Transformer insulation state evaluation method based on artificial intelligence multi-parameter fusion perception

Abstract

The application discloses a transformer insulation state evaluation method based on artificial intelligence multi-parameter fusion perception, relates to the technical field of electrical equipment detection, and comprises the following steps: step one, obtaining multi-modal monitoring data of the transformer and constructing a dynamic graph structure; wherein the multi-modal monitoring data comprises oil chromatogram data, partial discharge data and thermal image data; step two, coding the dynamic graph structure through an encoder embedding technology, performing feature aggregation and evolution modeling in the time and space dimensions based on the coded dynamic graph structure by using graph convolution and time convolution technologies, inputting preset transformer insulation positive and negative samples into the constructed model to perform unsupervised learning of InfoNCE loss, obtaining the state manifold of the normal transformer, and performing time-space alignment and abnormality differentiation of the multi-modal data of the state manifold; and step three, introducing a Bayesian weight after the time-space alignment and abnormality differentiation of the multi-modal data, and outputting the probability distribution of the insulation state embedding.

Description

Technical Field

[0001] This invention relates to the technical field of electrical equipment testing, specifically to a method for evaluating the insulation status of transformers based on artificial intelligence-based multi-parameter fusion perception. Background Technology

[0002] As a core hub in power transmission and transformation systems, the operational reliability of power transformers directly determines the safety and stability of the power grid and the quality of power supply. During long-term operation, the internal insulation system of transformers is continuously subjected to the coupling effects of multiple physical fields such as electricity, heat, machinery, and chemicals, making it highly susceptible to typical deterioration problems such as aging, moisture absorption, and partial discharge. These are not only key causes of insulation performance degradation but also the primary root cause of sudden transformer failures. Furthermore, insulation aging usually stems from the cumulative effect of long-term thermal and electrical stress, leading to a significant decrease in the dielectric strength and mechanical toughness of the oil-paper insulation material. At the same time, moisture intrusion into the insulation system severely weakens its insulation resistance and breakdown voltage. Partial discharge caused by early insulation defects is an important indicator. Although the initial energy is weak, it has the ability to continuously erode the insulation structure. If it is not identified and intervened in time, it can easily evolve into through discharge or even insulation breakdown accidents.

[0003] Traditional transformer insulation condition assessment has long relied on single indicators, such as furfural content or dielectric loss, which are insufficient to accurately reflect the nonlinear evolution characteristics and multi-physics coupling mechanisms during insulation degradation. Especially for slowly aging insulation systems, existing AI models face significant challenges in acquiring large numbers of well-labeled fault and insulation condition samples. This is not only costly and time-consuming, but also suffers from sample scarcity due to the rarity of actual degradation events, limiting the training quality and generalization ability of supervised learning models. Furthermore, dissolved gas components in transformer oil, partial discharge signals, and infrared thermography exhibit heterogeneity in terms of dimension, sampling frequency, and physical semantics. These cannot be simply pieced together and fused; they require inherent spatiotemporal correlations and physical coupling laws, leading to distorted feature representations and low information utilization efficiency. Existing AI models generally output black-box prediction results, lacking quantitative representation of assessment confidence and verification of reliability for power grid operation and maintenance decisions. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: Step 1: Acquire multimodal monitoring data of the transformer and construct a dynamic graph structure; the multimodal monitoring data includes oil chromatography data, partial discharge data and thermal imaging data; Step 2: The dynamic graph structure is encoded by the encoder extraction embedding technology. Based on the encoded dynamic graph structure, the spatiotemporal features are aggregated and evolved using graph convolution and temporal convolution techniques. The model is input with the pre-set positive and negative transformer insulation samples and InfoNCE loss is used for unsupervised learning to obtain the state manifold of the normal transformer. The state manifold is then spatiotemporally aligned with multimodal data and anomaly is distinguished. Step 3: After spatiotemporal alignment and anomaly differentiation of multimodal data, Bayesian weights are introduced to output the probability distribution of insulation state embedding. The cognitive uncertainty and data uncertainty of the probability distribution are decomposed and quantified. When the total uncertainty or embedding deviation exceeds the threshold, an insulation state warning is triggered.

[0005] Furthermore, the process of constructing the dynamic graph structure is as follows: The transformer multimodal monitoring data is acquired, and each parameter in the transformer multimodal monitoring data is used as a graph node. The initial edge weights between graph nodes are defined based on the transformer electrothermal-chemical coupling mechanism. The online dynamic adjustment edge connections and weights of the transformer operating conditions are obtained and dynamically adjusted according to the operating conditions to construct the physical coupling relationship of multiple parameters.

[0006] Furthermore, the process of encoding the dynamic graph structure using the encoder-extracted embedding technique is as follows: Based on the heterogeneous data characteristics of different types of graph nodes in a dynamic graph structure, a dedicated encoder is used to extract basic features and transform them into initial node embeddings of a unified dimension, resulting in a single-modal initial embedding. Based on the topological structure of the dynamic graph, the single-modal initial embeddings are fused into a global graph embedding. The temporal-aware embeddings of all graph nodes are weighted and pooled, and batch standardization is performed to eliminate the differences in feature distribution under different working conditions, outputting a dynamic graph global embedding with dimension D. This embedding simultaneously contains multimodal features, physical associations, and spatiotemporal evolution information. Through global pooling and standardization, the spatiotemporally fused node-level embeddings are aggregated into a graph-level global embedding.

[0007] Furthermore, the global graph embedding includes spatial association and temporal evolution: Spatial association: Using the edge weights of the dynamic graph structure as physical priors, the information of neighboring nodes is aggregated through a graph attention network. For each graph node, attention weights are calculated based on the initial embedding similarity between it and the neighboring graph nodes to strengthen the information transmission between graph nodes with strong physical associations. After multiple layers of graph convolution, the embedding of each node is updated to a spatially perceptual embedding that integrates its own features and the association features of neighboring nodes. Temporal Evolution: For the topological changes of the dynamic graph over time, a gated recurrent unit is used to capture the temporal evolution. The spatially aware embedding of continuous time steps is used as input. Through the update gate and reset gate of the GRU, redundant information is dynamically filtered and key changes are retained. The output is a temporally aware embedding that includes historical state dependencies.

[0008] Furthermore, the process of performing spatiotemporal feature aggregation and evolutionary modeling is as follows: Using the encoded dynamic graph node embedding as input, and based on the physical guidance edge weight topology, graph attention convolution is used to aggregate the association features between nodes. Based on the spatially aware embedding of continuous time steps, the temporal evolution law is captured by the causal extended temporal convolution network. The temporal features output by the temporal convolution and the spatial features output by the graph convolution are fused element-wise to generate a global feature vector that contains multiple physical coupling relationships and temporal evolution laws for model construction.

[0009] Furthermore, the process of obtaining the state manifold of a normal transformer is as follows: The positive and negative sample pairs are input into the model's contrastive learning head. Cosine similarity is used to calculate the similarity between the two augmented views in the positive sample pair, as well as the similarity between each positive sample and all negative samples in the negative sample set, quantifying the feature association strength between samples. The InfoNCE loss is used as the objective function, and the model parameters are optimized through gradient descent. The number of training iterations is set, and after each iteration, the average similarity of positive samples and the difference in similarity between positive and negative samples on the validation set are calculated. When the indicators are stable for 10 consecutive iterations and the loss value tends to level off, the model is considered to have converged. After the model converges, the encoding embeddings of all normal state samples in the training set are extracted, and their feature distribution is fitted by a Gaussian mixture model to obtain the normal state manifold.

[0010] Furthermore, the process of performing multimodal data spatiotemporal alignment and anomaly differentiation on the state manifold is as follows: Spatiotemporal alignment of multimodal data is achieved through time axis interpolation calibration and spatial correlation mapping, generating new sample embeddings. The deviation from the normal manifold is quantified by Mahalanobis distance, and normal, slight and severe anomalies are divided by the 3σ principle. The contribution of each modal anomaly is calculated, and the anomaly source is located by combining the insulation degradation mechanism.

[0011] Furthermore, the process of embedding the probability distribution of the output insulation state is as follows: After spatiotemporal alignment and anomaly differentiation of multimodal data, the core model for generating feature embeddings is modified by replacing the original fixed parameters with probability distribution weights. Multiple independent sets of insulation state feature embeddings are generated by randomly selecting values ​​from the probability distribution of Bayesian weights. Based on multiple sets of effective embedding samples, a probability distribution model of insulation state is constructed, and two sources of uncertainty are distinguished to form a complete probability assessment result.

[0012] Furthermore, the process of decomposing and quantifying the cognitive uncertainty and data uncertainty of the probability distribution is as follows: By quantifying the discreteness of the sampling results from the Bayesian model and the feature stability under noise interference, cognitive uncertainty and data uncertainty are added together by weight to obtain a total uncertainty value of 0 to 1. The deviation of the current average embedding from the normal state manifold is calculated, and a warning judgment is made based on the threshold.

[0013] The transformer insulation status assessment method based on artificial intelligence multi-parameter fusion perception provided by this invention has the following beneficial effects: (1) This invention constructs a dynamic graph structure based on the electrothermal-chemical coupling mechanism, and uses the parameters in the oil chromatography, partial discharge and thermal imaging data of transformer multimodal monitoring as graph nodes. The initial edge weights between graph nodes are defined according to physical priors and dynamically adjusted according to the operating conditions. This realizes the modeling of the physical coupling relationship of multiple parameters in the insulation system, overcomes the problem of feature expression distortion caused by the large differences in multimodal data dimension, sampling rate and physical semantics in traditional methods, and improves the feature expression ability and information utilization efficiency of insulation status assessment.

[0014] (2) This invention uses InfoNCE loss unsupervised learning to construct a normal state manifold, and achieves accurate quantitative assessment of insulation degradation state through multimodal data spatiotemporal alignment and anomaly differentiation mechanism. It does not require a large number of labeled samples, which solves the bottleneck problem of scarce real degradation samples and poor generalization ability of supervised learning model in slow aging insulation system. At the same time, it objectively classifies the degree of insulation degradation by dividing the anomaly level through Mahalanobis distance quantization and 3σ principle.

[0015] (3) This invention introduces the probability distribution of the insulation state embedding by Bayesian weight output, decomposes and quantifies the cognitive uncertainty and data uncertainty of the probability distribution, and provides credibility support for operation and maintenance decision-making; when the total uncertainty or embedding deviation exceeds the threshold, the insulation state warning is triggered, so that the evaluation result is transformed from a black box output to a credible evaluation with probabilistic interpretation, solving the key problem that the existing AI model lacks confidence quantification and is difficult to support operation and maintenance decision-making, and providing an interpretable decision basis for transformer insulation state evaluation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This application provides a method for evaluating the insulation status of transformers based on artificial intelligence multi-parameter fusion perception. The method includes: Step 1: Acquire multimodal monitoring data of the transformer and construct a dynamic graph structure; the multimodal monitoring data includes oil chromatography data, partial discharge data and thermal imaging data; Acquire multi-mode monitoring data of transformers: Oil chromatography data acquisition is achieved by using dissolved gas analysis sensors, such as gas chromatographs or online DGA monitoring devices, to take oil samples from the transformer oil tank or the main body and detect the concentration of characteristic dissolved gases in the oil, including H2, CH4, C2H2, C2H4, C2H6, etc. The online monitoring device automatically samples and analyzes according to a preset cycle, while the laboratory requires manual sampling and testing, and outputs time series data of each gas concentration.

[0019] Partial discharge data acquisition uses partial discharge sensors, such as ultra-high frequency sensors and ultrasonic sensors, which are installed in transformer tanks, bushings, and other parts to collect electrical or acoustic signals generated by partial discharge.

[0020] UHF sensors achieve non-contact monitoring by capturing ultra-high frequency electromagnetic waves of partial discharge (300MHz to 3GHz), while AE sensors monitor by detecting mechanical vibration waves of discharge (20kHz to 200kHz). The data sampling rate is typically at the MHz level, and the sensors output time-series data such as partial discharge quantity and discharge phase.

[0021] Infrared thermal imagers are used to acquire thermal data by scanning the transformer body, bushings, joints and other parts with infrared thermal imaging to collect temperature distribution images on the surface of the equipment. Fixed thermal imagers automatically take pictures every 15 minutes at fixed time intervals, while portable thermal imagers are taken by maintenance personnel during regular inspections, and output infrared thermal image data containing features such as temperature mean, temperature variance and hot spot location.

[0022] By synchronously collecting multimodal data or aligning it with timestamps, a multi-source dataset that matches in the time dimension is formed.

[0023] Constructed as a dynamic graph structure: The parameters in the transformer multimodal monitoring data are used as graph nodes, including oil chromatography nodes such as H2 concentration nodes, CH4 concentration nodes, and C2H2 concentration nodes. Each node corresponds to the concentration time series data of a characteristic gas. For partial discharge nodes, partial discharge quantity nodes and discharge phase distribution nodes are set to carry the quantitative characteristics of partial discharge signals. The thermal imaging nodes include thermal imaging temperature mean nodes, temperature variance nodes, hot spot location feature nodes, etc., to characterize the key information of infrared thermal imaging.

[0024] Based on the initial edge weights between nodes defined by the electrothermal-chemical coupling mechanism of transformers, and according to the gas generation mechanism of oil-immersed transformer faults, arc-induced partial discharge, with its high energy density, causes the C-C bonds in the oil to break at 800℃~1200℃, generating C2H2. Experiments show that C2H2 is a characteristic gas of partial discharge, especially arc discharge, and its concentration is strongly correlated with the intensity of partial discharge. Since partial discharge intensities and C2H2 concentration nodes trigger C2H2 generation, their edge weights are initially set to high values. The heat generated by the transformer windings is transferred to the oil temperature through thermal conduction and oil convection, and there is a clear thermal coupling relationship between the two. According to industry standards and thermal analysis, the winding temperature is usually higher than the oil temperature, and the oil temperature is an important factor affecting the winding temperature. For example, the temperature difference between the top oil temperature and the winding hot spot temperature in a natural oil circulation transformer is about 23°C. Due to the strong correlation between the oil temperature node and the winding temperature node, their edge weight can be initially set to 0.7. For parameters like H2 concentration nodes and average thermal imaging temperature nodes, which lack direct electrothermal-chemical coupling, their correlation stems only from weak correlations at the data level, such as accidental overlaps in operating conditions, lacking physical mechanism support. Quantitative studies of multi-parameter correlations using methods like grey relational analysis show that the weight of such weak correlations is typically much lower than that of physically coupled parameters. For nodes without direct physical coupling, such as H2 concentration nodes and average thermal imaging temperature nodes, the initial edge weight is set to 0.1.

[0025] It should be noted that the specific values ​​are related to the transformer model. This invention is mainly based on oil-immersed power transformers with voltage levels of 110kV and above, such as the common models SZ22-12500 / 110 and SFPZ9-120000 / 110.

[0026] The system dynamically adjusts edge connections and weights based on the transformer's operating conditions. Under conditions of sudden load changes, when the transformer load change rate exceeds a preset threshold, such as 10% / minute, the edge weights between nodes related to load current, winding temperature, and oil temperature are increased, for example, from 0.6 to 0.8, to reflect the strengthening effect of load changes on temperature correlation. Under conditions of sudden temperature changes, when the oil temperature change rate exceeds 5℃ / hour, the edge weights of nodes related to oil temperature and oil gas solubility are increased to reflect the influence of temperature on the solubility characteristics of gases in oil.

[0027] By transforming multimodal monitoring data into a graph structure that dynamically evolves with operating conditions, accurate modeling of the physical coupling relationships of multiple parameters in a transformer can be achieved.

[0028] Step 2: The dynamic graph structure is encoded by the encoder extraction embedding technology. Based on the encoded dynamic graph structure, the spatiotemporal features are aggregated and evolved using graph convolution and temporal convolution techniques. The model is input with the pre-set positive and negative transformer insulation samples and InfoNCE loss is used for unsupervised learning to obtain the state manifold of the normal transformer. The state manifold is then spatiotemporally aligned with multimodal data and anomaly is distinguished. The encoding process of dynamic graph structure: To address the heterogeneous data characteristics of different types of nodes in dynamic graph structures, such as oil chromatography, partial discharge, and infrared thermography, a dedicated encoder is used to extract basic features and transform them into initial node embeddings of a unified dimension.

[0029] When encoding time-series nodes for oil chromatography and dielectric response, which are low-sampling-rate time-series signals such as daily oil gas concentration and quarterly dielectric loss tangent, a time-series feature enhancement and 1D convolutional encoder are used. First, the original time-series data is statistically analyzed using a sliding window to calculate the mean, variance, and trend slope, thereby enhancing local time-series features. Then, the data is input into a 1D convolutional neural network, which uses multiple convolutional kernels with sizes of 3, 5, and 7 to capture short-term fluctuations and long-term trends at different time scales. After the dimensions are compressed by pooling layers, the output is a time-series node embedding with dimension d.

[0030] The encoding of high-frequency signal nodes in partial discharge uses a time-frequency conversion and attention mechanism encoder. The time-domain waveform is converted into a two-dimensional time-frequency matrix through short-time Fourier transform, preserving the frequency distribution and temporal evolution characteristics of the discharge signal. The input is a convolutional network with a self-attention mechanism, which focuses on the regions with significant discharge characteristics, such as the time-frequency points corresponding to the pulse peak, through attention weights. These are then mapped to partial discharge node embeddings of dimension d through a fully connected layer.

[0031] In the process of encoding infrared thermal images, the node data is a temperature distribution image. Image segmentation and a visual ViT encoder are used. First, the thermal image is segmented into regions, such as the body, sleeve, and connector regions. Prior features such as temperature extremes and gradients of each region are extracted. The segmented image blocks are input into ViT, and the temperature correlation between regions, such as the temperature difference between the sleeve and the body, is captured through multi-head self-attention. After global features are aggregated by clstoken, the thermal image node embedding of dimension d is output.

[0032] Spatiotemporal graph fusion encoding: Based on the topology of dynamic graphs, node and edge weights, the initial embedding of a single mode is fused into a global graph embedding that includes spatial association and temporal evolution, modeling the dynamic association and temporal evolution between nodes; In the spatial dimension, through physically guided graph attention convolution, with the edge weights of the dynamic graph structure as the physical prior, the graph attention network GAT aggregates the information of neighboring nodes. For each graph node, the attention weight is calculated based on its initial embedding similarity with neighboring graph nodes and the physical edge weights, thereby strengthening the information transmission between physically related graph nodes, such as increasing the attention weight of the partial discharge node for the C2H2 node. After multi-layer graph convolution, the embedding of each node is updated to a spatially perceptual embedding that integrates its own features and the features associated with neighboring nodes, preserving the physical coupling relationships from partial discharge to gas generation and from temperature to accelerated aging.

[0033] In the time dimension, a gated recurrent encoding with temporal differential awareness is used to capture temporal evolution of the dynamic graph as the topology changes and edge weights adjust over time. The spatial awareness embedding of continuous time steps is used as input, and redundant information, such as repetitive features under stable operating conditions, is dynamically filtered through the update gate and reset gate of the GRU, while retaining key changes, such as temperature node embedding changes during load mutations. The output contains a temporal awareness embedding that depends on historical states, reflecting the temporal correlation from short-term state fluctuations to long-term insulation degradation trends.

[0034] Animated graph embedding output: Global feature aggregation and dimensional unification are achieved through global pooling and standardization. The spatiotemporally fused node-level embeddings are aggregated into graph-level global embeddings. The temporal-aware embeddings of all graph nodes are weighted and pooled, with the weight being the degree centrality of the node in the graph, highlighting key parameter nodes. Batch standardization eliminates the differences in feature distribution under different working conditions, and the output dimension is D. For example, a dynamic graph global embedding with D=256 contains three layers of information: multimodal features, physical associations, and spatiotemporal evolution. Through unique coding adaptation heterogeneity, spatiotemporal fusion modeling associations, and the hierarchical relationship of consistent global aggregation output, the unique information of each modality data is preserved, and the physical coupling and temporal changes between nodes in the dynamic graph are captured.

[0035] Spatiotemporal feature aggregation and evolution modeling: Spatial dimension feature aggregation is achieved through graph convolution, and temporal dimension evolution modeling is achieved through temporal convolution, forming spatiotemporal fusion features.

[0036] Spatial dimension feature aggregation takes the encoded dynamic graph node embedding as input. Based on the physical guided edge weight topology, it realizes the aggregation of inter-node correlation features through graph attention convolution. Combined with the real-time edge weight of the dynamic graph, it reflects the physical correlation strength under the current working condition. The original attention weight between nodes is corrected by embedding similarity calculation to strengthen the information transmission weight of nodes with strong physical correlation, such as partial discharge and feature gas, temperature and dielectric loss. Through multi-layer graph convolution iteration, the first layer of convolution focuses on the feature interaction of directly adjacent nodes, such as oil temperature and winding temperature, and outputs intermediate embeddings containing local correlation. The second layer of convolution expands to second-order adjacent nodes, such as partial discharge, feature gas and dielectric loss, to capture indirect coupling relationships. Through 3 to 4 layers of convolution, the single-node features are fused into a spatially perceptual embedding of full-graph physical correlation.

[0037] Based on spatially perceptual embedding with continuous time steps, a causal dilated temporal convolutional network is used to capture temporal evolution patterns. Different dilation rates, such as 1, 2, and 4, are used to capture state fluctuations within one hour in the short term, trend changes within one day in the medium term, and degradation tendencies within one week in the long term, respectively, covering the multi-timescale features of transformer insulation state evolution. Each temporal convolutional layer retains the original temporal information through residual paths, avoiding long-term reliance on feature decay in modeling, and focusing on enhancing the temporal features of abnormal moments, such as sudden increases in partial discharge and jumps in gas concentration. For the differences in sampling rates of multimodal data, such as oil chromatography once a day and partial discharge 10 times per second, adaptive temporal interpolation is used to unify the spatial embeddings of different time lengths into a fixed time interval sequence, such as one sample every 15 minutes, to ensure the consistency of the input of temporal convolution.

[0038] The temporal features output by temporal convolution and the spatial features output by graph convolution are fused element-wise with weights dynamically allocated based on the spatiotemporal correlation strength under the current operating conditions, generating a global feature vector that simultaneously contains multiple physical coupling relationships and temporal evolution laws.

[0039] Obtain the state manifold of a normal transformer: By constructing positive and negative sample pairs and training the model based on InfoNCE loss, the unsupervised learning manifold of the normal transformer state is obtained.

[0040] Positive sample pair generation involves performing two different data augmentation operations, spatial augmentation and temporal augmentation, on the global embedding of the dynamic graph encoded at the same time to obtain two augmented views, which form a positive sample pair, so that although the two are different in form, they both come from the same normal state. Spatial enhancement: Randomly mask 10%–15% of non-critical node features, such as minor gas concentration nodes; Time augmentation: Apply minute time-order perturbations to the continuous time-step embedding, such as ±1-step time offset; Negative sample pairs are generated by selecting three types of negative samples: normal state embeddings from different time windows, such as samples with an interval of more than 24 hours; suspicious state embeddings marked in historical records, i.e., no clear fault but large parameter fluctuations; and invalid embeddings generated by randomly shuffling the connection relationships of graph nodes. Each positive sample pair is matched with 15 to 20 negative samples to form a negative sample pair set, which is significantly different from the positive samples in terms of state characteristics.

[0041] Unsupervised training using InfoNCE loss is employed, where positive and negative sample pairs are input into the model's comparative learning head. Cosine similarity is used to calculate the similarity between two augmented views in a positive sample pair, as well as the similarity between each positive sample and all negative samples in the negative sample set, quantifying the strength of feature associations between samples. Using InfoNCE loss as the objective function, gradient descent is used to optimize the model parameters. The loss function is designed to maximize the similarity between positive sample pairs and minimize the similarity between positive and negative samples. This means the model learns the feature distribution pattern that different augmented views of the same normal state should cluster together and be separated from samples of other states. The number of training iterations is set, such as 200-300 rounds. After each iteration, the average similarity of positive samples on the validation set is calculated, with a target of ≥0.85 and a target difference of ≥0.6 for the similarity between positive and negative samples. The model is considered converged when the metrics stabilize for 10 consecutive rounds and the loss value tends to level off.

[0042] After the model converges, the encoded embeddings of all normal state samples in the training set are extracted, and their feature distributions are fitted by a Gaussian mixture model to obtain the normal state manifold. The manifold uses the mean vector to represent the central features of the normal state and the covariance matrix to represent the fluctuation range of the normal state, forming a normal state feature cluster region, which provides a benchmark for subsequent anomaly detection: when a new sample embedding falls within the range of the manifold, it is determined to be a normal state; if it exceeds the range, it is marked as an anomaly.

[0043] Spatiotemporal alignment and anomaly differentiation of multimodal data: Based on a trained normal state manifold, this study addresses the spatiotemporal misalignment problem of multimodal data and performs accurate anomaly differentiation for feature matching, including spatiotemporal alignment of multimodal data, anomaly differentiation based on state manifold, and anomaly modality attribution.

[0044] Spatiotemporal alignment of multimodal data: To address the inconsistencies in sampling rates, timestamps, and feature dimensions of multimodal data such as oil chromatography, partial discharge, and infrared thermography, alignment is achieved through time axis calibration and spatial correlation mapping to ensure that the input features match the dimensions of the normal state manifold. S201: Analyze the original sampling cycles of each mode, such as partial discharge 10kHz, oil chromatography once / day, and infrared thermography once / 15 minutes. Using the mode with the highest sampling rate, such as partial discharge, as the benchmark, adaptive interpolation is performed to complete the low sampling rate modes. Oil chromatography data is generated by linear interpolation to generate pseudo-sampling points at 15 minutes / time, which retains the daily average trend while matching the thermography cycle. Thermography data is generated by neighbor interpolation to generate pseudo-sampling points at 1 minute / time, which fits the time density of partial discharge. S202: Based on the GPS timestamp of the transformer operation and maintenance system, perform timestamp calibration on all modal data to eliminate equipment clock deviation, such as the time difference between the partial discharge sensor and the thermal imager ≤ 1 second, to ensure that multimodal data in the same time slice correspond to the same operating time. S203: Slice the aligned time series data into fixed time windows, such as 15 minutes / window. Each window outputs a set of spatiotemporal samples containing all modal features, consistent with the time dimension of the normal state manifold, such as 15 minutes / sample.

[0045] Spatial correlation mapping is based on the physical edge weights of the dynamic graph, such as the correlation weight between partial discharge and C2H2, and the correlation weight between oil temperature and dielectric loss. A multimodal feature correlation matrix is ​​constructed. For each spatiotemporal sample, weakly correlated modal features, such as infrared thermography and H2 concentration, with a correlation weight < 0.3, are fused using features such as weighted summation. Strongly correlated modal features, such as partial discharge and C2H2, with a correlation weight ≥ 0.6, retain independent dimensions to avoid redundancy. The fused multimodal features are mapped to the same dimension as the normal state manifold, such as 256 dimensions, through a fully connected layer to ensure consistency of feature dimensions during subsequent comparisons.

[0046] Anomaly differentiation based on state manifold: Based on the mean vector and covariance matrix of the normal manifold fitted by the Gaussian mixture model, anomaly differentiation is performed through feature embedding alignment, deviation calculation, and threshold determination. The aligned multimodal spatiotemporal samples are input into the pre-trained encoder, including modality-specific encoding and spatiotemporal graph fusion encoding, and output new sample feature embeddings of the same dimension as the normal manifold, so that its feature distribution rules are consistent with those of the normal manifold, such as temporal trends and spatial correlation weight matching. Mahalanobis distance is used to calculate the deviation of the new sample embedding from the normal manifold. Using the mean vector of the normal manifold as the center, the covariance matrix is ​​combined to reflect the fluctuating correlation of features under normal conditions, and the Mahalanobis distance of the new sample embedding to the mean vector is calculated. Simultaneously, the mean cosine similarity between the new sample embedding and all samples within the normal manifold is calculated. ,like The bigger, The smaller the value, the more significant the deviation of the new sample characteristics from the normal manifold.

[0047] Mahalanobis distance: It can eliminate the influence of correlation between features and is more consistent with manifold distribution than Euclidean distance. Therefore, this invention uses Mahalanobis distance to calculate the embedding of new samples into the mean vector.

[0048] It should be noted that when performing multimodal data spatiotemporal alignment and anomaly differentiation on state manifolds, the inconsistency of sampling rate, timestamp, and feature dimension is specifically addressed. The goals and accuracy requirements of data processing at different stages are fundamentally different, and the processing in the early steps cannot meet the fine-grained requirements of state manifold matching. For example, in the early stages, it is only necessary to perform coarse-grained alignment of oil chromatography and partial discharge to ensure that each modal data corresponds to the same runtime, without pursuing extreme accuracy. The feature dimension only needs to be adapted to the definition of dynamic graph nodes. However, in the state manifold matching and anomaly differentiation stages, the feature distribution of the data to be evaluated must be fully adapted to the pre-trained normal state manifold. If the coarsely aligned data from the early stages is directly used, problems such as excessively coarse temporal granularity of oil chromatography data, mismatch of partial discharge feature dimensions, and accumulation of timestamp deviations will occur, leading to significant errors in deviation determination.

[0049] Based on the Mahalanobis distance and similarity of normal samples in the training set, a threshold is set using the 3σ principle: the mean Mahalanobis distance of the normal manifold is [value missing]. Standard deviation is Set anomaly detection threshold The mean similarity of normal samples is Standard deviation is Set a similarity threshold Judgment rule: If ≤ and ≥ If it is determined to be in a normal state; < ≤ and ≤ < If it is determined to be a minor anomaly, such as a characteristic shift caused by short-term operating condition fluctuations; if > and < If it is determined to be a serious abnormality, such as a characteristic mutation caused by insulation deterioration.

[0050] Anomalous modality attribution: After anomaly identification, key anomaly sources are located by calculating modal contribution, avoiding misjudgments caused by multimodal data confusion. For samples identified as anomaly, feature embeddings of each mode are extracted, such as oil chromatography mode embedding and partial discharge mode embedding. The Mahalanobis distance between the single-mode embedding and the corresponding mode embedding of the normal manifold is calculated. Where i is the modal index; according to Where n is the total number of modes, j is the summation index variable used to iterate from the 1st mode to the nth mode, and n is the total number of modes to calculate the anomaly contribution of each mode. The higher its contribution, the more significant the deviation of the mode from the normal state; combined with the transformer insulation degradation mechanism, for high contribution modes, such as... Attribution is performed for values ​​≥0.3: if the contribution of the partial discharge mode is the highest and the contribution of the corresponding C2H2 concentration mode is the second highest, the source of the anomaly is determined to be the risk of insulation breakdown caused by the aggravation of partial discharge; if the contributions of both the oil temperature mode and the dielectric loss mode are high, the source of the anomaly is determined to be the accelerated insulation aging due to the increase in temperature.

[0051] Step 3: After spatiotemporal alignment and anomaly differentiation of multimodal data, Bayesian weights are introduced to output the probability distribution of insulation state embedding. The cognitive uncertainty and data uncertainty of the probability distribution are decomposed and quantified. When the total uncertainty or embedding deviation exceeds the threshold, an insulation state warning is triggered. Probability distribution of insulating state embedding: By introducing Bayesian weights in the feature embedding generation stage, the deterministic features of the insulation state are transformed into a probability distribution form, while simultaneously quantifying the uncertainty in the evaluation: S301: Construct a Bayesian graph encoder to transform fixed weights into probability-distributed weights; after completing the spatiotemporal alignment and anomaly differentiation of multimodal data, reconstruct the core model for generating feature embeddings, namely the graph convolutional and temporal convolutional networks, replacing the original fixed parameters, such as convolutional kernels and attention weights, with probability-distributed weights, so that each parameter corresponds to a possible range of values, rather than a single numerical value: For the weights in the graph convolutional layer that characterize node associations, such as the association strength between partial discharge and characteristic gases, they are set to follow a certain probability distribution, such as a range similar to a normal distribution. The center value of the distribution is initialized based on the physical mechanism. For example, the center value of the association weight between partial discharge and acetylene is set to a relatively high level. The fluctuation range of the distribution, i.e., the standard deviation, is limited to a reasonable range to avoid negative associations, so that the weight changes conform to the electrothermal-chemical coupling law of transformer insulation. For the parameters in the time convolutional layer that capture time-series changes, such as the trend parameter of temperature changing with load, they are also set to a probability distribution. The center value reflects the typical time-series law during normal operation, such as the gradual trend of oil temperature rising with load. The fluctuation range covers reasonable operating condition fluctuations, such as temperature jumps caused by short-term load changes.

[0052] S302: Generate multiple sets of embedded samples through Monte Carlo sampling, randomly select values ​​multiple times from the probability distribution of Bayesian weights to generate multiple sets of independent insulation state feature embeddings, thereby approximating the overall distribution of features, and randomly select 100 to 200 different weight combinations from the probability distribution weights of graph convolution and temporal convolution.

[0053] The spatiotemporally aligned multimodal data, matched with the dimensions of the normal state manifold, were input into models corresponding to 100–200 sets of weights to obtain 100–200 sets of insulation state feature embeddings. Each set of embeddings contains spatiotemporal features fused from multimodal data. At the same time, abnormal embeddings that clearly violate physical common sense, such as embeddings with negative partial discharge values, were removed, and valid samples were retained.

[0054] S303: Based on multiple sets of valid embedded samples, a probability distribution model of the insulation state is constructed, and two sources of uncertainty are distinguished to form a complete probability assessment result: By fitting the effective embedded samples with a Gaussian mixture model, the probability distribution of the insulation state embedding is obtained. This distribution contains multiple feature cluster centers, corresponding to different insulation state subclasses and the feature fluctuation range of each center. It reflects both the typical characteristics of the current state and the possible range of feature changes.

[0055] The dispersion of multiple sets of embedded samples is measured by the degree of dispersion. The greater the difference between samples, the more ambiguous the model’s understanding of the current state is. This allows for the quantification of cognitive uncertainty. For example, when encountering a compound degradation pattern that has never been seen before, the sample dispersion will increase significantly.

[0056] By adding noise of different intensities to the original monitoring data, the uncertainty of the sensor error is simulated and quantified. Embedded samples are regenerated and their dispersion is calculated. The greater the dispersion, the stronger the interference of sensor noise on the evaluation results.

[0057] It should be noted that Bayesian probability weights, by introducing probability distributions, can quantify cognitive uncertainty, namely the fuzziness of the model regarding unknown insulation modes and data uncertainty, as well as the interference of sensor noise on the assessment. This compensates for the shortcomings of deterministic weights, which only output fixed features and fail to reflect the reliability of the assessment. For example, when a partial discharge mode anomaly is located, if the data uncertainty is high, it can suggest that the partial discharge sensor needs to be verified before confirming the anomaly; if the cognitive uncertainty is high, it can suggest the possible existence of an unknown composite degradation mode.

[0058] On the other hand, the probability distribution output by the Bayesian probability weights can dynamically adjust the stringency of the warning threshold. When the total uncertainty is high, the deviation threshold is appropriately relaxed to avoid misjudgment; when the uncertainty is low, the threshold is tightened to improve the warning sensitivity, making the threshold judgment more suitable for the actual assessment scenario. Judgment based on normal manifold: If the main feature center of the probability distribution falls within the range of the normal manifold and the total uncertainty is small, it is judged as a normal state with high confidence. If the feature center deviates from the normal manifold and the uncertainty is large, it is judged as an abnormal state and requires manual review, thus requiring dual verification based on probability and uncertainty.

[0059] Quantification of cognitive uncertainty: Cognitive uncertainty stems from the model's insufficient understanding of unseen insulation state patterns, such as novel composite degradation. This is addressed by quantifying the dispersion of Bayesian model sampling results. Based on 100-200 sets of insulation state embedding samples generated by the Bayesian graph encoder, the average embedding of all samples is calculated, i.e., the feature center of all samples. The difference between each embedded sample and the average embedding is statistically analyzed, such as the sum of differences in each dimension of the feature vector. The greater the difference, the more inconsistent the model's judgment of the current state, and the higher the cognitive uncertainty. Combining the distribution of cognitive uncertainty in historical normal states, such as the small sample differences during normal operation, the current cognitive uncertainty is converted into a quantized value of 0-1, where 0 represents complete certainty and 1 represents extreme uncertainty.

[0060] Quantification of data uncertainty: Data uncertainty stems from noise in the monitoring data, such as the accuracy error of the oil chromatography sensor and interference from partial discharge signals. This uncertainty is addressed by quantifying the stability of features under noise interference: For the original multimodal monitoring data, such as gas concentration in oil and partial discharge waveforms, simulated noise of varying intensities is added, with the noise amplitude matching the sensor's factory error range, such as ±5%. The noisy data is then input into a Bayesian model to generate new embedded samples. The degree of difference between the embedded samples before and after noise addition is calculated, such as the magnitude of feature changes. The greater the difference, the stronger the interference of data noise on the evaluation results, and the higher the data uncertainty. This is also converted into a quantification value of 0 to 1.

[0061] Calculation of total uncertainty and embedding deviation: Cognitive uncertainty and data uncertainty are weighted and added together. The weights are set based on operational experience, such as 60% for cognitive uncertainty and 40% for data uncertainty, resulting in a total uncertainty value of 0 to 1. The higher the value, the lower the reliability of the evaluation results. The deviation of the current average embedding from the normal state manifold is calculated. The normal state manifold is the feature clustering region of historical normal samples. By comparing the distance between the current average embedding and the center of the manifold, such as the feature matching degree, it is converted into a deviation value of 0 to 1, where 0 indicates a complete match to the normal state and 1 indicates a serious deviation. The warning also outputs key influencing factors, such as the possibility of unknown degradation patterns when cognitive uncertainty is high, and large fluctuations in oil chromatography sensor data when data uncertainty is high, to assist operation and maintenance personnel in accurate troubleshooting.

[0062] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0063] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the insulation condition of transformers based on multi-parameter fusion perception using artificial intelligence, characterized in that, The method includes: Step 1: Acquire multimodal monitoring data of the transformer and construct a dynamic graph structure; the multimodal monitoring data includes: oil chromatography data, partial discharge data, and thermal imaging data; The process of constructing the dynamic graph structure is as follows: Multimodal monitoring data of the transformer is acquired, and each parameter in the multimodal monitoring data is used as a graph node. The graph nodes include oil chromatography nodes, partial discharge nodes, and thermal imaging nodes. The oil chromatography nodes include H2 concentration nodes, CH4 concentration nodes, and C2H2 concentration nodes, each corresponding to the concentration time-series data of a characteristic gas. The partial discharge nodes, by setting partial discharge quantity nodes and discharge phase distribution nodes, carry the quantitative characteristics of the partial discharge signal. The thermal imaging nodes include thermal imaging temperature mean nodes, temperature variance nodes, and hotspot location feature nodes. Initial edge weights between graph nodes are defined based on the transformer's electrothermal-chemical coupling mechanism. The transformer's operating conditions are acquired in real time, and the edge connections and weights are dynamically adjusted online according to load and temperature change thresholds to construct multi-parameter physical coupling relationships. Step 2: The dynamic graph structure is encoded by the encoder extraction embedding technology. Based on the encoded dynamic graph structure, spatiotemporal feature aggregation and evolution modeling are performed using graph convolution and temporal convolution techniques. The model is input with pre-set positive and negative transformer insulation samples and InfoNCE loss is used for unsupervised learning to obtain the state manifold of the normal transformer. Based on the state manifold, spatiotemporal alignment and anomaly differentiation of multimodal data are performed. The process of spatiotemporal feature aggregation and evolutionary modeling is as follows: Using the encoded dynamic graph node embedding as input, based on physically guided edge weight topology, graph attention convolution is used to aggregate the correlation features between graph nodes; based on spatially aware embedding at continuous time steps, a causal dilated temporal convolutional network is used to capture temporal evolution patterns; the temporal features output by the temporal convolution and the spatial features output by the graph convolution are then fused element-wise with weights to generate a global feature vector that simultaneously contains multi-parameter physical coupling relationships and temporal evolution patterns, which is used to construct the model. The process of spatiotemporal alignment and anomaly differentiation of multimodal data based on state manifold is as follows: spatiotemporal alignment of multimodal data is performed through time axis interpolation calibration and spatial correlation mapping to generate new sample embeddings. The deviation from the normal manifold is quantified by Mahalanobis distance. Normal, slight and severe anomalies are divided by the 3σ principle. The contribution of each modality anomaly is calculated. The anomaly source is located by combining the insulation degradation mechanism. Step 3: After spatiotemporal alignment and anomaly differentiation of multimodal data, Bayesian weights are introduced to output the probability distribution of insulation state embedding. The cognitive uncertainty and data uncertainty of the probability distribution are decomposed and quantified. When the total uncertainty or embedding deviation exceeds the threshold, an insulation state warning is triggered.

2. The transformer insulation state assessment method based on artificial intelligence multi-parameter fusion perception according to claim 1, characterized in that, The process of encoding the dynamic graph structure using encoder extraction and embedding techniques is as follows: Based on the heterogeneous data characteristics of different types of graph nodes in the dynamic graph structure, a dedicated encoder is used to extract basic features and transform them into a unified-dimensional initial graph node embedding to obtain a single-modal initial embedding. Based on the topology of dynamic graphs, the initial embedding of a single mode is fused into a global graph embedding; We perform weighted pooling on the temporal-aware embeddings of all graph nodes, and then perform batch standardization to eliminate the differences in feature distribution under different working conditions. The output is a dynamic graph global embedding with dimension D. The dynamic graph global embedding contains three information: multimodal features, physical associations, and spatiotemporal evolution. Through global pooling and standardization, the spatiotemporally fused graph node-level embeddings are aggregated into graph-level global embeddings.

3. The transformer insulation state assessment method based on artificial intelligence multi-parameter fusion perception according to claim 2, characterized in that, The global graph embedding includes spatial correlation and temporal evolution: The spatial association uses the edge weights of the dynamic graph structure as the physical prior, aggregates the information of adjacent graph nodes through a graph attention network, and calculates the attention weight for each graph node based on its initial embedding similarity with the adjacent graph nodes to strengthen the information transmission between graph nodes with strong physical associations. After multiple layers of graph convolution, the embedding of each graph node is updated to a spatially perceptual embedding that integrates its own features and the association features of adjacent graph nodes. The aforementioned temporal evolution is based on the topological changes of the dynamic graph over time. It uses a gated recurrent unit to capture the temporal evolution, takes the spatially aware embedding of continuous time steps as input, and dynamically filters redundant information through the update gate and reset gate of the GRU to output a temporally aware embedding that includes historical state dependencies.

4. The transformer insulation state assessment method based on artificial intelligence multi-parameter fusion perception according to claim 3, characterized in that, The process of obtaining the state manifold of a normal transformer is as follows: The pre-set positive and negative sample pairs are input into the contrastive learning head of the model. Cosine similarity is used to calculate the similarity between two augmented views in a positive sample pair, and the similarity between each positive sample and all negative samples in the negative sample set, quantifying the strength of feature associations between samples. The model parameters are optimized using gradient descent with InfoNCE loss as the objective function. The number of training iterations is set, and after each iteration, the average similarity of positive samples and the difference in similarity between positive and negative samples on the validation set are calculated. Model convergence is determined based on the average similarity of positive samples, the similarity between positive and negative samples, and the InfoNCE loss. After convergence, the encoded embeddings of all normal state samples in the training set are extracted, and their feature distribution is fitted using a Gaussian mixture model. The state manifold of a normal transformer is obtained.

5. The transformer insulation state assessment method based on artificial intelligence multi-parameter fusion perception according to claim 4, characterized in that, The process of embedding the probability distribution of the output insulation state is as follows: After spatiotemporal alignment and anomaly differentiation of multimodal data, the core model for generating feature embeddings is modified by replacing the original fixed parameters with probability distribution weights. Multiple independent sets of insulation state feature embeddings are generated by randomly selecting values ​​from the probability distribution of Bayesian weights. Based on multiple sets of effective embedding samples, a probability distribution model of insulation state is constructed, and two sources of uncertainty are distinguished to form a complete probability assessment result.

6. The transformer insulation state assessment method based on artificial intelligence multi-parameter fusion perception according to claim 5, characterized in that, The process of decomposing and quantifying the cognitive uncertainty and data uncertainty of the probability distribution is as follows: The dispersion of the sampling results from the Bayesian model is quantified, and the feature stability under noise interference is quantified. Cognitive uncertainty and data uncertainty are added together by weight to obtain the total uncertainty value, which ranges from 0 to 1. The deviation of the current average embedding from the normal state manifold is calculated, and a warning judgment is made based on the threshold.

Images