are determined. These initial features include an initial first feature of the real operating data and an initial second feature of the virtual operating data. Each set of target operating data corresponds one-to-one with the multiple initial features. Based on the multiple initial features, a first data distribution difference between the real operating data and the virtual operating data is determined. If the first data distribution difference is greater than a preset difference threshold, the multiple initial features are filtered to obtain multiple filtered features. Drift detection is performed based on the multiple filtered features to obtain multiple target features. Based on the multiple target features, the sample library is expanded. The goal is to acquire multiple sets of target operating data for the target transformer, including real operating data and virtual operating data generated based on a digital twin. When the first data distribution difference between the real operating data and the virtual operating data is greater than a preset difference threshold, the sample library is expanded based on the initial features obtained from the multiple sets of target operating data. This achieves the technical effect of increasing the number and coverage of samples in the transformer sample library, thereby solving the technical problems of sample scarcity and insufficient sample coverage in related technologies. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0027] Figure 1 This is a flowchart of a method for expanding a transformer sample library based on digital twins, according to an embodiment of this application;

[0028] Figure 2 This is a flowchart of an optional digital twin-based transformer sample library expansion method provided according to an embodiment of this application;

[0029] Figure 3 This is a schematic diagram of an optional lightweight edge data fusion framework provided according to an embodiment of this application;

[0030] Figure 4 This is a schematic diagram of a transformer sample library expansion device based on digital twins according to an embodiment of this application;

[0031] Figure 5 This is a structural diagram of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] It should be noted that the information collected in this application (including but not limited to preset difference thresholds, preset drift thresholds, etc.) and data (including but not limited to multiple sets of target operation data, model performance data, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or institutions, providing users with corresponding operation entry points for them to choose to agree to or refuse automated decision results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0035] According to an embodiment of this application, a method embodiment for expanding a transformer sample library based on digital twins is provided. 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. Furthermore, 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.

[0036] Figure 1 This is a flowchart of a method for expanding a transformer sample library based on digital twins, according to an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0037] Step S102: Obtain multiple sets of target operating data for the target transformer. The multiple sets of target operating data include real operating data and virtual operating data. The real operating data is obtained by data acquisition from the target transformer, and the virtual operating data is obtained based on the digital twin of the target transformer.

[0038] It is understood that the aforementioned target operational data may include, but is not limited to, vibration data, acoustic data, infrared thermography data, and oil and gas chromatography data of the target transformer. By integrating virtual operational data generated by a high-fidelity digital twin with real operational data collected on-site, a hybrid data source covering all operating conditions and multiple types can be constructed. This effectively overcomes the limitations of single real operational data in terms of sample quantity and category balance, and improves the data completeness and distribution diversity of the sample library.

[0039] Optionally, to achieve high-fidelity virtual operation data generation and dynamic detection of initial characteristic differences between virtual and real systems, a high-fidelity simulation model of multi-physics fields, including electromagnetic-thermal-fluid coupling, can be constructed based on the digital twin of the transformer. This model can simulate normal, abnormal, and extreme operating conditions, generating virtual operation data with precise physical meaning and interpretable labels. For example, by simulating the winding hot spot temperature distribution under different load and cooling conditions, or the winding deformation and stress field evolution process under different short-circuit current impact conditions, a multi-dimensional characterization of the complex operating state of the transformer can be achieved.

[0040] Optionally, to improve the simulation fidelity and physical consistency of the virtual operation data, a multi-scale coupled simulation framework can be further constructed. The finite element method (FEM) can be used to simulate the overall distribution characteristics of the electromagnetic and temperature fields. Combined with the aging and degradation mechanism of the oil-paper insulation material, the synergistic coupling of the multi-scale models can be achieved through cross-scale parameter transfer. The resulting virtual operation data not only includes fault type labels but also interpretable physical parameters such as hotspot temperatures and partial discharge charge, providing a high-quality data foundation for the subsequent training and inference of the state-aware model.

[0041] Step S104: Based on multiple sets of target running data, determine multiple initial features, including the initial first feature of the real running data and the initial second feature of the virtual running data, with each set of target running data corresponding to one of the multiple initial features;

[0042] It is understandable that feature extraction is performed on multiple sets of target running data to obtain multiple initial features that correspond one-to-one with the multiple sets of target running data.

[0043] In one optional embodiment, multiple initial features are determined based on multiple sets of target operation data, including: for any set of target operation data, where the set of target operation data includes vibration data, acoustic data, infrared thermal imaging data, and oil and gas chromatography data, a one-dimensional depthwise separable convolutional module of the feature extraction model is used to extract time-domain and frequency-domain features from the vibration data and acoustic data to obtain initial vibration features and initial acoustic features; a two-dimensional depthwise separable convolutional module of the feature extraction model is used to extract spatial features from the infrared thermal imaging data to obtain initial infrared features; a fully connected layer of the feature extraction model is used to encode the oil and gas chromatography data to obtain initial oil and gas features; a multi-head cross-attention module of the feature extraction model is used to fuse the initial vibration features, initial acoustic features, initial infrared features, and initial oil and gas features to obtain the initial features of any set of target operation data; and multiple initial features are obtained by determining the initial features of any set of target operation data.

[0044] It is understood that, given that the target operational data includes vibration data, acoustic data, infrared thermography data, and oil and gas chromatography data, feature extraction is performed on any one of the multiple sets of target operational data using the following methods: For the vibration and acoustic data in any set of target operational data, a one-dimensional depthwise separable convolutional module of the feature extraction model is used to extract time-domain and frequency-domain features, obtaining initial vibration features and initial acoustic features. For the infrared thermography data in any set of target operational data, a two-dimensional depthwise separable convolutional module of the feature extraction model is used to extract spatial features, obtaining initial infrared features. For the oil and gas chromatography data in any set of target operational data, a fully connected layer of the feature extraction model is used for nonlinear mapping and encoding to extract the structured semantic features of the oil and gas chromatography data, obtaining initial oil and gas features. A multi-head cross-attention module of the feature extraction model is used to fuse the initial vibration features, initial acoustic features, initial infrared features, and initial oil and gas features, obtaining the initial features of any one set of target operational data. Using the above methods to obtain the initial features of any one set of target operational data, feature extraction is performed on multiple sets of target operational data to obtain multiple initial features. To address the challenge of directly fusing and aligning multi-source heterogeneous operational data (vibration, acoustic, infrared, and chromatographic) due to significant modal differences and varying physical meanings, this study extracts unique initial features for each modality to achieve standardized representation and semantic alignment of multimodal data. This provides a structured and comparable input basis for subsequent cross-modal attention fusion and calculation of differences in virtual and real data distribution.

[0045] Optionally, in the lightweight model (i.e., the feature extraction model), for transformer vibration data and acoustic data, which are two modes with strong temporal characteristics, a one-dimensional deep separable convolutional module is introduced at the front end of their respective modal coding branches to focus on extracting time-domain and frequency-domain features. For infrared thermal imaging data, a two-dimensional deep separable convolutional module is used for spatial feature extraction in the corresponding spatial feature branch. For structured data such as oil and gas chromatography, a fully connected layer is used for encoding in the corresponding structured coding branch. Through the feature extraction of the above multimodal operating data, a unified feature space for the transformer is constructed. A modal consistency constraint loss function is introduced at the end of the branch network. By maximizing the mutual information between different initial modal features (initial vibration features, initial acoustic features, initial infrared features, and initial oil and gas features), cross-modal semantic alignment is ensured, thereby improving the robustness and interpretability of the fused features (i.e., initial features). Subsequently, the initial modal features extracted from each branch are input into the cross-modal attention fusion module (i.e., the multi-head cross-attention module) to dynamically calculate the correlation weights between different modal features, thereby achieving feature fusion.

[0046] Optionally, the modal consistency constraint loss function It can be constructed in the following way:

[0047]

[0048] in, The number of sets of data to run. The total number of modes, For the first The first group running data Initial modal features of each modality, For the first The first group running data Initial modal features of each modality, This refers to the mutual information between two initial modal features.

[0049] Optionally, the multi-head cross-attention module can be constructed based on a multi-head cross-attention mechanism, where each attention head focuses on learning the correlation of the initial modal feature subspaces of different modalities. For the vibration-acoustic modal pair, an attention head based on joint attention in the time and frequency domains is constructed, which uses learnable convolutional kernels to dynamically adjust the feature response intensity in the time and frequency domains; for the thermal imaging-oil and gas chromatography modal pair, an attention head based on a heterogeneous graph attention network is constructed, which uses the temperature field distribution and chemical features as heterogeneous nodes to propagate correlation information in the graph structure.

[0050] Optionally, lightweight models can be pipelined for optimization based on a "pruning-distillation-quantization" process. This integrated "pruning-distillation-quantization" pipeline for lightweight model optimization is a fine-tuning sequence. Structured pruning assesses importance based on the L1 norm of the model's convolutional channels or the scaling factor of the BN (Batch Normalization) layer, iteratively removing redundant channels to accelerate the model. The pruned model faces accuracy loss, which is compensated for through knowledge distillation. Drawing inspiration from existing digital twin systems that use complex models to guide lightweight models, a large and accurate "teacher model" can be trained in the cloud, learning deep fusion representations of multimodal operational data. Then, the lightweight "student model" not only learns the labels of the original operational data but, more importantly, mimics the rich feature representations output by the "teacher model" in intermediate feature layers, thus transferring the teacher's knowledge to the student. Finally, quantization is performed. The entire pipeline should be automated to ensure efficient deployment after optimization for each model update. The total loss function of knowledge distillation... It can be constructed in the following way:

[0051]

[0052] in, The cross-entropy loss between the student model and the real labels, KL (Kullback-Leibler Divergence, relative entropy) loss for soft labels in student and teacher models. The MSE loss is the feature layer between the student model and the teacher model. , This is a hyperparameter used to balance the weights of different loss terms.

[0053] Optionally, lightweight models can be optimized through a collaborative edge-cloud pipeline and adaptive resource scheduling. This collaborative pipeline deploys lightweight models on the edge (sensors or smart gateways), performs real-time inference, and outputs high-dimensional fused features instead of raw operational data. The edge (substation server) receives the fused features from multiple edge devices and performs spatiotemporal correlation analysis. For example, by combining the fused features of phase A winding with those of phase B winding, it determines whether fault propagation exists. Simultaneously, the edge device needs dynamic resource scheduling capabilities to prioritize the operation of critical models during peak business periods. In the cloud, leveraging its powerful computing capabilities, it aggregates data from the entire network, performs continuous training, architecture search, and version management of the lightweight model, and periodically distributes the optimized model back to the edge device, forming a closed loop.

[0054] Optionally, in the collaborative mechanism, a feature transmission specification based on the lightweight protocol MQTT (Message Queuing Telemetry Transport) / CoAP (Constrained Application Protocol) can be designed. By defining feature compression encoding and differential update strategies, communication overhead can be further reduced. Edge-side resource scheduling introduces a reinforcement learning agent, using the priority of the end device (i.e., the corresponding physical hardware) and the length of the computation queue as the state space, and lightweight model switching and computing power allocation as the action space, to achieve real-time optimization of resource utilization. A lightweight model version repository is established in the cloud, managed using semantic version numbers. Each version is associated with training data snapshots, performance metrics, and a hardware compatibility matrix. Hot updates and canary releases of the lightweight model on the edge are achieved through OTA (Over-The-Air) differential update technology, ensuring zero-interruption upgrades.

[0055] Step S106: Based on multiple initial features, determine the first data distribution difference between real running data and virtual running data;

[0056] It is understandable that, in response to the problem that the distribution of virtual and real operating data is inconsistent due to modeling simplification or environmental differences, and that direct mixing can easily lead to training bias in state-aware models, the "virtual-real gap" can be accurately identified and quantified by quantifying the distribution differences between the two in the feature space. This provides an objective basis for triggering screening and drift detection mechanisms only when the differences are significant, thereby improving the sample coverage of the sample library.

[0057] In one optional embodiment, determining a first data distribution difference between real running data and virtual running data based on multiple initial features includes: determining the real data distribution based on an initial first feature; determining the virtual data distribution based on an initial second feature; and determining the first data distribution difference based on the real data distribution and the virtual data distribution.

[0058] It is understandable that, based on the initial first feature of the real running data among multiple initial features, the distribution of real data is determined; and based on the initial second feature of the virtual running data among multiple initial features, the distribution of virtual data is determined. Based on the distributions of real and virtual data, the first data distribution difference between real and virtual running data is determined. Addressing the issue that direct fusion of virtual and real running data, which exhibit significant differences in statistical characteristics, noise patterns, and structural features, can easily lead to biases in the subsequent training of the state-aware model, this paper constructs separate data distribution representations for real and virtual running data and quantifies their distribution difference indices such as Wasserstein distance. This accurately characterizes the degree of the "virtual-real gap" between real and virtual running data, thus providing an objective and quantifiable decision-making basis for the subsequent dynamic selection of high-value initial features.

[0059] Optionally, the Wasserstein distance can be used to characterize the difference in the first data distribution. The Wasserstein distance is a mathematical metric used to measure the difference between two probability distributions. It has continuity and translation invariance, and can effectively overcome the gradient vanishing problem caused by the non-overlapping support sets of real and virtual running data in the feature space. Thus, it can still provide a smooth optimization signal even when the fidelity of the virtual running data is low, and accurately quantify the gap between real and virtual running data.

[0060] Optionally, the first data distribution difference based on Wasserstein distance The following methods can be used to determine this:

[0061]

[0062] in, This represents the first data distribution difference between real operational status data and virtual operational status data. For the true data distribution, For virtual data distribution, Let be the set of all joint distributions. The initial first feature of the actual running state data, This is the initial second feature of the virtual running state data. For all possible and Combinations between As expected, This means selecting the combination that minimizes the expected value from all possible combinations.

[0063] Alternatively, the true data distribution can be determined as follows: Collect an initial set of first features from all real running data. Using kernel density estimation (KDE) or a Gaussian mixture model (GMM), a Gaussian kernel function is superimposed around each initial first feature in the feature space to construct the true data distribution that covers the probability density distribution of the initial first features in the multidimensional feature space of the real running data. .

[0064] Alternatively, the virtual data distribution can be determined by determining the distribution of the real data.

[0065] Step S108: If the difference in the first data distribution is greater than a preset difference threshold, multiple initial features are filtered to obtain multiple filtered features.

[0066] It is understood that the above-mentioned screening features include the corresponding fusion features and annotation labels. To address the issue that directly including virtual and real operational data in the sample library when there is a significant difference in distribution (i.e., a large "virtual-real gap") can easily introduce noise interference or lead to training bias in the state-aware model, a screening mechanism is used to remove low-value or distorted initial features. This ensures the quality and consistency of the sample library while accurately improving the quantity, quality, and coverage of samples.

[0067] In one optional embodiment, multiple initial features are filtered to obtain multiple filtered features, including: determining the uncertainty index values ​​corresponding to each of the multiple initial features; determining the spatial representativeness index values ​​corresponding to each of the multiple initial features; determining the comprehensive score of the initial first feature based on the uncertainty index values ​​of the initial first feature among the uncertainty index values ​​corresponding to the multiple initial features, and the spatial representativeness index value of the initial first feature among the spatial representativeness index values ​​corresponding to the multiple initial features; determining the comprehensive score of the initial second feature based on the uncertainty index values ​​of the initial second feature among the uncertainty index values ​​corresponding to the multiple initial features, the spatial representativeness index value of the initial second feature among the spatial representativeness index values ​​corresponding to the multiple initial features, and the transferability weight of the initial second feature, wherein the transferability weight is used to characterize the marginal contribution of the virtual operating data corresponding to the initial second feature to improving the performance of the state-aware model, and the state-aware model is used to determine the operating state of the target transformer; filtering a preset number of initial features in descending order based on the comprehensive scores of the initial first feature and the initial second feature to obtain multiple candidate features; and labeling the multiple candidate features to obtain multiple filtered features.

[0068] It can be understood that the uncertainty index values ​​corresponding to multiple initial features and the spatial representativeness index values ​​corresponding to multiple initial features are determined. The uncertainty index values ​​can be prediction entropy, confidence level, or a weighted value of both. From the uncertainty index values ​​corresponding to the multiple initial features, the uncertainty index value of the initial first feature is determined, and from the spatial representativeness index values ​​corresponding to the multiple initial features, the spatial representativeness index value of the initial first feature is determined. Based on the uncertainty index value and the spatial representativeness index value of the initial first feature, a weighted calculation method is used to obtain the comprehensive score of the initial first feature. From the uncertainty index values ​​corresponding to the multiple initial features, the uncertainty index value of the initial second feature is determined, and from the spatial representativeness index values ​​corresponding to the multiple initial features, the spatial representativeness index value of the initial second feature is determined. Based on the uncertainty index value, the spatial representativeness index value, and the transferability weight of the initial second feature, a weighted calculation method is used to obtain the comprehensive score of the initial second feature. Based on the comprehensive scores of the initial first and second features, a predetermined number of initial features are selected in descending order of their scores and used as multiple candidate features. Multiple candidate features are labeled to obtain corresponding labels for each candidate feature. These candidate features and their corresponding labels together constitute the screening features, resulting in multiple screening features. By integrating multi-dimensional indicators such as uncertainty, spatial representativeness, and transferability, accurate selection of initial features is achieved. While ensuring sample diversity and distribution coverage, this maximizes the effectiveness of transferring virtual running data to real-world scenarios, thereby improving the quality and coverage of the screening features.

[0069] Optionally, the aforementioned uncertainty index value is used to quantify the degree of confidence deficiency when the corresponding initial feature is predicted by the current state-aware model, reflecting whether the operating state of the target transformer represented by the initial feature is in the cognitive blind zone or high-risk boundary area of ​​the state-aware model. The spatial representativeness index value quantifies the distribution coverage and cluster center distance of the corresponding initial feature in the multimodal feature space, reflecting whether the initial feature can effectively represent the typical distribution pattern of a certain type of operating state of the target transformer.

[0070] Optionally, spatial representativeness can be represented by the average distance of an initial feature to its K nearest neighbor initial features in the feature space, or by the Euclidean distance to the cluster center. A smaller distance indicates that the initial feature is located in a high-density region (cluster center) and has high representativeness. A larger distance indicates that the initial feature is located in a low-density region or boundary (cluster edge / outlier), and has high diversity (representing rare or extreme conditions). Spatial representativeness can also be represented by the probability density value at the initial feature, either a local density estimate or a kernel density estimate (KDE). A high probability density value indicates that a large number of similar initial features are clustered around the initial feature, indicating high representativeness. A low probability density value indicates that the initial feature is sparsely surrounded, potentially representing a new pattern or noise. Spatial representativeness can also be represented by a greedy coverage metric based on core-set selection, used to measure how many other initial features the initial feature covers. If an initial feature can represent many other initial features, it indicates high representativeness.

[0071] Optionally, the logic of introducing transferability weights only for the initial second features of virtual operating data to calculate the comprehensive score is as follows: the initial first features of real operating data are themselves standard data reflecting the real operating state of the target transformer, and their value has been fully measured by uncertainty (reflecting the blind spots of model cognition, i.e., information gain) and spatial representativeness (reflecting the degree of coverage of data distribution), so no additional transfer weights are needed; while the initial second features of virtual operating data are obtained based on the digital twin of the target transformer, which may have domain bias. Introducing transferability weights aims to quantify their ability to adapt to the real operating state of the target transformer, so that in the screening process, high-fidelity initial second features that can not only narrow the gap between virtual and real distributions but also directly improve the performance of the state perception model are retained first, avoiding the contamination of the sample library by low-quality or distorted initial second features.

[0072] Optionally, a closed-loop feature optimization and annotation system centered on active learning can be constructed through active learning-guided feature selection and human-machine collaborative intelligent annotation. The sample library periodically acquires initial features from two sources: first, initial first features based on real-world operational data from the edge; and second, initial second features based on virtual operational data generated by digital twin simulation. The system uses the current state-aware model (such as a fault classification model) to infer the initial features, calculating their uncertainty index values ​​(such as prediction entropy and confidence) and spatial representativeness index values ​​representing their representativeness in the feature space. It prioritizes the initial features with the highest information gain in the state-aware model's cognitive blind spots as candidate features for subsequent annotation processes.

[0073] Optionally, a human-machine collaborative intelligent annotation mode can be adopted in the feature annotation stage. First, a pre-trained automatic annotation model is used to generate initial labels, which are then reviewed, corrected, and confirmed by domain experts through an interactive annotation interface. The feedback generated by experts during the annotation process will be continuously used to update the automatic annotation model, improving its accuracy over time. This significantly reduces reliance on manual annotation, allowing experts to focus their efforts on the most valuable and uncertain candidate features.

[0074] Optionally, the active learning strategy design integrates uncertainty sampling and diversity sampling mechanisms. On the one hand, the most uncertain initial features of the state-aware model are screened using indicators such as prediction entropy; on the other hand, a core set selection algorithm is introduced to select a small-scale representative initial feature set that can cover the feature distribution boundary and high-density regions in the initial feature space. For the initial second feature, a transferability weight is further introduced to measure its potential marginal contribution to improving the performance of the state-aware model, and the online estimation and dynamic adjustment of this weight are achieved through meta-learning methods.

[0075] Optionally, at the human-machine collaborative support platform level, based on interactive annotation tools, and utilizing functions such as text annotation, multimodal data linkage display, and label ambiguity identification, experts can be supported in annotating confidence intervals for uncertain labels. An active learning agent and expert behavior modeling mechanism are introduced to record experts' annotation habits and error correction patterns, gradually constructing an expert knowledge graph for personalized recommendation annotation strategies and sample distribution. After candidate feature annotation is completed, the results will automatically trigger incremental training and strategy updates of the automatic annotation model, forming a rapid closed loop of "annotation-learning-recommendation," continuously improving the sample quality of the sample library and the performance of the automatic annotation model.

[0076] Step S110: Drift detection is performed based on multiple screening features to obtain multiple target features;

[0077] As can be understood, drift refers to the significant statistical deviation between the distribution patterns of internal state parameters or external monitoring data of a target transformer and its historical baseline state during long-term operation, caused by equipment aging, changes in environmental conditions, or the evolution of faults. By detecting data drift in the screening features, subtle evolutionary trends in the health status of the target transformer can be captured in real time, providing a basis for subsequent correction of the screening features.

[0078] In one optional embodiment, drift detection is performed based on multiple screening features to obtain multiple target features, including: obtaining a comprehensive drift degree based on multiple screening features; if the comprehensive drift degree is less than or equal to a preset drift threshold, determining that no data drift has occurred, and identifying the multiple screening features as multiple target features; or, if the comprehensive drift degree is greater than the preset drift threshold, determining that data drift has occurred, and correcting the multiple screening features respectively to obtain multiple target features.

[0079] It is understandable that the overall drift level is determined based on multiple screening features, and then compared with a preset drift threshold. If the overall drift level is less than or equal to the preset drift threshold, it indicates that the target transformer is operating stably and no data drift has occurred, and the multiple screening features are identified as multiple target features. If the overall drift level is greater than the preset drift threshold, it indicates that the target transformer may have undergone state evolution due to equipment aging, sudden environmental changes, or new fault modes, and data drift is determined to have occurred. The multiple screening features are then corrected accordingly to obtain multiple target features. By setting a preset drift threshold to classify and judge the overall drift level, a dual mechanism is achieved: efficient direct data entry when the feature distribution is stable and targeted correction when drift occurs. This ensures both the efficiency and real-time nature of sample database expansion and the data quality of the entered features, effectively preventing distorted features from contaminating the sample database.

[0080] In one optional embodiment, the comprehensive drift degree is obtained based on multiple screening features, including: determining a first drift degree based on the model performance data of the state-aware model in a first time period; dividing the multiple screening features into multiple first screening features, multiple second screening features, and multiple third screening features using a sliding window method, wherein the first time window corresponding to the multiple first screening features is smaller than the second time window corresponding to the multiple second screening features, and the second time window is smaller than the third time window corresponding to the multiple third screening features; determining a second drift degree based on the multiple first screening features, multiple second screening features, multiple third screening features, and multiple historical features in the sample library; dividing the multiple screening features into multiple screening feature subsets according to a preset time window; determining the outlier density corresponding to each of the multiple screening feature subsets, wherein the outlier density is used to characterize the degree of data distribution anomaly of the corresponding screening feature subset; determining a third drift degree based on the outlier density corresponding to each of the multiple screening feature subsets; and determining the comprehensive drift degree based on the first drift degree, the second drift degree, and the third drift degree.

[0081] The overall drift level is determined as follows: First, based on the model performance data of the state-aware model in the first time period, such as model accuracy, loss function value, and prediction confidence, the first drift level is determined. This first time period corresponds to the collection period of multiple sets of target running data. Second, multiple filtering features are divided into multiple first filtering features, multiple second filtering features, and multiple third filtering features using a sliding window approach. For example, filtering features from the most recent hour are used as first filtering features, filtering features from the most recent 24 hours are used as first filtering features, and filtering features from the most recent week are used as third filtering features. Then, based on the multiple first filtering features, multiple second filtering features, and multiple third filtering features, as well as multiple historical features in the sample library, the second drift level is determined. Next, the multiple filtering features are divided into equal time windows according to a preset time window, resulting in multiple subsets of filtering features. For example, multiple filtering features are divided into multiple subsets at one-hour intervals, where each subset includes filtering features acquired within one hour. Then, the outlier densities corresponding to multiple screening feature subsets are determined, and the third drift degree is determined based on the outlier densities corresponding to these subsets. Finally, the comprehensive drift degree is determined by weighted summation based on the first, second, and third drift degrees. By fusing the first drift degree, which reflects the performance degradation of the short-term state-aware model, the second drift degree, which reflects the macroscopic evolution of data distribution across multiple time scales, and the third drift degree, which reflects the aggregation of local anomalous patterns, a multi-scale, multi-dimensional comprehensive drift evaluation mechanism is constructed. This mechanism can comprehensively capture various forms of conceptual drift, from sudden performance degradation to gradual distribution shifts and local outlier patterns, improving the sensitivity, robustness, and adaptability to complex operating conditions of drift detection.

[0082] Optionally, an online concept drift detection module can be embedded in the sample library service interface to continuously monitor the system's operating status. This module focuses on tracking two key signals: First, the difference in data distribution between the distribution of the selected features obtained from the real-time input data stream and the historical data distribution of the historical features in the sample library (i.e., the second, third, and fourth data distribution differences) determines the degree of second drift. The aforementioned data distribution differences are quantified using a distribution metric based on Wasserstein distance. Second, the performance change of the downstream state-aware model within a preset sliding time window (i.e., the first time period), such as a decrease in classification accuracy or a continuous increase in the loss function, determines the degree of first drift.

[0083] Optionally, to improve the sensitivity and robustness of drift detection, a multi-window, multi-scale drift detection framework can be introduced when determining the second degree of drift: a short time window (hourly) is used to capture sudden drifts, and a long time window (weekly) is used to identify gradual evolutionary trends. Simultaneously, an isolated forest can be used to detect outlier patterns in the feature space, determining the outlier densities corresponding to multiple subsets of selected features obtained by partitioning multiple selected features, and thus determining the third degree of drift. When the outlier density continuously increases, it is determined as potential distribution drift. The first, second, and third degrees of drift can be combined to determine the comprehensive drift degree, and drift detection is performed based on the comprehensive drift degree. Once a significant drift phenomenon is detected, the system will promptly trigger an early warning mechanism, realizing online detection and root cause analysis of concept drift.

[0084] In one optional embodiment, determining a second degree of drift based on multiple first screening features, multiple second screening features, multiple third screening features, and multiple historical features in a sample library includes: determining a first screening data distribution of multiple first screening features, a second screening data distribution of multiple second screening features, a third screening data distribution of multiple third screening features, and a historical data distribution of multiple historical features; determining a second data distribution difference based on the first screening data distribution and the historical data distribution; determining a third data distribution difference based on the second screening data distribution and the historical data distribution; determining a fourth data distribution difference based on the third screening data distribution and the historical data distribution; and determining a second degree of drift based on the second data distribution difference, the third data distribution difference, and the fourth data distribution difference.

[0085] It is understandable that, based on multiple first-selection features, multiple second-selection features, multiple third-selection features, and multiple historical features, the first-selection data distribution of the multiple first-selection features, the second-selection data distribution of the multiple second-selection features, the third-selection data distribution of the multiple third-selection features, and the historical data distribution of the multiple historical features are determined respectively. The second data distribution difference is determined based on the first-selection data distribution and the historical data distribution. The third data distribution difference is determined based on the third-selection data distribution and the historical data distribution. The fourth data distribution difference is determined based on the second, third, and fourth data distribution differences. A weighted summation method is used to determine the second degree of drift. By introducing a comparison between multi-timescale (short, medium, and long-term) selection features and the baseline historical data distribution, it is possible to simultaneously capture recent abrupt changes, medium-term evolutions, and long-term trend shifts in feature distributions, thereby comprehensively and meticulously depicting the evolution characteristics of the target transformer's operating state across different time dimensions, improving the dimensional coverage and assessment accuracy of drift detection.

[0086] Optionally, the distribution of the first, second, third, and historical data can be determined by determining the distribution of real and virtual data.

[0087] Optionally, the second, third, and fourth data distribution differences can be determined in the following ways:

[0088]

[0089] in, At time t, the first The data distribution differences between multiple filtering features and historical features within a time window (including second, third, and fourth data distribution differences). For the first The distribution of filtering data for multiple filtering features within a given time window, where w is the distribution of the filtering data for the first filtering feature. The size of a time window, This represents the distribution of historical data.

[0090] In one optional embodiment, multiple screening features are modified to obtain multiple target features, including: determining the target source causing data drift based on a first drift degree, a second drift degree, and a third drift degree; and modifying multiple screening features based on a comprehensive drift degree and the target source to obtain multiple target features.

[0091] Understandably, based on the first, second, and third degrees of drift, a Bayesian causal graph-based inference model is used to identify the target sources causing data drift in the target transformer. These sources could include factors such as decreased cooling system fan speed, reduced oil pump efficiency, changes in load levels, or increased ambient temperature. Based on the combined degree of drift and the target source, multiple screening features are modified to obtain multiple target features. By accurately locating the physical root causes of data drift and accordingly modifying the screening features, data distortion caused by equipment aging, environmental changes, or fault evolution is effectively eliminated. This ensures that the target features entering the database retain true physical patterns while eliminating abnormal interference, thereby improving the sample quality of the sample library.

[0092] Optionally, at the root cause analysis level, an inference model based on Bayesian causal graphs can be constructed. The observed data drift levels (including first, second, and third drift levels) are used as evidence nodes. The most likely root cause node (i.e., the target source) is inferred using a belief propagation algorithm, such as a decrease in cooling system fan speed, a decline in oil pump efficiency, a change in load level, or an increase in the ambient temperature baseline. The posterior probability of each potential factor is quantified. The digital twin has a built-in parameter inversion and calibration engine that can automatically correct uncertain parameters in the inference model based on observed data, gradually approximating the behavior of the real equipment and achieving synchronous evolution between the "entity" and the "twin."

[0093] Optionally, the target source can be determined as follows: First, construct and initialize a Bayesian causal network. Construct a Bayesian causal graph reflecting the multi-physics coupling relationship of the target transformer. Nodes include potential root variables (such as cooling fan status, oil pump efficiency, load level, ambient temperature, insulation aging degree, etc.) and observed variables (such as winding temperature, vibration, oil chromatography, etc.). Assign prior probabilities to each node and train a conditional probability table based on historical normal operation data to establish causal dependencies between nodes (e.g., a decrease in fan speed leads to an increase in winding temperature). Second, construct a multi-dimensional drift evidence vector based on the first, second, and third drift degrees. The first drift degree serves as evidence of performance degradation nodes, reflecting the difficulty of the state-aware model fitting the features, and is usually strongly correlated with sudden faults or severe operating condition deviations. The second drift degree serves as evidence of distribution offset nodes, reflecting the macroscopic changes in the overall distribution of the selected features over time, used to distinguish between long-term trends (such as aging) and recent abrupt changes. The third drift degree serves as evidence of anomalous clustering nodes, reflecting outliers in local selected features, used to capture intermittent faults or local hotspots. Then, Bayesian belief propagation inference is performed. The aforementioned multidimensional drift evidence vector is substituted into the Bayesian causal graph, and backward inference is performed using the belief propagation algorithm or variable elimination algorithm. The posterior probability distribution of each potential root cause node (e.g., "fan failure," "abnormal load") is calculated given the observed data. Finally, the target source is identified and determined. The node with the highest posterior probability is selected as the target source.

[0094] Optionally, when a gradual change in a known pattern is detected, feature correction focuses on simulating such an evolutionary process to correct the selected features; when a completely new unknown pattern is identified, a conservative correction strategy is adopted, relying more on the selected features from real operating data and the selected features from high-fidelity simulation data (i.e., virtual operating data) to carry out exploratory enhancement to obtain the target features.

[0095] Optionally, the selected features can be modified as follows: First, map the modification strategy. Based on the determined target source, match the corresponding modification model from a pre-set modification strategy library. For example, if the source is "ambient temperature baseline rise," then match a temperature compensation strategy; if it is "load change," then match a load normalization strategy. Second, quantify the modification strength. Determine the magnitude or weighting coefficient of the modification based on the numerical value of the overall drift. The higher the overall drift, the more significant the data drift, requiring a stronger modification force to eliminate the bias. At the same time, the modification coefficient is weighted in conjunction with the confidence level of the target source to ensure the reliability of the modification. Then, perform feature transformation. Input the selected features into the corresponding modification model for processing, including detrending / de-interference and normalization / standardization. Detrending / de-interference refers to using physical models (such as heat balance equations) or statistical methods to remove non-entropy state changes caused by the target source from the original selected features (such as subtracting the influence of ambient temperature and compensating for heat dissipation efficiency loss). Normalization refers to mapping the selected feature values ​​back to their equivalent values ​​under the standard baseline operating conditions (such as rated load and standard ambient temperature) of the digital twin. Finally, target features are generated and verified. The corrected output features are used as target features, and their compliance with physical constraints (such as temperature not exceeding limits) and distribution consistency requirements is checked. If the corrected features are still abnormal, they are marked as high-risk features requiring manual intervention; otherwise, they are directly used as target features in the subsequent data entry process.

[0096] Step S112: Expand the sample database based on multiple target features.

[0097] It is understandable that multiple target features are encapsulated into structured samples and added to the sample library using an incremental update method, thereby expanding the sample library. By dynamically integrating target features into the sample library using an incremental update mechanism, the computational overhead and storage redundancy caused by full data reconstruction are avoided, achieving low-latency and high-efficiency expansion of the sample library.

[0098] Optionally, to achieve incremental updates, structured storage, and version management of the sample library, a sample library data structure supporting efficient retrieval and fine-grained management is designed to ensure the controllability and traceability of the sample library during continuous evolution. Each feature and its related information are uniformly encapsulated into a structured sample, containing the original data or its feature index, the corresponding feature (e.g., target feature), high-quality label information, and rich metadata such as collection time, operating conditions, sensor identification, and label confidence level. New samples are added to the library incrementally, avoiding full reconstruction of existing historical samples, and small-scale, targeted data augmentation operations are automatically triggered based on the sample category. For example, for newly added "partial discharge" samples, augmentation strategies such as noise addition and scaling can be adaptively executed to generate diverse sample variants, thereby improving the sample coverage and robustness of this category.

[0099] Optionally, an elastic forgetting mechanism can be introduced to calculate the information value weight of historical samples in the sample library based on their collection time, access frequency, and similarity to the data distribution of the latest target features. Long-term low-value outdated historical samples can be automatically archived or eliminated to maintain the effectiveness and dynamic update capability of the sample library.

[0100] Optionally, the continuous evolution of the sample library can be tightly coupled with the update mechanism of the downstream state-aware model, forming a strong closed loop of data-model collaborative optimization. Each incremental update or cleanup operation of the sample library triggers incremental learning or periodic fine-tuning of the state-aware model, enabling it to quickly adapt to new operating environments. Conversely, changes in the performance of the state-aware model provide feedback for active learning strategies and concept drift evaluation, guiding subsequent sample selection and enhancement strategy adjustments. By establishing a two-way closed-loop mechanism of "data evolution driving model optimization, and model optimization guiding data evolution in turn," the sample library possesses continuous evolution and long-term vitality. The system fully records and logs each evolution process, providing traceable evidence for long-term operational analysis and strategy optimization, ultimately achieving a closed loop of model-data collaborative optimization.

[0101] Optionally, incremental update refers to adding new target features to the sample library in an "append-write" manner without reprocessing, restoring, or reindexing existing historical features, and simultaneously updating its metadata, version snapshot, and index structure, thereby achieving efficient, low-overhead, and traceable dynamic expansion of the sample library.

[0102] Through the above steps S102 to S112, multiple sets of target operating data of the target transformer, including real operating data and virtual operating data generated based on digital twins, can be obtained. When the first data distribution difference between real operating data and virtual operating data is greater than a preset difference threshold, the sample library is expanded based on the initial features obtained from the multiple sets of target operating data. This achieves the technical effect of increasing the number and coverage of samples in the transformer sample library, thereby solving the technical problems of sample scarcity and insufficient sample coverage in the transformer sample library in related technologies.

[0103] Based on the above embodiments and optional embodiments, this application proposes an optional implementation method for expanding a transformer sample library based on digital twins. This optional implementation method can be understood as a method for constructing a high-quality sample library with self-learning and evolution capabilities for digital twins.

[0104] The sample library construction method is a fundamental supporting technology for the long-term evolution of data-driven models and digital twins. It aims to achieve the systematic organization, continuous expansion, and efficient utilization of data resources while ensuring sample quality, representativeness, and reliability. As operating scenarios become increasingly complex and monitoring methods become more sophisticated, transformers generate a large amount of multi-source, multi-modal, and spatiotemporal scale data during operation. This data is not only the core foundation for constructing state-aware models and digital twins but also a key factor restricting the accuracy and generalization ability of intelligent analysis. Therefore, how to construct a high-quality, evolvable sample library system has become one of the key issues supporting the implementation of intelligent operation and maintenance and digital twin applications.

[0105] In real-world engineering scenarios, transformer sample libraries commonly suffer from problems such as sample scarcity, class imbalance, high annotation costs, and unstable operating condition distribution. Traditional sample management methods centered on static datasets are insufficient to meet the continuous learning and performance evolution requirements of state-aware models. By systematically governing and organizing multimodal raw data, scattered and heterogeneous data resources are transformed into well-structured, quality-controllable, and sustainably updated sample assets. This provides unified data support for the training, evaluation, and online updates of state-aware models. The design revolves around the entire process of "data acquisition - quality control - structured storage - incremental evolution," making the sample library not only a data storage container but also an infrastructure supporting the learning and knowledge accumulation of state-aware models.

[0106] Based on this, a method for constructing a high-quality sample library with self-learning and evolutionary capabilities for digital twins is proposed. The construction of a high-quality, self-evolving sample library for transformers in digital twins is a systematic project spanning the entire "data-model-service" chain. The sample library is not merely a static data warehouse, but an intelligent entity with the capabilities of perception, thinking, action, and evolution. First, the core challenge in constructing the data entry point for the sample library lies in achieving high-fidelity, low-latency feature extraction and fusion of multimodal operational data under the stringent resource constraints of edge devices. Second, addressing the core bottlenecks of sample scarcity and high annotation costs, the sample library's automatic and intelligent growth is driven by the interaction between simulation data (i.e., virtual operational data) generated by the digital twin and real operational data. Furthermore, the sample library must possess long-term adaptability to address the "concept drift" problem (i.e., data drift caused by transformer equipment aging and environmental changes) resulting from transformer equipment aging and environmental changes, enabling autonomous evolution. Finally, all components are integrated into an organic whole that is operable, evaluable, and evolvable.

[0107] Figure 2 This is a flowchart of an optional digital twin-based transformer sample library expansion method provided according to an embodiment of this application, such as... Figure 2As shown, the steps of a method for constructing a high-quality sample library with self-learning and evolution capabilities for digital twins include:

[0108] Step S1: Design and Implementation of a Lightweight Edge Data Fusion Framework. This step aims to study modality-specific encoding and cross-modal attention fusion techniques, design a lightweight "pruning-distillation-quantization" integrated model (i.e., feature extraction model) with optimized pipeline, and construct a layered data fusion architecture that coordinates edge, cloud, and device to achieve high-fidelity, low-latency feature extraction and fusion of multimodal operational data such as vibration, acoustics, infrared thermography, and oil and gas chromatography.

[0109] Figure 3 This is a schematic diagram of an optional lightweight edge data fusion framework provided according to an embodiment of this application, as shown below. Figure 3 As shown, the lightweight edge data fusion framework consists of three parts: multimodal perception data (i.e., operational data, including real operational data and virtual operational data) collection and modality-specific encoding, cross-modal attention fusion and spatiotemporal correlation modeling mechanism, and edge-cloud collaborative deployment and lightweight model optimization. These will be introduced separately below.

[0110] Multimodal sensing data acquisition and modality-specific encoding. For transformer vibration data and acoustic data, which are two types of modal operation data with strong temporal characteristics, one-dimensional depthwise separable convolutions can be introduced at the front end of the branch network to focus on extracting time-domain and frequency-domain features. For infrared thermal imaging data, two-dimensional depthwise separable convolutions are used for spatial feature extraction. For structured data such as oil and gas chromatography, fully connected layers are used for encoding. Subsequently, a cross-modal attention mechanism is introduced to dynamically calculate the correlation weights between feature maps of different modalities.

[0111] In the lightweight model, for transformer vibration and acoustic data—two modes with strong temporal characteristics—one-dimensional deep separable convolutional modules are introduced at the front end of their respective modal coding branches to focus on extracting time-domain and frequency-domain features. For infrared thermal imaging data, a two-dimensional deep separable convolutional module is used for spatial feature extraction in the corresponding spatial feature branch. For structured data such as oil and gas chromatography, fully connected layers are used for encoding in the corresponding structured coding branch. Through feature extraction from the above multimodal operating data, a unified feature space for the transformer is constructed. A modal consistency constraint loss function is introduced at the end of the branch network to maximize the mutual information between different initial modal features (initial vibration features, initial acoustic features, initial infrared features, and initial oil and gas features), ensuring cross-modal semantic alignment and thus improving the robustness and interpretability of the fused features (i.e., initial features). Subsequently, the initial modal features extracted from each branch are input into a cross-modal attention fusion module (i.e., a multi-head cross-attention module) to dynamically calculate the correlation weights between different modal features, thereby achieving feature fusion.

[0112] Modal consistency constraint loss function It is constructed in the following way:

[0113]

[0114] in, The number of sets of data to run. The total number of modes, For the first The first group running data Initial modal features of each modality, For the first The first group running data Initial modal features of each modality, This refers to the mutual information between two initial modal features.

[0115] A cross-modal attention fusion and spatiotemporal correlation modeling mechanism is proposed. The multi-head cross-attention module is constructed based on this mechanism, where each attention head focuses on learning the correlations of the initial modal feature subspaces of different modalities. For the vibration-acoustic modal pair, an attention head based on joint attention in the time and frequency domains is constructed, which dynamically adjusts the feature response intensity in the time and frequency domain using learnable convolutional kernels. For the thermal imaging-oil and gas chromatography modal pair, an attention head based on a heterogeneous graph attention network is constructed, which uses temperature field distribution and chemical features as heterogeneous nodes to propagate correlation information within the graph structure.

[0116] Edge-Cloud Collaborative Deployment and Lightweight Model Optimization. The "Pruning-Distillation-Quantization" three-in-one lightweight model optimization pipeline is a fine-tuning sequence. Structured pruning assesses importance based on the L1 norm of the model's convolutional channels or the scaling factor of the BN (Batch Normalization) layer, iteratively removing redundant channels to accelerate the model. The pruned model faces accuracy loss, which is compensated for through knowledge distillation. Drawing inspiration from existing digital twin systems that use complex models to guide lightweight models, a large and accurate "teacher model" is trained in the cloud, learning deep fusion representations of multimodal operational data. Then, the lightweight "student model" not only learns the labels of the original operational data but, more importantly, mimics the rich feature representations output by the "teacher model" in the intermediate feature layers, thus transferring the teacher's knowledge to the student. Finally, quantization is performed. The entire pipeline should be automated to ensure efficient deployment after optimization for each model update. The total loss function of knowledge distillation... It is constructed in the following way:

[0117]

[0118] in, The cross-entropy loss between the student model and the real labels, KL (Kullback-Leibler Divergence, relative entropy) loss for soft labels in student and teacher models. The MSE loss is the feature layer between the student model and the teacher model. , This is a hyperparameter used to balance the weights of different loss terms.

[0119] Through a converged pipeline and adaptive resource scheduling via edge-cloud collaboration, lightweight models are deployed on the edge (sensors or smart gateways) to perform real-time inference and output high-dimensional fused features rather than raw operational data. The edge (substation server) is responsible for receiving fused features from multiple edge devices and performing spatiotemporal correlation analysis. For example, by combining the fused features of phase A winding with those of phase B winding, it determines whether fault propagation exists. Simultaneously, the edge must possess dynamic computing resource scheduling capabilities to prioritize the operation of critical models during peak business periods. In the cloud, leveraging its powerful computing capabilities, it aggregates data from the entire network, performs continuous training, architecture search, and version management of lightweight models, and periodically distributes optimized models to the edge, forming a closed loop.

[0120] In the collaborative mechanism, a feature transmission specification based on the lightweight protocol MQTT (Message Queuing Telemetry Transport) / CoAP (Constrained Application Protocol) needs to be designed. This involves defining feature compression encoding and differential update strategies to further reduce communication overhead. Edge-side resource scheduling introduces a reinforcement learning agent, using the priority of the end device (i.e., the corresponding physical hardware) and the length of the computation queue as the state space, and lightweight model switching and computing power allocation as the action space, to achieve real-time optimization of resource utilization. A lightweight model version repository is established in the cloud, managed using semantic version numbers. Each version is associated with training data snapshots, performance metrics, and a hardware compatibility matrix. Hot updates and canary releases of the lightweight model on the edge are achieved through OTA (Over-The-Air) differential update technology, ensuring zero-interruption upgrades.

[0121] Step S2 involves constructing a self-learning and incremental update mechanism for the sample library based on "virtual-real interaction." This step aims to study the generation of high-fidelity virtual operation data in multiphysics and the dynamic detection technology of initial virtual-real feature differences (i.e., differences in the first data distribution), construct an active learning-guided feature optimization and human-machine collaborative intelligent annotation system, design a sample structured storage and full lifecycle version management scheme that supports efficient retrieval, and realize the automatic and intelligent growth and high-quality incremental update of the sample library.

[0122] To achieve high-fidelity virtual operation data generation and dynamic detection of initial feature differences between virtual and real data, a high-fidelity simulation model of multiphysics fields, including electromagnetic-thermal-fluid coupling, is constructed based on a digital twin of the transformer. This model simulates normal, abnormal, and extreme operating conditions, generating virtual operation data with precise physical meaning and interpretable labels. For example, by simulating the winding hot spot temperature distribution under different load and cooling conditions, or the winding deformation and stress field evolution under different short-circuit current impact conditions, a multi-dimensional characterization of the complex operating state of the transformer is achieved. Based on this, a dynamic detection model of initial feature differences between virtual and real data is constructed. By modeling the joint distribution of the initial first feature of the real operation data and the initial second feature of the virtual operation data, the differences between the two in statistical characteristics, noise patterns, and structural features are quantitatively characterized. The difference between virtual and real initial features is calculated using the Wasserstein distance. The difference between virtual and real initial features based on the Wasserstein distance is then presented. The following method is used to determine:

[0123]

[0124] in, This represents the first data distribution difference between real operational status data and virtual operational status data. For the true data distribution, For virtual data distribution, Let be the set of all joint distributions. The initial first feature of the actual running state data, This is the initial second feature of the virtual running state data. For all possible and Combinations between As expected, This means selecting the combination that minimizes the expected value from all possible combinations.

[0125] To improve the simulation fidelity and physical consistency of virtual operational data, a multi-scale coupled simulation framework was further constructed. The Finite Element Method (FEM) was used to simulate the overall distribution characteristics of the electromagnetic and temperature fields. Combined with the aging and degradation mechanisms of oil-paper insulation materials, cross-scale parameter transfer was used to achieve the synergistic coupling of multi-scale models. The resulting virtual operational data not only includes fault type labels but also interpretable physical parameters such as hotspot temperatures and partial discharge charge, providing a high-quality data foundation for the subsequent training and inference of the state-aware model.

[0126] By employing active learning-guided feature optimization and human-machine collaborative intelligent annotation, a closed-loop feature optimization and annotation system centered on active learning is constructed. The sample library periodically acquires initial features from two sources: first, initial first features based on real-world operational data from the edge; and second, initial second features based on virtual operational data generated from digital twin simulation. The system uses the current state-aware model (such as a fault classification model) to infer the initial features, calculating their uncertainty indices (such as prediction entropy and confidence) and spatial representativeness indices representing their representativeness in the feature space. It prioritizes the initial features with the highest information gain within the state-aware model's cognitive blind spots as candidate features for subsequent annotation processes.

[0127] In the feature annotation stage, a human-machine collaborative intelligent annotation mode is adopted. First, a pre-trained automatic annotation model is called to generate preliminary labels, which are then reviewed, corrected, and confirmed by domain experts through an interactive annotation interface. The feedback information generated by experts during the annotation process will be continuously used to update the automatic annotation model, so that its annotation accuracy will continue to improve with use, thereby significantly reducing the dependence on manual annotation and allowing experts to focus their energy on the most valuable and uncertain candidate features.

[0128] In designing the active learning strategy, an uncertainty sampling and diversity sampling mechanism are integrated. On the one hand, the most uncertain initial features of the state-aware model are screened using indicators such as prediction entropy; on the other hand, a core set selection algorithm is introduced to select a small-scale representative initial feature set that can cover the feature distribution boundary and high-density regions in the initial feature space. For the initial second feature, a transferability weight is further introduced to measure its potential marginal contribution to improving the performance of the state-aware model, and the online estimation and dynamic adjustment of this weight are achieved through meta-learning methods.

[0129] At the human-machine collaborative support platform level, based on interactive annotation tools, and utilizing functions such as text annotation, multimodal data linkage display, and label ambiguity identification, experts are supported in annotating confidence intervals for uncertain labels. An active learning agent and expert behavior modeling mechanism are introduced to record experts' annotation habits and error correction patterns, gradually constructing an expert knowledge graph for personalized recommendation annotation strategies and sample distribution. After candidate feature annotation is completed, the results automatically trigger incremental training and strategy updates for the automatic annotation model, forming a rapid closed loop of "annotation-learning-recommendation," continuously improving the sample quality of the sample library and the performance of the automatic annotation model.

[0130] To achieve incremental updates, structured storage, and versioned management of the sample library, a data structure supporting efficient retrieval and fine-grained management is designed to ensure the controllability and traceability of the sample library during its continuous evolution. Each feature and its related information are uniformly encapsulated into a structured sample, containing the original data or its feature index, the corresponding feature (e.g., target feature), high-quality label information, and rich metadata such as collection time, operating conditions, sensor identification, and label confidence level. New samples are added to the library incrementally, avoiding a full reconstruction of existing historical samples, and small-scale, targeted data augmentation operations are automatically triggered based on the sample category. For example, for newly added "partial discharge" samples, augmentation strategies such as noise addition and scaling can be adaptively executed to generate diverse sample variants, thereby improving the sample coverage and robustness of this category.

[0131] To ensure the long-term maintainability of the sample library and the controllability of state-aware model training, a strict version management mechanism is implemented. Each critical sample change (such as adding fault types, adjusting annotation rules, or upgrading enhancement strategies) generates an independent version snapshot. This snapshot depicts the phased state of the sample library structure and sample distribution, supporting performance evolution analysis, version rollback, and comparative evaluation of the state-aware model. The version snapshot also synchronously records data health indicators such as sample distribution histograms, class imbalance indices, and the proportion of samples corresponding to virtual running data, providing sample quality early warnings and decision-making basis for subsequent state-aware model training and deployment.

[0132] In terms of storage architecture design, a tiered storage strategy is adopted to balance performance and cost. The hot data layer stores frequently accessed features and tag information based on an in-memory database, supporting online inference and rapid training; the warm data layer uses a time-series database to manage raw runtime data, facilitating time-dimensional querying and analysis; the cold data layer relies on a distributed file system to archive historical data, meeting the needs of long-term retention and auditing. Each sample is assigned a unique digital object identifier, and a complete data lineage record is established, covering the entire process of data collection, transmission, annotation, enhancement, and storage, achieving full lifecycle traceability management of samples.

[0133] In terms of version management and storage optimization, a data version control mechanism based on the Git-LFS (Git Large File Storage) concept is introduced. Version iteration is completed by recording data difference patches instead of full copying, significantly reducing storage overhead and improving management efficiency. For the sample augmentation stage, a conditional generative adversarial network is introduced. This network learns the multimodal manifold distribution of specific fault types, generating synthetic samples that conform to physical laws and possess diversity. This avoids the distribution distortion problem caused by traditional simple augmentation methods, thereby continuously improving the quality and application value of the sample library.

[0134] Step S3: Implementation of the self-evolution strategy and closed-loop management mechanism for the sample library. This step aims to study online detection of multi-scale concept drift and Bayesian causal root cause analysis techniques, propose adaptive data augmentation strategies and elastic forgetting mechanisms, and construct a two-way collaborative closed loop of "data evolution driving model optimization, and model optimization guiding data evolution in turn," thereby achieving the autonomous adaptation and continuous evolution of the sample library.

[0135] An online concept drift detection module is embedded in the sample library service interface to continuously monitor the system's operating status. This module focuses on tracking two key signals: First, it determines the second degree of drift by identifying the data distribution differences between the screening data distribution of the screening features obtained from the real-time input data stream and the historical data distribution of the historical features in the sample library (i.e., the second, third, and fourth data distribution differences). These data distribution differences are quantified using a distribution metric based on Wasserstein distance. Second, it determines the first degree of drift by analyzing the performance changes of the downstream state-aware model within a preset sliding time window (i.e., the first time period), such as a decrease in classification accuracy or a continuous increase in the loss function.

[0136] To improve the sensitivity and robustness of drift detection, a multi-window, multi-scale drift detection framework is introduced when determining the second drift level: a short time window (hourly) is used to capture sudden drifts, and a long time window (weekly) is used to identify gradual evolutionary trends. Simultaneously, an isolated forest can be used to detect outlier patterns in the feature space, determining the outlier density corresponding to multiple subsets of selected features obtained by partitioning multiple selection features, and then determining the third drift level. When the outlier density continuously increases, it is identified as potential distribution drift. The first, second, and third drift levels can be combined to determine the comprehensive drift level, and drift detection is performed based on the comprehensive drift level. Once a significant drift phenomenon is detected, the system will promptly trigger an early warning mechanism, realizing online detection and root cause analysis of concept drift.

[0137] The second, third, and fourth data distribution differences can be determined in the following way:

[0138]

[0139] in, At time t, the first The data distribution differences between multiple filtering features and historical features within a time window (including second, third, and fourth data distribution differences). For the first The distribution of filtering data for multiple filtering features within a given time window, where w is the distribution of the filtering data for the first filtering feature. The size of a time window, This represents the distribution of historical data.

[0140] Building upon drift detection, root cause analysis is further conducted by leveraging the simulation capabilities of digital twins. For detected anomalies, the twin is used to perform comparative simulations of different operating scenarios to help determine the true cause of the drift. When the winding temperature prediction result remains consistently high, the twin can simulate scenarios such as decreased cooling efficiency, load changes, or increased ambient temperature to differentiate the impact of equipment malfunctions on external environmental changes.

[0141] At the root cause analysis level, a Bayesian causal graph-based inference model is constructed. The observed data drift levels (including first, second, and third drift levels) are used as evidence nodes. A belief propagation algorithm is used to infer the most likely root cause node (i.e., the target source), such as a decrease in cooling system fan speed, a decline in oil pump efficiency, changes in load levels, or an increase in the ambient temperature baseline. The posterior probability of each potential factor is quantified. The digital twin has a built-in parameter inversion and calibration engine that automatically corrects uncertain parameters in the inference model based on observed data, gradually approximating the behavior of the real equipment and achieving synchronous evolution between the "entity" and the "twin."

[0142] To investigate the adaptive enhancement and correction strategy and the elastic forgetting mechanism, based on the detection results of concept drift and root cause analysis, the adaptive enhancement and correction strategy module dynamically adjusts the focus of feature enhancement. When a gradual change in a known pattern is detected, the enhancement and correction strategy focuses on simulating this evolutionary process to correct the selected features. When a completely new unknown pattern is identified, a conservative correction strategy is adopted, relying more on the selected features from real-world data and high-fidelity simulation data for exploratory enhancement to obtain the target features. Simultaneously, an elastic forgetting mechanism is introduced. Based on the collection time, access frequency, and similarity of historical samples in the sample library to the data distribution of the latest target features, the information value weight is calculated. Long-term low-value, outdated historical samples are automatically archived or discarded, maintaining the effectiveness and dynamic update capability of the sample library.

[0143] By tightly coupling the continuous evolution of the sample library with the update mechanism of the downstream state-aware model, a strong closed loop of data-model collaborative optimization is formed. Each incremental update or cleanup operation of the sample library triggers incremental learning or periodic fine-tuning of the state-aware model, enabling it to quickly adapt to new operating environments. Conversely, changes in the performance of the state-aware model provide feedback for active learning strategies and concept drift evaluation, guiding subsequent sample selection and enhancement strategy adjustments. By establishing a two-way closed-loop mechanism of "data evolution driving model optimization, and model optimization guiding data evolution in turn," the sample library possesses continuous evolution and long-term vitality. The system fully records and logs each evolution process, providing traceable evidence for long-term operational analysis and strategy optimization, ultimately achieving a closed loop of model-data collaborative optimization.

[0144] Step S4: Deployment, verification, and continuous iterative optimization of the sample library. A microservice architecture is adopted to achieve modular decoupling and elastic deployment of the system. A multi-dimensional performance verification and benchmarking system covering effectiveness, efficiency, and evolutionary capabilities is established. Combined with reinforcement learning automatic tuning and privacy-preserving knowledge federation technology, a full-process continuous iterative optimization mechanism is constructed to ensure the long-term stable operation and capability improvement of the sample library system.

[0145] A unified, high-quality, self-evolving sample library adopts a microservice architecture, decoupling modules such as lightweight fusion feature access, virtual and real data management, active learning workflow, sample storage and retrieval, and self-evolving strategy engine, and encapsulating them as independent microservices. Unified data services, model training services, and operation and maintenance monitoring services are provided externally through an API (Application Programming Interface) gateway. This architecture ensures the system's elasticity, scalability, and maintainability, facilitating future feature iterations and upgrades, and enabling the deployment of the sample library.

[0146] Through multi-dimensional performance verification and benchmark testing, a systematic deployment and verification were conducted in a pilot environment for a transformer digital twin. Verification should cover three aspects: first, effectiveness, comparing the accuracy, recall, and timeliness of the fault diagnosis model (a state-aware model for transformer fault diagnosis) before and after using a self-evolving sample library; second, efficiency, evaluating the inference latency and power consumption of the lightweight end-side model, as well as the reduction in manual annotation workload due to the active learning strategy; and third, evolutionary capability, verifying the system's full-cycle performance from detection, early warning, sample updates to model adaptation by simulating sudden changes in the operating environment or introducing new fault modes. Clear benchmark testing indicators were established to compare with traditional static sample libraries and quantify its advantages.

[0147] The problems and experiences discovered during the verification process will form a feedback loop, driving continuous optimization in each of the aforementioned stages. For example, the complexity of the lightweight model will be adjusted based on the actual operation at the end-user side; the uncertainty quantification model will be optimized based on the feedback from the digital twin to the actively learned recommended samples; and the sensitivity of the drift detection algorithm and the threshold of the forgetting mechanism will be fine-tuned based on the effectiveness of concept drift handling, achieving continuous iteration and optimization of the digital twin's feedback. Ultimately, this self-evolving sample library will no longer be an isolated system, but will be deeply embedded in the transformer digital twin, becoming a dynamically growing "intelligent data foundation" for it to achieve accurate perception, intelligent prediction, and autonomous decision-making, promoting the transformation of transformer operation and maintenance models from "traditional preventative" to "state-based predictive" and even "autonomous adaptive."

[0148] The continuous iterative process can be structured as a reinforcement learning problem, using system hyperparameters (lightweight model width, active learning batch size, forgetting threshold, etc.) as the action space, and automatically optimizing the configuration through policy gradient methods. The digital twin is not only a source of virtual operational data but also a validation sandbox, where extreme failure scenarios can be safely simulated, and the response strategies of the sample library and various models can be pre-tested. Risks are assessed before cautious deployment to the physical system. The system incorporates a knowledge federation mechanism, enabling cross-site experience transfer by sharing model update gradients or knowledge distillation among sample libraries from multiple substations, while protecting data privacy, thus accelerating the cold start of sample libraries at new sites. The ultimate goal is to build a self-reflective system capable of automatically generating optimization suggestion reports, providing decision support for human-machine hybrid intelligent operation and maintenance.

[0149] The above-mentioned optional implementation methods achieve at least the following effects: By integrating real operating data and digital twin simulation data, and through the synergistic guidance of three core technologies—lightweight end-side data fusion, virtual-real interactive self-learning, and self-evolving closed-loop management—a multimodal self-evolving sample library system with deep integration of mechanisms and data is constructed, realizing the essential transformation of the sample library from a static data storage container to an intelligent data infrastructure; by actively learning and intelligently labeling, the workload and cost of manual labeling are significantly reduced; by using a self-evolving mechanism, the challenge of concept drift brought about by changes in the operating environment is accurately addressed; the state perception accuracy, data reliability, and long-term evolution support capability of the digital twin are improved; and a solid data foundation is provided for the transformation of transformers from preventive operation and maintenance to predictive and autonomous adaptive operation and maintenance.

[0150] 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.

[0151] This embodiment also provides a transformer sample library expansion device based on digital twins, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the terms "module" and "device" can refer to a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0152] According to an embodiment of this application, an apparatus embodiment for implementing a transformer sample library expansion method based on digital twins is also provided. Figure 4 This is a schematic diagram of a transformer sample library expansion device based on digital twins according to an embodiment of this application, as shown below. Figure 4 As shown, the above-mentioned transformer sample library expansion device based on digital twins includes a data acquisition module 402, a first determination module 404, a second determination module 406, a screening module 408, a drift detection module 410, and an expansion module 412. The device will be described below.

[0153] The data acquisition module 402 is used to acquire multiple sets of target operating data of the target transformer. The multiple sets of target operating data include real operating data and virtual operating data. The real operating data is obtained by data acquisition of the target transformer, and the virtual operating data is obtained based on the digital twin of the target transformer.

[0154] The first determining module 404 is connected to the data acquisition module 402 and is used to determine multiple initial features based on multiple sets of target running data. The multiple initial features include the initial first feature of the real running data and the initial second feature of the virtual running data. The multiple sets of target running data correspond one-to-one with the multiple initial features.

[0155] The second determining module 406, connected to the first determining module 404, is used to determine the first data distribution difference between real running data and virtual running data based on multiple initial features.

[0156] The filtering module 408 is connected to the second determining module 406 and is used to filter multiple initial features when the difference in the first data distribution is greater than a preset difference threshold, thereby obtaining multiple filtered features.

[0157] The drift detection module 410 is connected to the filtering module 408 and is used to perform drift detection based on multiple filtering features to obtain multiple target features.

[0158] The expansion module 412, connected to the drift detection module 410, is used to expand the sample library based on multiple target features.

[0159] This application provides a transformer sample library expansion device based on digital twins. By setting up a data acquisition module 402, a first determination module 404, a second determination module 406, a screening module 408, a drift detection module 410, and an expansion module 412, the device acquires multiple sets of target operating data of the target transformer, including real operating data and virtual operating data generated based on the digital twin. When the first data distribution difference between the real operating data and the virtual operating data is greater than a preset difference threshold, the device expands the sample library based on the initial features obtained from the multiple sets of target operating data. This achieves the technical effect of increasing the number and coverage of transformer sample libraries, thereby solving the technical problems of sample scarcity and insufficient sample coverage in related technologies.

[0160] It should be noted that the above modules can be implemented by software or hardware. For example, for the latter, it can be implemented in the following ways: the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0161] It should be noted that the data acquisition module 402, the first determination module 404, the second determination module 406, the filtering module 408, the drift detection module 410, and the extension module 412 mentioned above correspond to steps S102 to S112 in the embodiments. The instances and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules, as part of the device, can run in a computer terminal.

[0162] It should be noted that the optional or preferred implementation methods of this embodiment can be found in the relevant descriptions in the embodiments, and will not be repeated here.

[0163] The aforementioned transformer sample library expansion device based on digital twins may also include a processor and a memory. The data acquisition module 402, the first determination module 404, the second determination module 406, the screening module 408, the drift detection module 410, the expansion module 412, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize the corresponding functions.

[0164] The processor contains a core that retrieves the corresponding program unit from memory. One or more cores may be configured. Memory may include non-persistent memory in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory includes at least one memory chip.

[0165] This application provides a non-volatile storage medium storing a program that, when executed by a processor, implements a method for expanding a transformer sample library based on digital twins.

[0166] This application provides an electronic device. Figure 5 This is a structural diagram of an electronic device provided according to an embodiment of this application. For example... Figure 5 As shown, the electronic device may include: one or more ( Figure 5 (Only one is shown in the diagram) Processor 502, memory 504, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module, and display. The electronic device includes a processor, memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs the following steps: acquiring multiple sets of target operating data for the target transformer, wherein the multiple sets of target operating data include real operating data and virtual operating data. The real operating data is obtained by data acquisition from the target transformer, and the virtual operating data is obtained based on a digital twin of the target transformer; determining multiple initial features based on the multiple sets of target operating data, wherein the multiple initial features include an initial first feature of the real operating data and an initial second feature of the virtual operating data, with each set of target operating data corresponding one-to-one with the multiple initial features; determining a first data distribution difference between the real operating data and the virtual operating data based on the multiple initial features; filtering the multiple initial features when the first data distribution difference is greater than a preset difference threshold to obtain multiple filtered features; performing drift detection based on the multiple filtered features to obtain multiple target features; and expanding the sample library based on the multiple target features. The device in this document can be a server, PC, etc.

[0167] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: acquiring multiple sets of target operating data for a target transformer, wherein the multiple sets of target operating data include real operating data and virtual operating data, the real operating data being obtained through data acquisition of the target transformer, and the virtual operating data being obtained based on a digital twin of the target transformer; determining multiple initial features based on the multiple sets of target operating data, wherein the multiple initial features include an initial first feature of the real operating data and an initial second feature of the virtual operating data, and the multiple sets of target operating data correspond one-to-one with the multiple initial features; determining a first data distribution difference between the real operating data and the virtual operating data based on the multiple initial features; if the first data distribution difference is greater than a preset difference threshold, filtering the multiple initial features to obtain multiple filtered features; performing drift detection based on the multiple filtered features to obtain multiple target features; and expanding the sample library based on the multiple target features.

[0168] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0173] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0174] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0175] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0176] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0177] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A transformer sample library extension method based on digital twinning, characterized in that, include: Multiple sets of target operating data for the target transformer are acquired, including real operating data and virtual operating data. The real operating data is obtained by data collection from the target transformer, and the virtual operating data is obtained based on a digital twin of the target transformer. Based on the multiple sets of target running data, multiple initial features are determined, wherein the multiple initial features include an initial first feature of the real running data and an initial second feature of the virtual running data, and the multiple sets of target running data correspond one-to-one with the multiple initial features; Based on the aforementioned initial features, a first data distribution difference between the real operating data and the virtual operating data is determined; When the difference in the first data distribution is greater than a preset difference threshold, the multiple initial features are filtered to obtain multiple filtered features; Drift detection is performed based on the multiple screening features to obtain multiple target features; The sample database is expanded based on the aforementioned multiple target features.

2. The method of claim 1, wherein, Based on the multiple sets of target operation data, several initial features are determined, including: For any set of target operation data in the multiple sets of target operation data, if the set of target operation data includes vibration data, acoustic data, infrared thermal imaging data and oil and gas chromatography data, a one-dimensional depthwise separable convolution module of the feature extraction model is used to extract time-domain and frequency-domain features from the vibration data and the acoustic data to obtain initial vibration features and initial acoustic features. The two-dimensional depthwise separable convolutional module of the feature extraction model is used to extract spatial features from the infrared thermal image data to obtain initial infrared features. The fully connected layer of the feature extraction model is used to encode the oil and gas chromatographic data to obtain initial oil and gas features; The multi-head cross-attention module of the feature extraction model is used to fuse the initial vibration features, the initial acoustic features, the initial infrared features, and the initial oil and gas features to obtain the initial features of any set of target operation data. The plurality of initial features are obtained by determining the initial features of any set of target running data.

3. The method of claim 1, wherein, The step of determining the first data distribution difference between the real operating data and the virtual operating data based on the multiple initial features includes: Based on the initial first feature, the true data distribution is determined; Based on the initial second feature, the virtual data distribution is determined; The difference in the first data distribution is determined based on the real data distribution and the virtual data distribution.

4. The method according to claim 1, characterized in that, The process of filtering the multiple initial features yields multiple filtered features, including: Determine the uncertainty index values ​​corresponding to the plurality of initial features; Determine the spatial representativeness index values ​​corresponding to the plurality of initial features; Based on the uncertainty index values ​​corresponding to the plurality of initial features, the uncertainty index value of the initial first feature, and the spatial representativeness index values ​​corresponding to the plurality of initial features, the spatial representativeness index value of the initial first feature, the comprehensive score of the initial first feature is determined. Based on the uncertainty index values ​​corresponding to the plurality of initial features, the uncertainty index value of the initial second feature, the spatial representativeness index value of the initial second feature, and the transferability weight of the initial second feature, the comprehensive score of the initial second feature is determined. The transferability weight is used to characterize the marginal contribution of the virtual operating data corresponding to the initial second feature to improving the performance of the state-aware model. The state-aware model is used to determine the operating state of the target transformer. Based on the combined score of the initial first feature and the combined score of the initial second feature, a preset number of initial features are selected in descending order to obtain multiple candidate features; The candidate features are labeled to obtain the multiple screening features.

5. The method according to any one of claims 1 to 4, characterized in that, The drift detection based on the multiple screening features yields multiple target features, including: Based on the aforementioned multiple screening features, the overall drift degree is obtained; If the overall drift level is less than or equal to a preset drift threshold, it is determined that no data drift has occurred, and the multiple screening features are identified as the multiple target features; or... If the overall drift level is greater than the preset drift threshold, data drift is determined to have occurred, and the multiple screening features are corrected accordingly to obtain the multiple target features.

6. The method according to claim 5, characterized in that, The comprehensive drift degree obtained based on the multiple screening features includes: Based on the model performance data of the state-aware model in the first time period, the first degree of drift is determined; The multiple filtering features are divided into multiple first filtering features, multiple second filtering features, and multiple third filtering features using a sliding window method. The first time window corresponding to the multiple first filtering features is smaller than the second time window corresponding to the multiple second filtering features, and the second time window is smaller than the third time window corresponding to the multiple third filtering features. Based on the plurality of first screening features, the plurality of second screening features, the plurality of third screening features, and the plurality of historical features in the sample library, the second degree of drift is determined; The multiple filtering features are divided into multiple filtering feature subsets according to a preset time window; Determine the outlier density corresponding to each of the plurality of screening feature subsets, wherein the outlier density is used to characterize the degree of data distribution anomaly of the corresponding screening feature subset; The third degree of drift is determined based on the outlier densities corresponding to the multiple subsets of selected features. The overall drift degree is determined based on the first drift degree, the second drift degree, and the third drift degree.

7. The method according to claim 6, characterized in that, The determination of the second drift degree based on the plurality of first screening features, the plurality of second screening features, the plurality of third screening features, and the plurality of historical features in the sample library includes: Determine the first screening data distribution of multiple first screening features, the second screening data distribution of multiple second screening features, the third screening data distribution of multiple third screening features, and the historical data distribution of multiple historical features; Based on the first filtered data distribution and the historical data distribution, the difference in the second data distribution is determined; Based on the second filtered data distribution and the historical data distribution, the third data distribution difference is determined; Based on the third filtered data distribution and the historical data distribution, the fourth data distribution difference is determined; The degree of the second drift is determined based on the second data distribution difference, the third data distribution difference, and the fourth data distribution difference.

8. The method according to claim 5, characterized in that, The step of modifying the plurality of screening features to obtain the plurality of target features includes: Based on the first degree of drift, the second degree of drift, and the third degree of drift, the target source causing the data drift is determined; Based on the overall drift level and the target source, the multiple screening features are modified to obtain the multiple target features.

9. A transformer sample library expansion device based on digital twins, characterized in that, include: The data acquisition module is used to acquire multiple sets of target operating data of the target transformer. The multiple sets of target operating data include real operating data and virtual operating data. The real operating data is obtained by data collection from the target transformer, and the virtual operating data is obtained based on the digital twin of the target transformer. The first determining module is used to determine multiple initial features based on the multiple sets of target running data, wherein the multiple initial features include an initial first feature of the real running data and an initial second feature of the virtual running data, and the multiple sets of target running data correspond one-to-one with the multiple initial features; The second determining module is used to determine a first data distribution difference between the real running data and the virtual running data based on the multiple initial features; The filtering module is used to filter the multiple initial features when the difference in the first data distribution is greater than a preset difference threshold, so as to obtain multiple filtered features; The drift detection module is used to perform drift detection based on the multiple screening features to obtain multiple target features; An extension module is used to expand the sample library based on the multiple target features.

10. The apparatus according to claim 9, characterized in that, The first determining module includes: The first feature extraction unit is used to extract time-domain and frequency-domain features from any one set of target operation data in the plurality of target operation data, where the target operation data includes vibration data, acoustic data, infrared thermal imaging data and oil and gas chromatography data, using a one-dimensional depthwise separable convolution module of the feature extraction model to obtain initial vibration features and initial acoustic features. The second feature extraction unit is used to extract spatial features from the infrared thermal image data using the two-dimensional depthwise separable convolution module of the feature extraction model to obtain initial infrared features. The encoding unit is used to encode the oil and gas chromatographic data using the fully connected layer of the feature extraction model to obtain initial oil and gas features; The feature fusion unit is used to perform feature fusion on the initial vibration features, the initial acoustic features, the initial infrared features, and the initial oil and gas features using the multi-head cross-attention module of the feature extraction model to obtain the initial features of any set of target operation data; Multiple initial feature determination units are used to obtain the multiple initial features by determining the initial features of any set of target running data.

11. The apparatus according to claim 9, characterized in that, The second determining module includes: The first determining unit is used to determine the true data distribution based on the initial first feature; The second determining unit is used to determine the virtual data distribution based on the initial second feature; The third determining unit is used to determine the difference in the first data distribution based on the real data distribution and the virtual data distribution.

12. The apparatus according to claim 9, characterized in that, The filtering module includes: The fourth determining unit is used to determine the uncertainty index values ​​corresponding to the plurality of initial features respectively; The fifth determining unit is used to determine the spatial representativeness index values ​​corresponding to the plurality of initial features respectively; The sixth determining unit is used to determine the comprehensive score of the initial first feature based on the uncertainty index value of the initial first feature among the uncertainty index values ​​corresponding to the plurality of initial features, and the spatial representative index value of the initial first feature among the spatial representative index values ​​corresponding to the plurality of initial features. The seventh determining unit is used to determine the comprehensive score of the initial second feature based on the uncertainty index value of the initial second feature among the uncertainty index values ​​corresponding to the plurality of initial features, the spatial representative index value of the initial second feature among the spatial representative index values ​​corresponding to the plurality of initial features, and the transferability weight of the initial second feature. The transferability weight is used to characterize the marginal contribution of the virtual operating data corresponding to the initial second feature to improving the performance of the state-aware model. The state-aware model is used to determine the operating state of the target transformer. The filtering unit is used to filter out a preset number of initial features in descending order based on the comprehensive score of the initial first feature and the comprehensive score of the initial second feature, thereby obtaining multiple candidate features; The annotation unit is used to annotate the multiple candidate features to obtain the multiple filtered features.

13. The apparatus according to claim 9, characterized in that, The drift detection module includes: A comprehensive drift degree determination unit is used to obtain the comprehensive drift degree based on the multiple screening features; The eighth determining unit is configured to, when the overall drift level is less than or equal to a preset drift threshold, determine that no data drift has occurred, and determine the plurality of filtering features as the plurality of target features; or, The ninth determining unit is used to determine that data drift has occurred when the overall drift level is greater than the preset drift threshold, and to correct the plurality of screening features to obtain the plurality of target features.

14. The apparatus according to claim 13, characterized in that, The comprehensive drift degree determination unit includes: The tenth determining unit is used to determine the first degree of drift based on the model performance data of the state-aware model in the first time period; The first division unit is used to divide the plurality of filtering features into a plurality of first filtering features, a plurality of second filtering features and a plurality of third filtering features in a sliding window manner, wherein the first time window corresponding to the plurality of first filtering features is smaller than the second time window corresponding to the plurality of second filtering features, and the second time window is smaller than the third time window corresponding to the plurality of third filtering features. The eleventh determining unit is used to determine the second degree of drift based on the plurality of first screening features, the plurality of second screening features, the plurality of third screening features, and the plurality of historical features in the sample library; The second partitioning unit is used to divide the multiple filtering features into multiple filtering feature subsets according to a preset time window; The twelfth determining unit is used to determine the outlier density corresponding to each of the plurality of screening feature subsets, wherein the outlier density is used to characterize the degree of data distribution anomaly of the corresponding screening feature subset; The thirteenth determining unit is used to determine the third drift degree based on the outlier densities corresponding to the multiple screening feature subsets respectively; The fourteenth determining unit is used to determine the comprehensive drift degree based on the first drift degree, the second drift degree, and the third drift degree.

15. The apparatus according to claim 14, characterized in that, The eleventh determining unit includes: The fifteenth determining unit is used to determine the first screening data distribution of multiple first screening features, the second screening data distribution of multiple second screening features, the third screening data distribution of multiple third screening features, and the historical data distribution of multiple historical features. The sixteenth determining unit is used to determine the difference in the second data distribution based on the first filtered data distribution and the historical data distribution; The seventeenth determining unit is used to determine the third data distribution difference based on the second filtered data distribution and the historical data distribution; The eighteenth determining unit is used to determine the fourth data distribution difference based on the third filtered data distribution and the historical data distribution; The nineteenth determining unit is used to determine the second degree of drift based on the second data distribution difference, the third data distribution difference, and the fourth data distribution difference.

16. The apparatus according to claim 13, characterized in that, The ninth determining unit includes: The twentieth determining unit is used to determine the target source causing the data drift based on the first drift degree, the second drift degree, and the third drift degree. The correction unit is used to correct the plurality of screening features based on the overall drift degree and the target source, respectively, to obtain the plurality of target features.

17. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores multiple instructions adapted for loading and execution by a processor of the transformer sample library expansion method based on digital twins as described in any one of claims 1 to 8.

Images