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

Abstract

The invention discloses a transformer insulation state evaluation method based on artificial intelligence multi-parameter fusion perception, and relates to the technical field of electrical equipment detection, and the method comprises the steps: 1, obtaining the multi-modal monitoring data of a transformer, and constructing a dynamic graph structure; wherein the multi-modal monitoring data comprises oil chromatography data, partial discharge data and thermal image data; 2, encoding the dynamic graph structure through an encoder extraction embedding technology, carrying out space-time dimension feature aggregation and evolution modeling based on the encoded dynamic graph structure by utilizing graph convolution and time convolution technologies, inputting preset transformer insulation positive and negative sample pairs into a constructed model, carrying out InfoNCE loss unsupervised learning, and carrying out infoNCE loss unsupervised learning; obtaining a state manifold of a normal transformer, and carrying out multi-modal data space-time alignment and anomaly distinguishing on the state manifold; and step 3, introducing a Bayesian weight after multi-modal data space-time alignment and anomaly distinguishing, and outputting probability distribution of insulation state embedding.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electrical equipment detection, in particular to a transformer insulation state evaluation method based on artificial intelligence multi-parameter fusion perception. BACKGROUND

[0002] As a core pivotal device in a power transmission and transformation system, the operation reliability of a power transformer directly determines the safety and stability of a power grid and the level of power supply quality. In a long-term operation process, the insulation system inside the transformer continuously bears the coupling effects of multiple physical fields such as electricity, heat, machinery and chemistry, and typical degradation problems such as aging, dampness and partial discharge are prone to occur, which are not only the key inducement of insulation performance degradation but also the primary root cause of sudden transformer failure. Moreover, insulation aging is usually caused by the cumulative effect of long-term thermal stress and electric stress, which significantly reduces the dielectric strength and mechanical toughness of oil-paper insulation materials. At the same time, the invasion of water into the insulation system seriously weakens the insulation resistance and breakdown voltage thereof. Partial discharge caused by early insulation defects, although the initial energy is weak, has the ability to continuously erode the insulation structure, and if not identified and intervened in time, it is easy to evolve into a through discharge or even an insulation breakdown accident.

[0003] Traditional transformer insulation state evaluation has long relied on a single index such as furfural content or dielectric loss value, which is difficult to accurately reflect the non-linear evolution characteristics and multi-physical field coupling mechanism in the insulation degradation process. Especially for slowly aging insulation systems, the acquisition of a large number of labeled fault and insulation state samples by existing AI models not only has high cost and long cycle, but also causes the sample scarcity problem to be prominent due to the rarity of real degradation events, which limits the training quality and generalization ability of the supervised learning model. At the same time, the oil-dissolved gas components, partial discharge signals and infrared thermal images in the transformer have heterogeneity in terms of dimension, sampling frequency and physical semantic level, and cannot be simply spliced and fused. The inherent spatio-temporal correlation and physical coupling law are needed, which leads to distortion of feature expression and low information utilization efficiency. The existing AI model generally outputs black-box prediction results, lacks quantitative characterization of evaluation confidence, and lacks reliable verification of the decision of power grid operation and maintenance. SUMMARY

[0004] To achieve the above purpose, the application is implemented by the following technical solutions: Step one: obtaining transformer multi-modal monitoring data and constructing a dynamic graph structure; wherein the multi-modal monitoring data includes oil chromatographic data, partial discharge data and thermal image data; Step two: encode the dynamic graph structure through an encoder extraction embedding technology, based on the encoded dynamic graph structure, use graph convolution and time convolution technology to model feature aggregation and evolution in space-time dimensions, input the pre-set transformer insulation positive and negative sample pairs into the model constructed by the transformer insulation state manifold obtained by unsupervised learning of InfoNCE loss, and perform spatiotemporal alignment and abnormality distinction on the state manifold; Step three: introduce Bayesian weight after spatiotemporal alignment and abnormality distinction of multi-modal data, output the probability distribution of insulation state embedding, decompose and quantify the cognitive uncertainty and data uncertainty of the probability distribution, and trigger insulation state warning when the total uncertainty or embedding deviation exceeds the threshold.

[0005] Further, the process of constructing a dynamic graph structure is: Obtain transformer multi-modal monitoring data, take each parameter in the transformer multi-modal monitoring data as a graph node, define the initial edge weight between the graph nodes based on the transformer electro-thermal chemical coupling mechanism, obtain the online dynamic adjustment of the edge connection and weight under the transformer operating condition, and dynamically adjust the edge connection and weight under the transformer operating condition, and construct the multi-parameter physical coupling relationship.

[0006] Further, the process of encoding the dynamic graph structure through the encoder extraction embedding technology is: Based on the heterogeneous data characteristics of different types of graph nodes in the dynamic graph structure, the exclusive encoder is used to extract the basic features, which are converted into initial node embeddings of a unified dimension to obtain single-modal initial embeddings; based on the topology structure of the dynamic graph, the single-modal initial embeddings are fused into global graph embeddings; the time sequence perception embedding of all graph nodes is weighted and pooled, and the feature distribution difference under different conditions is eliminated through batch normalization processing, and the dynamic graph global embedding with a dimension of D is output, which contains three information of multi-modal features, physical correlation and spatiotemporal evolution. Through global pooling and standardization, the node-level embedding fused in space and time is aggregated into a global graph embedding.

[0007] Further, the global graph embedding includes spatial correlation and temporal evolution: Spatial correlation: taking the edge weight of the dynamic graph structure as the physical prior, the adjacent node information is aggregated through the graph attention network, for each graph node, the attention weight is calculated according to the similarity of the initial embedding of the adjacent graph node, the information transmission between the graph nodes with strong physical correlation is strengthened, and through multi-layer graph convolution, the embedding of each node is updated to the spatial perception embedding that fuses its own features and adjacent node correlation features; Temporal evolution: for the topology change of the dynamic graph over time, the gated recurrent unit is used to capture the temporal evolution, taking the spatial perception embedding at consecutive time steps as the input, through the update gate and reset gate of GRU, the redundant information is dynamically filtered and the key changes are retained, and the time sequence perception embedding containing historical state dependence is output.

[0008] Further, the process of performing the feature aggregation and evolution modeling in the space-time dimension is: With the encoded dynamic graph node embedding as input, based on the physical guided edge weight topology, the associated feature aggregation between nodes is realized through graph attention convolution; based on the spatial perception embedding of continuous time steps, the time evolution law is captured through the causal expansion time convolution network, the time sequence features output by the time convolution and the spatial features output by the graph convolution are element-wise weighted and fused to generate a global feature vector containing the multi-parameter physical coupling relationship and the time evolution law, which is used to construct the model.

[0009] Further, the process of obtaining the state manifold of the normal transformer is: The contrast learning head of the model is input with positive and negative sample pairs, the similarity between the two enhanced views in the positive sample pair and the similarity between each positive sample and all negative samples in the negative sample set are calculated through cosine similarity, the feature correlation strength between samples is quantified, the InfoNCE loss is taken as the objective function, the model parameters are optimized through gradient descent, the training iteration number is set, the average similarity of positive samples and the similarity difference between positive and negative samples on the validation set are calculated after each iteration, when the indicators are stable and the loss value tends to be flat for 10 consecutive times, the model is determined to be converged; after the model converges, the encoding embedding of the normal state samples in all training sets is extracted, the feature distribution is fitted through the Gaussian mixture model, and the normal state manifold is obtained.

[0010] Further, the process of performing space-time alignment and abnormality distinction of the state manifold on multi-modal data is: The space-time alignment of multi-modal data is realized through time axis interpolation calibration and spatial correlation mapping, new sample embedding is generated, the deviation degree from the normal manifold is quantified by Mahalanobis distance, the normal, slight and severe abnormalities are divided by the 3σ principle, the abnormal contribution degree of each modality is calculated, and the abnormal source is located combined with the insulation deterioration mechanism.

[0011] Further, the process of outputting the probability distribution of the insulation state embedding is: After the space-time alignment and abnormality distinction of multi-modal data, the core model of generating feature embedding is modified, the original fixed parameters are replaced by probability distribution weights, multiple independent insulation state feature embeddings are generated by randomly taking values from the probability distribution of Bayesian weights multiple times, based on multiple effective embedding samples, a probability distribution model of insulation state is constructed, and two sources of uncertainty are distinguished, forming a complete probability evaluation result.

[0012] Further, the process of decomposing and quantifying the cognitive uncertainty and data uncertainty of the probability distribution is: The cognitive uncertainty and data uncertainty are added according to weights by quantifying the dispersion degree of the Bayesian model sampling result and quantifying the feature stability under noise interference, a total uncertainty value of 0-1 is obtained, the deviation degree of the current average embedding from the normal state manifold is calculated, and early warning judgment is made based on a threshold.

[0013] The transformer insulation state evaluation method based on artificial intelligence multi-parameter fusion perception provided by the application has the following beneficial effects: (1) The application constructs a dynamic graph structure based on an electro-thermal chemical coupling mechanism, takes each parameter in transformer multi-modal monitoring data, oil chromatogram, partial discharge and thermal image as a graph node, defines the initial edge weight between the graph nodes according to physical priori, and dynamically adjusts the initial edge weight with the running condition, realizes modeling of the multi-parameter physical coupling relationship of the insulation system, overcomes the feature expression distortion problem caused by the large difference between the multi-modal data dimension, sampling rate and physical semantics in the traditional method, and improves the feature expression ability and information utilization efficiency of the insulation state evaluation.

[0014] (2) The application adopts InfoNCE loss unsupervised learning to construct a normal state manifold, and realizes precise quantitative evaluation of the insulation deterioration state through a multi-modal data space-time alignment and abnormality distinguishing mechanism, and without a large number of labeled samples, solves the bottleneck problem of the poor generalization ability of the supervised learning model in the slowly aging insulation system, and objectively grades the insulation deterioration degree by quantifying the Mahalanobis distance and dividing the abnormality level according to the 3σ principle.

[0015] (3) The application introduces the Bayesian weight to output the probability distribution of the insulation state embedding, decomposes and quantifies the cognitive uncertainty and data uncertainty of the probability distribution, provides credibility support for operation and maintenance decision-making, triggers the insulation state early warning when the total uncertainty or embedding deviation degree exceeds the threshold, changes the evaluation result from black box output to credible evaluation with probability explanation, solves the key problem that the existing AI model lacks confidence quantification and is difficult to support operation and maintenance decision-making, and provides an interpretable decision basis for transformer insulation state evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The figure is a schematic diagram of the method of the application. DETAILED DESCRIPTION

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

[0018] Please refer toFigure 1 The embodiment of the application provides a transformer insulation state evaluation method based on artificial intelligence multi-parameter fusion perception, and the method comprises the following steps: Step 1: Obtain transformer multi-modal monitoring data and construct a dynamic graph structure; wherein the multi-modal monitoring data comprises oil chromatogram data, partial discharge data and thermal image data; Obtain transformer multi-modal monitoring data: Obtain the oil chromatogram data by using an oil dissolved gas analysis sensor such as a gas chromatograph or an online DGA monitoring device, take an oil sample from a transformer oil pillow or body, detect the concentration of characteristic gases dissolved in the oil, including H2, CH4, C2H2, C2H4, C2H6 and the like, automatically sample and analyze the online monitoring device at a preset period, manually sample and detect in a laboratory, and output time series data of the concentration of each gas.

[0019] Obtain the partial discharge data by using a partial discharge sensor such as an ultrahigh frequency sensor or an ultrasonic sensor, install the sensor on a transformer box, a bushing or the like, and collect electric signals or acoustic signals generated by partial discharge.

[0020] The UHF sensor captures ultrahigh frequency electromagnetic waves 300MHz-3GHz of partial discharge to realize non-contact monitoring, and the AE sensor detects mechanical vibration waves 20kHz-200kHz of discharge to monitor, and the data sampling rate is usually MHz level, and time series data such as partial discharge quantity and discharge phase are output.

[0021] Obtain the thermal image data by using an infrared thermal imager to perform infrared thermal imaging scanning on the body, bushing, joint and the like of the transformer, collect the temperature distribution image of the surface of the equipment, and the fixed thermal imager automatically photographs every 15 minutes at a fixed time interval, and the portable thermal imager is photographed by operation and maintenance personnel at regular intervals, and infrared thermal image data containing temperature mean, temperature variance, hot spot position and the like are output.

[0022] Synchronously collect multi-modal data or align the multi-modal data through a time stamp to form a multi-source data set matched in the time dimension.

[0023] Construct a dynamic graph structure: Each parameter in the transformer multi-modal monitoring data is taken as a graph node, the oil chromatogram node contains H2 concentration nodes, CH4 concentration nodes, C2H2 concentration nodes and the like, each node corresponds to time series data of the concentration of one characteristic gas, for the partial discharge node, the partial discharge quantity node and the discharge phase distribution node are set to carry quantitative characteristics of the partial discharge signal, and the thermal image node contains the thermal image temperature mean node, the temperature variance node and the hot spot position characteristic node to represent key information of the infrared thermal image.

[0024] The initial edge weight between the graph nodes is defined based on the electro-thermal chemical coupling mechanism of the transformer. According to the oil-immersed transformer fault gas generation mechanism, the energy density of arc partial discharge is high, which will promote the rupture of C-C bond in oil at 800-1200 DEG C and generate C2H2. In the experiment, it is shown that C2H2 is the characteristic gas of partial discharge, especially arc discharge, and the concentration of C2H2 has a strong correlation with the intensity of partial discharge; the partial discharge amount node and the C2H2 concentration node are initially set to a high value because partial discharge will cause the generation of C2H2; The heat generated by the transformer winding is transferred to the oil temperature through heat conduction and oil convection, and there is a clear thermal coupling relationship between them. According to the industry standard 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, such as the natural oil circulation transformer top oil temperature and winding hot spot temperature difference of about 23 DEG C. The oil temperature node and the winding temperature node have a strong correlation due to heat conduction, and the initial edge weight can be set to 0.7; For the H2 concentration node and the thermal image temperature average node, there is no direct electro-thermal chemical coupling between the parameters, and the correlation only comes from the weak correlation on the data level, such as accidental working condition overlap, and lacks physical mechanism support. Combined with gray correlation analysis and other methods for quantitative research on the correlation of multiple parameters, the weight of such weak correlation is usually much lower than that of the parameters with physical coupling. The nodes without direct physical coupling, such as the H2 concentration node and the thermal image temperature average node, are initially set to an edge weight of 0.1.

[0025] It should be noted that the specific numerical value is related to the transformer model, and the present application is mainly based on oil-immersed power transformers of 110kV and above voltage grade, such as common SZ22-12500 / 110, SFPZ9-120000 / 110 and other models.

[0026] The edge connection and weight are obtained by online dynamic adjustment of the transformer operating condition. Under the condition of sudden load change, when the transformer load change rate exceeds the preset threshold, such as 10% / minute, the edge weight between the load current, winding temperature and oil temperature nodes is enhanced, such as from 0.6 to 0.8, to reflect the strengthening effect of load change on temperature correlation; under the condition of sudden temperature change, when the oil temperature change rate exceeds 5 DEG C / hour, the edge weight of the related nodes of oil temperature and gas solubility in oil is increased, reflecting the influence of temperature on the solubility characteristics of gas in oil.

[0027] By converting the multi-modal monitoring data into a graph structure that evolves dynamically with the working condition, the precise modeling of the multi-parameter physical coupling relationship of the transformer is realized.

[0028] Step two: encode the dynamic graph structure by an encoder extraction embedding technology, based on the encoded dynamic graph structure, use graph convolution and time convolution technology to model the feature aggregation and evolution in space-time dimension, input the pre-set transformer positive and negative sample pairs into the model constructed by the transformer, and perform unsupervised learning based on the InfoNCE loss to obtain the state manifold of the normal transformer, and perform space-time alignment and abnormality distinction on the state manifold; Dynamic graph structure encoding process: For different types of nodes in the dynamic graph structure, such as oil chromatogram, partial discharge, and infrared thermal image, etc., the characteristics of heterogeneous data are extracted by using dedicated encoders to transform into initial node embeddings of uniform dimensions.

[0029] For oil chromatogram and dielectric response type nodes, low sampling rate time series signals such as daily gas concentration in oil and quarterly dielectric loss tangent are encoded using time series feature enhancement and 1D convolution encoder. The original time series data is first statistically processed by sliding window to calculate the mean, variance, and trend slope to enhance local time series features. The 1D convolution neural network is inputted, and through multiple convolution kernels with sizes of 3, 5, and 7 corresponding to different time scales, short-term fluctuations and long-term trends are captured. After dimension compression by the pooling layer, the output dimension is d of the time series node embedding.

[0030] For partial discharge type high frequency signal node encoding, time-frequency conversion and attention mechanism encoder are used. The time domain waveform is converted into a two-dimensional time-frequency matrix by short-time Fourier transform to preserve the frequency distribution and time evolution characteristics of the discharge signal. The convolution network with self-attention mechanism is inputted, and the attention weight is focused on the significant area of the discharge feature, such as the time-frequency point corresponding to the pulse peak. The fully connected layer is mapped to the partial discharge node embedding with dimension d.

[0031] In the process of infrared thermal image type image node encoding, the node data is a temperature distribution image. Image segmentation and visual ViT encoder are used. The thermal image is first segmented into regions, such as body, sleeve, and joint regions, to extract prior features such as temperature extrema and gradients in each region. The segmented image blocks are inputted into ViT, and the temperature correlation between regions such as the temperature difference between the sleeve and the body is captured by multi-head self-attention. After clstoken aggregates the global features, the output is the thermal image node embedding with dimension d.

[0032] Space-time graph fusion encoding: Based on the topology structure of the dynamic graph, the node and edge weights, the single modal initial embedding is fused into a global graph embedding containing spatial correlation and temporal evolution to model the dynamic correlation and time evolution between nodes. In the spatial dimension, through the physical guidance of the graph attention convolution, the edge weight of the dynamic graph structure is taken as the physical prior, the adjacent node information is aggregated through the graph attention network GAT, for each graph node, the attention weight is calculated according to the similarity of the initial embedding of the adjacent graph node, and the physical edge weight is modified to strengthen the information transmission between the graph nodes with strong physical correlation, such as the attention weight of the partial discharge node to the C2H2 node is improved; after multi-layer graph convolution, each node embedding is updated to a spatial perception embedding that integrates its own features and the associated features of adjacent nodes, and the physical coupling relationships such as partial discharge to gas generation and temperature to aging acceleration are preserved.

[0033] In the time dimension, through the time difference perception of the gated recurrent encoding, for the topology change and edge weight adjustment of the dynamic graph over time, the gated recurrent unit GRU is adopted to capture the time evolution, and the spatial perception embedding at the continuous time step is taken as the input. Through the update gate and the reset gate of the GRU, redundant information is dynamically filtered, such as repeated features under stable working conditions, and key changes are preserved, such as the mutation of the temperature node embedding when the load suddenly changes; the time perception embedding containing historical state dependence is output, reflecting the time correlation from short-term state fluctuation to long-term insulation deterioration trend.

[0034] Dynamic graph embedding output: Global feature aggregation and dimension unification: through global pooling and standardization, the node-level embedding after spatio-temporal fusion is aggregated into a graph-level global embedding, the time perception embedding of all graph nodes is weighted pooled, and the weight is the degree centrality of the node in the graph, which highlights the key parameter nodes; after batch normalization processing, the feature distribution difference under different working conditions is eliminated, and the output dimension is D, such as the dynamic graph global embedding with D being 256. This embedding contains three pieces of information: multi-modal features, physical correlations, and spatio-temporal evolution. Through the unique encoding of heterogeneity, spatio-temporal fusion modeling correlation, and global aggregation output consistency, the unique information of each modal data is preserved, and the physical coupling and time evolution between nodes in the dynamic graph are captured.

[0035] Spatio-temporal dimension feature aggregation and evolution modeling: Spatio-temporal fusion features are formed through graph convolution for spatial dimension feature aggregation and time convolution for time dimension evolution modeling.

[0036] The spatial dimension feature aggregation takes the encoded dynamic graph node embedding as input, realizes the inter-node correlation feature aggregation through graph attention convolution based on the physical guidance edge weight topology, combines the dynamic graph real-time edge weight to reflect the physical correlation strength under the current working condition, corrects the original attention weight between nodes by embedding similarity calculation, and strengthens the information transmission weight of nodes with strong physical correlation such as partial discharge, characteristic gas and temperature and dielectric loss; through multi-layer graph convolution iteration, the first layer of convolution focuses on directly adjacent nodes, such as the feature interaction of oil temperature and winding temperature, and outputs the intermediate embedding containing local correlation; the second layer of convolution is extended to the second-order adjacent nodes, such as partial discharge, characteristic gas and dielectric loss, to capture the indirect coupling relationship; through 3-4 layers of convolution, the single node features are fused into the spatial perception embedding of the whole graph physical correlation.

[0037] Based on the spatial perception embedding of continuous time steps, the time series evolution law is captured through the causal dilated temporal convolution network, different dilated rates such as 1, 2 and 4 are used, and short-term state fluctuations within 1 hour, medium-term trend changes within 1 day and long-term deterioration tendency within 1 week are captured respectively to cover the multi-time scale features of transformer insulation state evolution; each time convolution layer retains the original time series information through the residual path to avoid feature attenuation in long-term dependence modeling, and focuses on strengthening the time series features of abnormal moments such as sudden increase of partial discharge and jump of gas concentration; in view of the difference in sampling rate of multi-modal data, such as 1 time per day for oil chromatogram and 10 times per second for partial discharge, the spatial embedding of different time intervals is unified into a fixed time interval sequence, such as 1 sample per 15 minutes, to ensure the consistency of the input of time convolution.

[0038] The time series features output by the time convolution and the spatial features output by the graph convolution are element-wise weighted and fused, the weight is dynamically allocated according to the time-space correlation strength under the current working condition, and a global feature vector containing the multi-parameter physical coupling relationship and the time series evolution law is generated.

[0039] Obtaining 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 normal transformer state manifold is obtained.

[0040] Positive sample pair generation: through two different data enhancement operations on the encoded dynamic graph global embedding at the same time, spatial enhancement and time enhancement, two enhanced views are obtained to form a positive sample pair, so that both of them are from the same normal state although they are different in form; Spatial enhancement: randomly mask 10%-15% of non-key node features, such as secondary gas concentration nodes; Time enhancement: a small time series disturbance is made to the continuous time step embedding, such as ±1 step time offset; Negative sample pair generation: By selecting three types of negative samples, including normal state embeddings with different time windows, such as samples with intervals exceeding 24 hours; suspicious state embeddings marked in historical records, which are not explicitly faulty but have large parameter fluctuations; and invalid embeddings generated by randomly disrupting the connection between graph nodes. Each positive sample pair is matched with 15-20 negative samples to form a negative sample pair set, and there is a significant difference in state characteristics between the positive sample and the negative sample.

[0041] Unsupervised training through InfoNCE loss: The positive and negative sample pairs are input into the contrastive learning head of the model. The similarity between the two augmented views in the positive sample pair and the similarity between each positive sample and all negative samples in the negative sample set are calculated using cosine similarity to quantify the feature correlation strength between samples. The InfoNCE loss is used as the objective function, and the model parameters are optimized through gradient descent. The loss function is designed to maximize the similarity of positive sample pairs and minimize the similarity between positive samples and negative samples, i.e., the model learns that different augmented views of the same normal state should be clustered together and that other state samples should be separated from each other. The training iteration number is set, such as 200-300 rounds. After each iteration, the average similarity of positive samples on the validation set is calculated, and the target is ≥0.85 and the difference between positive and negative sample similarities is ≥0.6. When the indicators are stable for 10 consecutive rounds and the loss value tends to be flat, the model is considered to have converged.

[0042] After the model converges, extract the encoding embeddings of normal state samples in all training sets, fit their feature distribution through Gaussian mixture model, get the normal state manifold, the manifold represents the center feature of normal state with mean vector, and represents the fluctuation range of normal state with covariance matrix, forming the normal state feature clustering area, providing a reference for subsequent anomaly detection: when the new sample embedding falls within the manifold range, it is determined to be normal state; if it exceeds the range, it is marked as abnormal.

[0043] Multi-modal data spatio-temporal alignment and anomaly discrimination: Using the trained normal state manifold as a reference, solve the spatio-temporal misalignment problem of multi-modal data and accurately distinguish abnormal features, including multi-modal data spatio-temporal alignment, anomaly discrimination based on state manifold, and abnormal modality attribution.

[0044] Multi-modal data spatio-temporal alignment: For multi-modal data, such as oil chromatogram, partial discharge, and infrared thermal image, the sampling rate, timestamp, and feature dimension are inconsistent. Through time axis calibration and spatial correlation mapping, alignment is achieved to ensure that the input features match the normal state manifold dimension; S201: Comb the original sampling period of each modality, such as partial discharge 10kHz, oil chromatogram 1 time / day, infrared thermal image 1 time / 15 minutes, take the highest sampling rate modality as the reference, such as partial discharge, and perform adaptive interpolation completion on low sampling rate modalities. Linear interpolation is used to generate pseudo-sampling points for oil chromatogram data at 15 minutes / time. The daily trend is preserved while matching the thermal image cycle. Thermal image data is generated through adjacent interpolation to generate pseudo-sampling points at 1 minute / time, which matches the time density of partial discharge; S202: Based on the GPS timestamp of the transformer operation and maintenance system, timestamp calibration is performed on all modal data to eliminate device clock deviation. The time difference between partial discharge sensor and thermal imager is less than or equal to 1 second, ensuring that multi-modal data within the same time slice corresponds to the same operating time. S203: Slice the aligned time series data in fixed time windows, such as 15 minutes / window. Each window outputs 1 set of spatiotemporal samples containing all modal features, with the same time dimension as the normal state manifold, such as 15 minutes / sample.

[0045] Spatial correlation mapping is based on the physical edge weight of dynamic graph, such as the correlation weight of partial discharge and C2H2, and the correlation weight of oil temperature and dielectric loss. A multi-modal feature correlation matrix is constructed. For each spatiotemporal sample, weakly correlated modal features, such as infrared thermal image and H2 concentration, with a correlation weight <0.3, are fused, such as weighted summation, and strongly correlated modal features, such as partial discharge and C2H2, with a correlation weight ≥0.6, are kept independent dimensions to avoid redundancy. Through a fully connected layer, the fused multi-modal features are mapped to the same dimension as the normal state manifold, such as 256 dimensions, ensuring consistent feature dimensions for subsequent comparison.

[0046] Abnormality differentiation based on state manifold: Take the mean vector and covariance matrix fitted by the Gaussian mixture model as the reference. Through feature embedding comparison, deviation calculation, and threshold judgment, abnormality differentiation is performed. Align the multi-modal spatiotemporal samples and input them into the trained encoder, including modal-specific encoding and spatiotemporal graph fusion encoding. Output new sample feature embeddings with the same dimension as the normal state manifold, making their feature distribution regular and consistent with the normal manifold, such as time trend and spatial correlation weight matching. Calculate the deviation of the new sample embedding from the normal manifold using Mahalanobis distance. Take the mean vector of the normal manifold as the center and combine the covariance matrix to reflect the correlation of feature fluctuations under normal conditions. Calculate the Mahalanobis distance of the new sample embedding to the mean vector At the same time, calculate the average cosine similarity between the new sample embedding and all samples in the normal manifold The larger is and the smaller is, the more obvious the deviation of the new sample feature from the normal manifold.

[0047] Mahalanobis distance: can eliminate the influence of correlation between features, more suitable for manifold distribution than Euclidean distance, so the invention uses Mahalanobis distance as the calculation of embedding new samples into the mean vector.

[0048] It should be noted that when the state manifold is aligned and distinguished in time and space with multi-modal data, the problems of sampling rate, timestamp and feature dimension inconsistency are specially processed, and there are essential differences in the target and accuracy requirement of data processing in different stages. The processing of the previous steps cannot meet the fine requirement of state manifold matching. For example, only rough alignment of oil chromatogram and partial discharge is required in the early stage, which can ensure that each modal data corresponds to the same running period, and there is no need to pursue extreme accuracy. The feature dimension only needs to adapt to the dynamic graph node definition. In the state manifold matching and abnormality distinguishing stage, the feature distribution of the data to be evaluated is completely adapted to the pre-trained normal state manifold. If the data aligned directly in the early stage is used, problems such as coarse time granularity of oil chromatogram data, mismatch of partial discharge feature dimension, and cumulative timestamp deviation will occur, resulting in significant errors in deviation judgment.

[0049] Based on the Mahalanobis distance and similarity of normal samples in the training set, the threshold is set by the 3σ principle: the mean of the Mahalanobis distance of the normal manifold is , the standard deviation is , and the abnormal judgment threshold is set as ; the average similarity of normal samples is , the standard deviation is , and the similarity threshold is set as ; the judgment rule is: if ≤ and ≥ , it is judged as normal state; if < ≤ and ≤ < , it is judged as slight abnormality, such as feature deviation caused by short-term working condition fluctuation; if > and < , it is judged as serious abnormality, such as feature mutation caused by insulation deterioration.

[0050] Abnormality attribution: After abnormality distinguishing, the key abnormality source is located by calculating the modal contribution degree, avoiding misjudgment caused by multi-modal data confusion. The feature embedding of each modal of the sample judged as abnormal is extracted, such as oil chromatogram modal embedding and partial discharge modal embedding. The Mahalanobis distance between single modal embedding and corresponding modal embedding of normal manifold is calculated , where i is the modal number; according to where n is the total number of modes, j is the summation index variable, and n is the total number of modes, to calculate the abnormal contribution of each mode The higher the contribution, the more significant the deviation of the mode from the normal state; combined with the insulation degradation mechanism of the transformer, the high-contribution mode, such as ≥0.3 is attributed: if the partial discharge mode has the highest contribution and the corresponding C2H2 concentration mode has the second highest contribution, it is determined that the abnormal source is the insulation breakdown risk caused by the aggravation of partial discharge; if the oil temperature mode and the dielectric loss mode have high contributions, it is determined that the abnormal source is the accelerated insulation aging caused by temperature rise.

[0051] Step three: after the spatiotemporal alignment and abnormality distinction of the multi-modal data, the Bayesian weight is introduced to output the probability distribution of the insulation state embedding, and the cognitive uncertainty and data uncertainty of the probability distribution are decomposed and quantified, and the insulation state warning is triggered when the total uncertainty or embedding deviation exceeds the threshold; Probability distribution of insulation state embedding: By introducing the Bayesian weight in the feature embedding generation link, the certainty features of the insulation state are converted into probability distribution form, and the uncertainty in the evaluation is quantified: S301: Construct a Bayesian graph encoder to convert fixed weights into probability distribution weights; after completing the spatiotemporal alignment and abnormality distinction of the multi-modal data, the core model for generating feature embedding, i.e., the graph convolution and time convolution network, is reconstructed, and the original fixed parameters such as convolution kernel and attention weight are replaced by probability distribution weights, so that each parameter corresponds to a possible value range rather than a single numerical value: For the weights in the graph convolution layer that depict node association, such as the association strength between partial discharge and characteristic gas, set them to follow a certain probability distribution, such as a normal distribution value range, and the center value of the distribution is initialized based on the physical mechanism, such as setting the association weight center value between partial discharge and acetylene to a high level, and the fluctuation range of the distribution, i.e., the standard deviation, is limited to a reasonable range to avoid negative association, so that the weight change conforms to the electrical-thermal-chemical coupling law of transformer insulation; for the parameters in the time convolution layer that capture the time sequence changes, such as the trend parameter of temperature change with load, also set them to probability distribution, and the center value reflects the typical time sequence law in normal operation, such as the gentle trend of oil temperature rise with load, and the fluctuation range covers reasonable working condition fluctuations, such as temperature jump caused by short-term load surge.

[0052] S302: Generate multiple sets of embedding samples by Monte Carlo sampling, randomly take values from the probability distribution of the Bayesian weight multiple times, generate multiple independent insulation state feature embeddings, and approximate the overall distribution of the features. From the probability distribution weights of the graph convolution and time convolution, randomly extract 100-200 different weight combinations.

[0053] The spatio-temporally aligned multi-modal data, which has been matched with the normal state manifold dimension, is input into 100-200 groups of weight corresponding models respectively, to obtain 100-200 groups of insulation state feature embeddings, each of which contains multi-modal fused spatio-temporal features; at the same time, abnormal embeddings that obviously violate physical common sense, such as embeddings with negative partial discharge values, are removed, and valid samples are retained.

[0054] S303: Based on the plurality of valid embedding samples, a probability distribution model of the insulation state is constructed, and two sources of uncertainty are distinguished to form a complete probability evaluation result: The valid embedding samples are fitted by using a Gaussian mixture model to obtain a probability distribution of the insulation state embedding, which contains a plurality of feature aggregation centers corresponding to different insulation state sub-classes and the feature fluctuation range of each center, which reflects not only the typical features of the current state but also the possible variation interval of the features.

[0055] By calculating the dispersion degree of a plurality of embedding samples, the greater the difference between the samples, the more ambiguous the model's understanding of the current state, and the quantitative cognitive uncertainty is obtained, for example, when encountering a never-before-seen compound degradation mode, the sample dispersion will increase significantly.

[0056] By adding different intensities of noise to the original monitoring data, the sensor error is quantified to quantify the data uncertainty, the embedding samples are regenerated and the dispersion degree is calculated, and the greater the dispersion, the stronger the interference of the sensing noise on the evaluation result.

[0057] It should be noted that the Bayesian probability weight can quantify cognitive uncertainty by introducing a probability distribution, i.e., the ambiguity of the model to unknown insulation modes and data uncertainty, and the interference of sensing noise on the evaluation, which makes up for the shortcomings of deterministic weights that only output fixed features and reflect the reliability of the evaluation; for example, when locating a partial discharge modal anomaly, if the data uncertainty is high, it may be necessary to first verify the partial discharge sensor before confirming the anomaly; if the cognitive uncertainty is high, it may indicate the presence of an unknown compound degradation mode.

[0058] On the other hand, the probability distribution output by the Bayesian probability weight can dynamically adjust the strictness of the warning threshold. When the total uncertainty is high, the deviation threshold is appropriately relaxed to avoid false positives; when the uncertainty is low, the threshold is tightened to improve the warning sensitivity, making the threshold judgment more suitable for the actual evaluation scenario Combined with the normal state manifold judgment: If the main feature center of the probability distribution falls within the normal manifold range and the total uncertainty is small, it is determined to be in a normal state with high confidence; If the feature center deviates from the normal manifold and the uncertainty is large, it is determined to be in an abnormal state and needs to be manually reviewed, thereby achieving double verification based on probability and uncertainty.

[0059] Quantification of cognitive uncertainty: Cognitive uncertainty is caused by the lack of understanding of unobserved insulation state patterns, such as new composite degradation. The dispersion of the results of Bayesian model sampling is quantified. Based on the Bayesian graph encoder, 100-200 groups of insulation state embedding samples are generated. The average embedding of all samples is calculated, i.e. the feature center of all samples. The difference between each embedding sample and the average embedding is calculated, such as the sum of the differences in each dimension of the feature vector. The greater the difference, the less consistent the model's judgment of the current state, and the higher the cognitive uncertainty. The cognitive uncertainty of the historical normal state is combined, such as the small difference between normal operation samples, to convert the current cognitive uncertainty into a quantitative value of 0-1, where 0 represents complete certainty and 1 represents extreme ambiguity.

[0060] Quantification of data uncertainty: Data uncertainty is caused by noise in monitoring data, such as oil chromatographic sensor accuracy error and partial discharge signal interference. The stability of the features under noise interference is quantified: add different intensities of simulated noise to the original multi-modal monitoring data, such as gas concentration in oil and partial discharge waveform. The noise amplitude matches the sensor factory error range, such as ±5%. The noisy data is input into the Bayesian model to generate new embedding samples. The difference between the embedding samples before and after noise is calculated, such as the amplitude of feature change. The greater the difference, the stronger the interference of data noise on the evaluation results, and the higher the data uncertainty, which is also converted into a quantitative value of 0-1.

[0061] Calculation of total uncertainty and embedding deviation: The cognitive uncertainty and data uncertainty are added according to the weight, and the weight is set according to the operation and maintenance experience, such as cognitive uncertainty accounting for 60% and data uncertainty accounting for 40%. The total uncertainty value of 0-1 is obtained, and the higher the value, the lower the credibility 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 aggregation area of historical normal samples. The distance between the current average embedding and the manifold center is calculated, such as feature matching degree, to convert it into a deviation value of 0-1. 0 represents complete matching with normal state, and 1 represents serious deviation. The warning also outputs key influencing factors, such as unknown degradation patterns when cognitive uncertainty is high, and large fluctuations in oil chromatographic sensor data when data uncertainty is high, to assist operation and maintenance personnel in accurate troubleshooting.

[0062] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art can be aware that units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed by hardware or software depends on the specific application and design constraints of the technical solutions.

[0063] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, and can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiments.

[0064] The above describes only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A transformer insulation state evaluation method based on artificial intelligence multi-parameter fusion perception, characterized in that, The method comprises: Step 1: Obtain the multi-modal monitoring data of the transformer, and construct a dynamic graph structure; wherein the multi-modal monitoring data comprises oil chromatographic data, partial discharge data and thermal image data; Step 2: encode the dynamic graph structure through an encoder extraction embedding technology, based on the encoded dynamic graph structure, use graph convolution and time convolution technology for feature aggregation and evolution modeling in space-time dimension, input the pre-set positive and negative samples of the transformer insulation into the constructed model, perform unsupervised learning of InfoNCE loss, obtain the state manifold of the normal transformer, and perform space-time alignment and abnormality differentiation of the state manifold on the multi-modal data; Step 3: introduce the Bayesian weight after the space-time alignment and abnormality differentiation of the multi-modal data, output the probability distribution of the insulation state embedding, decompose and quantify the cognitive uncertainty and data uncertainty of the probability distribution, and trigger the insulation state warning when the total uncertainty or embedding deviation degree exceeds the threshold.

2. The transformer insulation condition evaluation method based on artificial intelligence multi-parameter fusion perception according to claim 1, characterized in that, The process of constructing a dynamic graph structure is: Obtain the multi-modal monitoring data of the transformer, take each parameter in the multi-modal monitoring data of the transformer as a graph node, define the initial edge weight between the graph nodes based on the transformer electro-thermal chemical coupling mechanism, obtain the online dynamic adjustment edge connection and weight under the transformer operating condition, and construct the multi-parameter physical coupling relationship.

3. The transformer insulation state assessment method based on artificial intelligence multi-parameter fusion perception according to claim 2, characterized in that, The process of encoding the dynamic graph structure through the encoder extraction embedding technology is: Based on the heterogeneous data characteristics of different types of graph nodes in the dynamic graph structure, the exclusive encoder is used to extract the basic features, which are converted into initial graph node embeddings of a unified dimension to obtain single-modal initial embeddings; Based on the topology structure of the dynamic graph, the single-modal initial embeddings are fused into global graph embeddings; Weighted pooling is performed on the time sequence perception embedding of all graph nodes, batch normalization processing is performed to eliminate the feature distribution difference under different conditions, and the dynamic graph global embedding with a dimension of D is output, which contains three information of multi-modal features, physical correlation and space-time evolution, and the graph node level embedding fused in space-time is aggregated into global graph embedding through global pooling and standardization.

4. The transformer insulation condition evaluation method based on artificial intelligence multi-parameter fusion perception according to claim 3, characterized in that, The global graph embedding includes spatial correlation and time evolution: Spatial correlation: taking the edge weight of the dynamic graph structure as the physical prior, the information of adjacent graph nodes is aggregated through the graph attention network, for each graph node, the attention weight is calculated according to the similarity of the initial embedding of the adjacent graph nodes, the information transmission between the graph nodes with strong physical correlation is strengthened, and through multi-layer graph convolution, the embedding of each graph node is updated to the spatial perception embedding fused with its own features and the correlation features of adjacent graph nodes; Time evolution: for the topology change of the dynamic graph over time, the gated recurrent unit is used to capture the time evolution, the spatial perception embedding at continuous time steps is input, the redundant information is dynamically filtered through the update gate and reset gate of the GRU, and the time sequence perception embedding containing the historical state dependence is output.

5. The transformer insulation condition evaluation method based on artificial intelligence multi-parameter fusion perception according to claim 4, characterized in that, The process of performing feature aggregation and evolution modeling in space-time dimension is: With the encoded dynamic graph node embedding as input, based on the physical guidance edge weight topology, the correlation feature aggregation between graph nodes is carried out through graph attention convolution; based on the spatial perception embedding of continuous time steps, the time evolution law is captured through the causal expansion time convolution network, the time sequence features output by the time convolution and the spatial features output by the graph convolution are element-wise weighted and fused to generate a global feature vector containing the multi-parameter physical coupling relationship and the time evolution law, which is used to build the model.

6. The transformer insulation condition evaluation method based on artificial intelligence multi-parameter fusion perception according to claim 1, characterized in that, The process of obtaining the state manifold of the normal transformer is: The preset positive and negative sample pairs are input into the contrast learning head of the model, the similarity of the two enhanced views in the positive sample pair and the similarity of each positive sample and all negative samples in the negative sample set are calculated by cosine similarity, the feature correlation strength between the samples is quantified, the InfoNCE loss is taken as the objective function, the model parameters are optimized by gradient descent, the training iteration number is set, and the average similarity of the positive samples and the similarity difference between the positive and negative samples on the validation set are calculated after each iteration, when the indicators are stable and the loss value tends to be flat for 10 consecutive times, the model is determined to be converged; after the model converges, the encoding embedding of the normal state samples in all training sets is extracted, the feature distribution is fitted by a Gaussian mixture model, obtain the state manifold of the normal transformer.

7. The transformer insulation condition evaluation method based on artificial intelligence multi-parameter fusion perception according to claim 1, characterized in that, The process of the state manifold for multi-modal data space-time alignment and abnormality distinction is: Through time axis interpolation calibration and spatial correlation mapping, the multi-modal data space-time alignment is performed, the new sample embedding is generated, the deviation degree from the normal manifold is quantified by Mahalanobis distance, the normal, slight and severe abnormalities are divided by the 3σ principle, the abnormal contribution degree of each modality is calculated, and the abnormal source is located combined with the insulation deterioration mechanism.

8. The transformer insulation condition evaluation method based on artificial intelligence multi-parameter fusion perception according to claim 1, characterized in that, The process of outputting the probability distribution of the insulation state embedding is: After the multi-modal data space-time alignment and abnormality distinction, the core model of the generated feature embedding is modified, the original fixed parameters are replaced by probability distribution weights, multiple independent insulation state feature embeddings are generated by randomly taking values from the probability distribution of the Bayesian weight multiple times, and based on the multiple effective embedding samples, the probability distribution model of the insulation state is constructed, two sources of uncertainty are distinguished, and a complete probability evaluation result is formed.

9. The transformer insulation condition evaluation method based on artificial intelligence multi-parameter fusion perception according to claim 1, characterized in that, The process of decomposing and quantifying the cognitive uncertainty and data uncertainty of the probability distribution is: The dispersion degree of the Bayesian model sampling result is quantified, the feature stability under noise interference is quantified, the cognitive uncertainty and data uncertainty are added according to the weight, the total uncertainty value of 0-1 is obtained, the deviation degree of the current average embedding from the normal state manifold is calculated, and the warning judgment is made based on the threshold.

Images