Power transformer fault diagnosis method based on image analysis

Abstract

The invention provides a power transformer fault diagnosis method based on image analysis, relates to the technical field of fault diagnosis and health management, and realizes accurate, dynamic and root diagnosis of power transformer faults through an image analysis technology. Initial spatio-temporal features are extracted and an attention weight map is generated by using a spatio-temporal attention 3D convolutional neural network, so that key abnormal regions are effectively focused, and the extraction efficiency and pertinence of fault features are improved; a fault propagation network is constructed in combination with a causal discovery module, and causal relationship strength between potential fault areas is quantified, so that a fault root and a derivative phenomenon can be distinguished, and a root cause area can be accurately positioned; furthermore, neighborhood information is aggregated and node representation is updated through the graph neural network, so that the analysis capability of the fault propagation path is enhanced, a diagnosis result including root cause space positioning, visual propagation path and risk level evaluation is finally output, and the accuracy of fault diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and health management technology, specifically to a power transformer fault diagnosis method based on image analysis. Background Technology

[0002] Power transformers are key equipment in power transmission and distribution systems, enabling voltage transformation, power distribution, and isolation. Operating under complex electromagnetic and mechanical stress environments for extended periods, they are prone to failure due to potential defects such as insulation aging, poor contact, winding deformation, and localized overheating. Common fault types include increased contact resistance in windings and outgoing terminals, circulating current overheating caused by multiple grounding points in the core, increased bushing dielectric loss, and blockage or failure of the cooling system. These faults typically manifest as abnormal temperatures in specific areas of the equipment surface.

[0003] Currently, non-contact diagnosis of power transformer faults mainly relies on infrared thermal imaging technology. This technology acquires infrared radiation images of the transformer surface and converts them into temperature distribution maps, thus visually identifying overheated areas. However, most existing diagnostic methods are based on the analysis of single or a few static infrared images. This method only provides a "snapshot" of the temperature field at a specific moment, and is essentially a static and isolated analysis.

[0004] Therefore, existing technologies have the following problems at the image analysis level: static image-based analysis methods cannot extract and analyze the dynamic evolution characteristics of the fault area and the spatiotemporal correlation between them from time-series infrared image sequences. This makes it impossible to distinguish between the root cause of the fault and the derived phenomena, so that the diagnostic results remain at the level of superficial recognition. They lack the ability to locate the fault propagation chain and the root cause, and cannot achieve fault prediction and health management. As a result, maintenance decisions lack foresight and pertinence, and are inefficient. Summary of the Invention

[0005] This invention aims to address the lack of dynamic fault prediction and health management capabilities in existing technologies by providing a power transformer fault diagnosis method based on image analysis. This method not only achieves accurate fault diagnosis but also possesses fault prediction and health management functions.

[0006] The technical solution adopted in this invention is as follows:

[0007] A power transformer fault diagnosis method based on image analysis includes the following steps:

[0008] S1: Acquire a high-frequency time-series infrared image sequence of a power transformer, the high-frequency time-series infrared image sequence containing infrared images collected at multiple consecutive time points, covering multiple potential fault areas on the surface of the power transformer;

[0009] S2: Based on the spatiotemporal attention 3D convolutional neural network, the initial spatiotemporal features of the high-frequency temporal infrared image sequence are extracted through its spatiotemporal feature extraction layer. Then, the attention weight calculation layer generates an attention weight map reflecting the anomalous significance of each spatial location at different time points based on the initial spatiotemporal features. The attention weight map is then weighted and fused with the initial spatiotemporal features to output an attention weighted feature map.

[0010] S3: Based on the attention-weighted feature map, the causal relationship strength between the temporal feature vectors of each potential fault region derived from the attention-weighted feature map is calculated by the causal discovery module, and a fault propagation network with potential fault regions as nodes and causal relationship strength as directed edges is constructed according to the causal relationship strength.

[0011] S4: The attention weights representing regional saliency in the attention weight graph are weighted and combined with the causal relationship strength to obtain a comprehensive importance score for each potential fault region; the fault propagation network containing the comprehensive importance score is input into the graph neural network; the neighborhood information of each node in the fault propagation network is aggregated and the node representation is updated through a message passing mechanism; the root cause region as the origin of the fault is identified based on the updated node representation.

[0012] S5: Based on the root cause region and the fault propagation path extracted from the fault propagation network, generate and output a diagnostic result including the spatial location information of the root cause region, a visualized fault propagation path map, and a risk level assessment of the future deterioration of the root cause region.

[0013] Furthermore, in step S2, the attention weight calculation layer of the spatiotemporal attention 3D convolutional neural network adopts a soft attention mechanism, which calculates the attention weight of each spatial location at each time point through learnable query vectors, key vectors, and value vectors to generate the attention weight map.

[0014] Further, in step S3, the fault propagation network is a directed weighted graph; the nodes of the directed weighted graph correspond to the spatial location of each potential fault region, and the weight of the directed edge connecting two nodes is defined by the causal relationship strength.

[0015] Furthermore, in step S3, the causal discovery module employs Granger causality analysis to quantify the strength of the causal relationship through a vector autoregression model and an F-test.

[0016] Furthermore, in step S4, the attention weights and the causal relationship strengths are weighted and combined using an adaptive weighting method based on the sigmoid function.

[0017] Furthermore, the graph neural network adopts a graph attention network architecture, which enhances the expressive ability of neighborhood information aggregation during message passing through a multi-head attention mechanism.

[0018] Furthermore, the risk level assessment for the future deterioration of the root cause region includes:

[0019] Based on a historical failure case library, a mapping relationship between temperature change rate and failure development speed is established, and the probability of deterioration within a specified time window is predicted by combining the current temperature trend of the root cause region.

[0020] The beneficial effects of this invention are:

[0021] This invention achieves accurate, dynamic, and root cause diagnosis of power transformer faults through image analysis technology. Specifically, it acquires high-frequency temporal infrared image sequences and utilizes a spatiotemporal attention 3D convolutional neural network to extract initial spatiotemporal features and generate attention weight maps, effectively focusing on key abnormal areas and improving the efficiency and targeting of fault feature extraction. Combined with a causal discovery module, it constructs a fault propagation network to quantify the causal relationship strength between potential fault areas, thereby distinguishing between the root cause and derivative phenomena and accurately locating the root cause region. Furthermore, it aggregates neighborhood information and updates node representations through a graph neural network, enhancing the ability to analyze fault propagation paths. The final output includes diagnostic results such as root cause spatial location, visualized propagation paths, and risk level assessment. This overcomes the limitations of traditional static image analysis methods, which cannot capture dynamic evolution, lack focus on key areas, and fail to analyze fault propagation paths, thus improving the accuracy of fault diagnosis. Attached Figure Description

[0022] Figure 1 This is a flowchart of an image analysis-based power transformer fault diagnosis method according to an embodiment of the present invention;

[0023] Figure 2 This is an overall flowchart of transformer fault diagnosis according to an embodiment of the present invention;

[0024] Figure 3 A flowchart of an embodiment of the present invention;

[0025] Figure 4 This is a flowchart of causal discovery and network construction according to an embodiment of the present invention;

[0026] Figure 5 This is a flowchart illustrating the fusion diagnosis and root cause localization of an embodiment of the present invention. Detailed Implementation

[0027] 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.

[0028] like Figures 1-5 As shown, the image analysis-based power transformer fault diagnosis method of this invention includes the following steps:

[0029] S1: Acquire a high-frequency time-series infrared image sequence of the power transformer. The high-frequency time-series infrared image sequence contains infrared images collected at multiple consecutive time points, covering multiple potential fault areas on the surface of the power transformer.

[0030] The image acquisition equipment can employ a high-precision infrared thermal imager with a spectral response range set to the 8~14μm mid-wave infrared band, effectively adapting to power equipment temperature detection scenarios. The temperature measurement range covers -20℃ to 600℃, encompassing both normal operating temperatures of power transformers and capturing overheating temperatures under fault conditions. Furthermore, the image resolution is no less than 640×512 pixels, ensuring detailed identification of small-sized potential fault areas. The frame rate is ≥25fps, meeting the requirements for high-frequency time-series acquisition. During installation, a gimbal bracket with image stabilization can be used to fix the high-precision infrared thermal imager 3~5m directly in front of the transformer, balancing the integrity of the field of view with measurement accuracy. An integrated temperature and humidity sensor provides environmental parameters for subsequent temperature field calibration.

[0031] In a specific embodiment of the present invention, the acquisition period needs to be set to 10 min to 2 h based on the fault development characteristics of the power transformer. Within this period, no fewer than 100 consecutive time points are set, with the time interval between adjacent time points controlled between 1 s and 10 s, i.e., the acquisition frequency is 0.1 to 1 Hz. This ensures the capture of the continuous trend of temperature changes in the fault area and avoids missing key nodes in the fault propagation process due to excessively long time intervals. The acquired image sequence can adopt a single-frame image and metadata storage format. The single-frame image is a 16-bit grayscale image, with each pixel corresponding to a 16-bit value to represent the quantified temperature value. The metadata needs to include the acquisition timestamp (accurate to milliseconds), ambient temperature, ambient humidity, and the relative distance and angle between the high-precision infrared thermal imager and the power transformer.

[0032] It should also be noted that the potential fault area must encompass all critical heat-generating components of the power transformer, including winding terminals, bushing joints, tap changer contacts, core grounding points, tank welds and valve interfaces, and cooler fin connections. In practice, the elevation, horizontal, and focal length of the thermal imager are adjusted to ensure that a single frame image covers more than 80% of the transformer's critical components. The remaining areas are captured and stitched together in 2-3 groups, with an overlap rate of ≥10% at the stitching points to avoid omissions. Before acquisition, each potential fault area must be pre-marked using the thermal imager's annotation function, and the marking information is stored in the metadata. Furthermore, invalid images are removed from the acquired image sequence to ensure data source reliability; invalid images include frames with abnormal temperature data, blurred frames, and obstructed frames.

[0033] In this embodiment of the invention, after acquiring the high-frequency time-series infrared image sequence, the images in the sequence need to be preprocessed to eliminate spatial deviations, noise interference, and temperature quantization errors during the acquisition process, thereby establishing a unified high-quality data benchmark. The preprocessing includes image registration, noise filtering, and temperature field calibration. Specifically, image registration can eliminate spatial position shifts in the time-series images caused by slight equipment vibrations, installation displacements, etc., ensuring that the same potential fault area is in uniform spatial coordinates in images at different time points. Specifically, the first frame in the sequence is used as the reference frame, and all subsequent frames are aligned to this reference frame to avoid cumulative offset between multiple frames. The SIFT algorithm, which has scale invariance and rotation invariance, is used to extract key feature points between the reference frame and the frame to be registered, such as gray-level abrupt change areas like transformer edges and connector corners. Feature point matching is performed using the KNN algorithm, and mismatched pairs are eliminated with a threshold of Euclidean distance ≤ 1.5 times the minimum distance. Based on the effective matching point pairs, the affine transformation matrix is ​​solved using the least squares method to ensure that the spatial deviation of the registered feature points is ≤ 1 pixel. After applying this transformation matrix to the unregistered frames, pixel values ​​are filled by bilinear interpolation to ensure the continuity of image gray levels. Finally, the structural similarity index between the registered image and the reference frame is used as the verification standard to ensure that the spatial alignment accuracy meets the requirements of subsequent temporal feature extraction.

[0034] Noise filtering can eliminate interference from equipment electronic noise, environmental radiation noise, and other factors introduced during infrared image acquisition, avoiding temperature data distortion caused by random grayscale fluctuations while preserving temperature details in fault areas. Specifically, this invention employs a combination of Gaussian filtering and median filtering. First, the noise intensity is determined by statistically analyzing the variance of the image grayscale values. For weak noise (variance ≤ 5), only Gaussian filtering is used; for strong noise (variance > 5), a 3×3 median filtering window is used to remove salt-and-pepper noise before Gaussian filtering for smoothing. The Gaussian kernel size is selected as 3×3 or 5×5 based on the noise intensity, and an adaptive adjustment strategy is used for the standard deviation to ensure that temperature boundaries are not blurred while suppressing noise. The filtering effect is evaluated using the signal-to-noise ratio (SNR), requiring an improvement of ≥3dB compared to the original image. Temperature field calibration establishes a precise mapping relationship between the infrared image grayscale values ​​and the actual surface temperature of the power transformer, eliminating temperature quantization errors caused by factors such as equipment parameters, environmental conditions, and acquisition distance.

[0035] S2: Based on the spatiotemporal attention 3D convolutional neural network, the initial spatiotemporal features of the high-frequency temporal infrared image sequence are extracted through its spatiotemporal feature extraction layer. Then, the attention weight calculation layer generates an attention weight map that reflects the anomalous significance of each spatial location at different time points based on the initial spatiotemporal features. The attention weight map is then weighted and fused with the initial spatiotemporal features to output the attention weighted feature map.

[0036] This step involves synchronously capturing the spatial distribution characteristics and temporal evolution patterns of potential fault areas from preprocessed high-frequency temporal infrared image sequences, and using an attention mechanism to focus on the spatiotemporal locations of key anomalies to generate attention-weighted feature maps that are both targeted and identifiable.

[0037] In this embodiment of the invention, the input to the spatiotemporal feature extraction layer is the preprocessed high-frequency temporal infrared image sequence tensor. Where T is the total number of time points (time series length), H and W are the height and width of the original image, and C is the number of image channels (the infrared image defaults to 1 channel of temperature data).

[0038] Initial spatiotemporal feature extraction involves performing spatiotemporal joint convolution on the preprocessed high-frequency temporal infrared image sequence using 3D convolution operations. The formula is as follows:

[0039]

[0040] In the formula, The initial spatiotemporal feature map at time t, spatial location The eigenvalues ​​of the d-th channel; , , The time, height, and width dimensions of the 3D convolution kernel are specified, such as 3×3×3; C is the number of input image channels (1 by default for infrared images); D is the number of output feature channels. For the preprocessed infrared image sequence tensor ( ); 3D convolution kernel weights ( ), This is the bias term for the d-th channel.

[0041] The spatiotemporal feature extraction layer network structure of this invention employs stacked units of 3D convolutional layers, batch normalization, and activation functions, which ensures feature extraction capability while avoiding gradient vanishing. The 3D convolutional kernel size is... Time dimension It covers continuous time windows, capturing short-term temperature change trends, and has spatial dimensions. Adaptable to potential fault regions at different scales; time dimension Preserve complete temporal information without losing details in the temporal and spatial dimensions. (Balancing feature map resolution and computational cost to avoid redundancy); the number of feature channels D is gradually increased from the number of input channels C to 64~256, and the feature dimension is increased through multi-layer convolution stacking to enhance expressive power; space filling number This avoids the loss of image edge features and ensures that the input and output spatial dimensions match.

[0042] After processing by the spatiotemporal feature extraction layer, the initial spatiotemporal feature map is output. The initial spatiotemporal feature map is a four-dimensional tensor ( ), each of its elements Let be an initial spatiotemporal feature value, which simultaneously contains the time t and the spatial location of the feature map. Spatial features (such as temperature distribution differences and regional texture correlations) and temporal features; initial spatiotemporal feature map height and the width of the initial spatiotemporal feature map They are respectively:

[0043] ,

[0044] ,

[0045] In the formula, This indicates rounding down (integer division).

[0046] In one embodiment of the present invention, the attention weight calculation layer of the spatiotemporal attention 3D convolutional neural network adopts a soft attention mechanism, which uses a learnable query vector. Key vector Sum value vector Calculate the attention weight for each spatial location at each time point to generate an attention weight map. Specifically, this includes:

[0047] First, feature projection is performed to transform the initial spatiotemporal features. Each vector is mapped to a three-dimensional vector in the attention space using three independent learnable projection matrices and bias vectors, as shown in the formula:

[0048] ,

[0049] ,

[0050] ,

[0051] In the formula, , For attention feature dimension, ; For bias vectors, These respectively represent feature requirements, feature attributes, and feature content to be weighted.

[0052] Then, weight calculation is performed. In this invention, the attention weight for each spatial location is calculated using the scaled dot product attention formula and then normalized. The formula is as follows:

[0053] ,

[0054] In the formula, Let t be the spatial location. Attention weights, weight values This value reflects the significance of the anomaly at that location; the closer the value is to 1, the stronger the correlation between the spatial location and the fault. , The query vector and key vector are obtained by projecting the initial features, with dimensions... ; The attention feature dimension is (D / 8~D / 2). This is a scaling factor used to mitigate gradient vanishing caused by excessively large inner product results in high-dimensional features; The function ensures that the sum of the weights of all spatial locations is 1.

[0055] Finally, an attention weight map is generated. The elements of the attention weight map directly correspond to the calculated attention weights, and its spatiotemporal dimension is the same as the initial spatiotemporal features. Figure 1 One-to-one correspondence, intuitively presenting the distribution of anomalies at each location.

[0056] In this embodiment of the invention, attention-weighted fusion applies the focusing effect of the attention mechanism to the feature level, strengthening the feature signals of key abnormal regions through weighted operations while weakening interference from irrelevant regions. Specifically, due to the attention weight map... It is a single channel (containing only weight information), and the initial spatiotemporal feature map for The channels contain multi-dimensional features; therefore, the attention weight map needs to be processed first. Perform channel expansion to obtain This involves copying the single-channel weights to D feature channels to ensure dimensionality matching without changing the relative magnitudes of the weights; subsequently, element-wise multiplication is used. The initial spatiotemporal features are fused with the expanded attention weight map using the following formula:

[0057] ,

[0058] In the formula, For attention-weighted feature maps, the fusion process amplifies the feature values ​​at high-weight (high anomaly salience) locations in the initial spatiotemporal features, while suppressing the feature values ​​at low-weight (low anomaly salience) locations, thereby... It not only preserves complete spatiotemporal structural information, but also highlights the core features directly related to the fault; This is the initial spatiotemporal feature map; The expanded attention weight map is created by copying the single-channel weights to the D channel, with dimensions equal to... Consistent; This is element-wise multiplication (weighted per pixel).

[0059] It should also be noted that all learnable parameters of the spatiotemporal attention 3D convolutional neural network, including the 3D convolutional kernel... Projection matrix , , The bias vector and other parameters were optimized through end-to-end training using a historical infrared image sequence dataset labeled with fault type, root cause region, and propagation path. The training process used the cross-entropy loss for root cause region identification as the optimization objective, employing the Adam optimizer to iteratively update parameters. This ensured that the network could adaptively learn the spatiotemporal characteristics and anomaly significance criteria of power transformer faults, thereby accurately extracting effective features.

[0060] S3: Attention-weighted feature map based The attention-weighted feature map is calculated through the causal discovery module. The causal relationship strength between the temporal feature vectors of each potential fault region is determined, and a fault propagation network is constructed based on the causal relationship strength, with potential fault regions as nodes and causal relationship strength as directed edges.

[0061] This step involves analyzing the attention-weighted feature map. The dynamic causal relationships between potential fault regions are explored, and the dispersed regional characteristics are transformed into a structured fault propagation network by quantifying the strength of causal relationships.

[0062] In this embodiment of the invention, temporal feature vector extraction is the foundation of causal analysis. First, attention-weighted feature maps are extracted. According to predefined potential fault areas Spatial aggregation is performed (N is the total number of regions) to obtain the temporal feature representation of each potential fault region; then, for each potential fault region... For i=1,2,...,N, determine its position in the attention-weighted feature map. The set of corresponding pixel positions on It maps one-to-one with the spatial locations of the pre-labeled regions in S1, and obtains the precise location on the attention-weighted feature map through coordinate scaling; finally, it targets the potential fault regions. The average eigenvalue of each time point t across all feature channels is calculated to form a single-time-point feature vector, thereby obtaining the temporal feature vector of the potential fault region. The specific calculation formula is as follows:

[0063] ,

[0064] In the formula, Potential fault area The average feature value of the d-th channel at time t not only preserves the abnormal features of attention focus, but also reflects the dynamic changes in the region state. d=1,2,...,D is the feature channel index. Potential fault area The set of pixels at time t in the attention-weighted feature map; Potential fault area The number of pixels in the pixel set at time t of the feature map; This is an attention-weighted feature map.

[0065] In one embodiment of the present invention, in step S3, the causal discovery module uses Granger causality analysis to quantify the strength of causality through a vector autoregression (VAR) model and an F-test.

[0066] Specifically, firstly, a VAR model is constructed for any two potential fault regions. (Candidate reason region) and (Candidate result region), using the temporal feature vectors of both as input, a VAR(p) model is established. The temporal feature vectors are simplified to single-dimensional features, and the mean of D channels is taken. , Its model expression is:

[0067] ,

[0068] In the formula, Potential fault area The single-dimensional temporal feature value of the (candidate result region) at time t is obtained by dimensionality reduction of the region temporal feature vector in step S3, reflecting the state of the potential fault region at time t; p is the lag order of the VAR model, indicating that the current state is affected by the states of the previous p time points, and its value range is... Determined through the AIC criteria; Potential fault area Its own hysteresis k-th order coefficient quantifies the potential fault region. The strength of the influence of the state at time tk on its own state at time t, such as the continuity of the historical state on the current state; Potential fault area The single-dimensional temporal feature value at time tk (its own state lagged by k steps). Potential fault area (Candidate reason region) The lag k-th order influence coefficient is used to determine Is it The Granger cause; if the coefficient is significantly non-zero, it indicates... The historical situation will affect The current state; Potential fault area The single-dimensional temporal feature value at time tk (the state of the cause region lagging k steps). White noise error term ( ), representing random disturbances that the model cannot explain, with a mean of 0, constant variance, and independent of errors at other times.

[0069] An F-test was then performed to check the null hypothesis. ,Right now right Without Granger causality, calculate the validity of the null hypothesis separately (excluding...). Sum of squared residuals (lagging terms) The alternative hypothesis holds true (including) Sum of squared residuals (lagging terms) Construct the F-statistic, whose mathematical expression is:

[0070] ,

[0071] In the formula, Let F be the statistic, and let F be the number of degrees of freedom. The F-distribution is used to test the significance of causal relationships; This is the sum of squared residuals under the null hypothesis (no causal relationship); The sum of squared residuals under the alternative hypothesis (causal relationship); λ represents the lag order of the VAR model; T represents the total number of time points.

[0072] like ,in, At the significance level, If the value is critical, then the null hypothesis is rejected, and the determination is made. right There is a significant causal relationship, and the expression for the strength of the causal relationship is:

[0073] ,

[0074] Indicates potential fault areas right The causal relationship is strong, and its value ranges from [0,1]. The larger the value, the stronger the causal relationship.

[0075] If the test fails, then By traversing all potential fault areas This yields the complete causal strength matrix. .

[0076] In one embodiment of the present invention, in step S3, the fault propagation network is a directed weighted graph; the nodes of the directed weighted graph correspond to the spatial location of each potential fault region, and the weight of the directed edge connecting two nodes is defined by the strength of the causal relationship.

[0077] In this embodiment of the invention, the fault propagation network construction transforms the strength of causal relationships into a visualized, structured network model, wherein the directed weighted graph is... ,in, For a set of nodes, Each node corresponds one-to-one with a potential fault area. Bearing potential fault areas The temporal characteristics and comprehensive attributes; Let be the set of edges. Only when right When a significant causal relationship exists, there exists from point to The directed edge, whose direction is the direction of fault propagation; This is the weight matrix. In, among which elements This directly quantifies the causal strength of edges. The network clearly shows the propagation path of faults from the cause region to the result region.

[0078] It should be added that, in the process of calculating the strength of causal relationship, for time series vectors with multiple feature channels, dimensionality reduction can be achieved through principal component analysis (PCA) before constructing the VAR model, thus avoiding computational redundancy caused by high dimensionality.

[0079] S4: The attention weights representing the saliency of a region in the attention weight map are weighted and combined with the causal relationship strength to obtain the comprehensive importance score of each potential fault region; the fault propagation network containing the comprehensive importance score is input into the graph neural network; the neighborhood information of each node in the fault propagation network is aggregated through the message passing mechanism and the node representation is updated; the root cause region as the origin of the fault is identified based on the updated node representation.

[0080] In this embodiment of the invention, the attention weight and the causal relationship strength are weighted and combined using an adaptive weighting method based on the sigmoid function. The weighting coefficients of the attention weight and the causal relationship strength are dynamically optimized through the training process of the graph neural network.

[0081] Specifically, firstly, regarding the attention weight map Causal strength matrix Set of potential fault areas For each potential fault area The mean attention weights at all time points and all corresponding pixel locations are calculated to obtain the region aggregation attention weights. The corresponding calculation formula is:

[0082] ,

[0083] in, Potential fault area Aggregated attention weights This reflects the overall significance of anomalies in the region; the larger the value, the higher the overall significance of anomalies in the region. This is an attention weight map; Potential fault area The set of pixels in the feature map at time t; T is the total number of time points.

[0084] Secondly, based on the causal relationship strength matrix Calculate the aggregation causality strength for each region. The calculation formula is:

[0085] ,

[0086] in, Potential fault area The strength of aggregation causality, Positive values ​​indicate potential root cause attributes, while negative values ​​indicate that a region is more likely to affect other regions (derived fault attributes). Potential fault area right The strength of the causal relationship; N is the total number of potential fault areas.

[0087] Finally, an adaptive weighting method based on the sigmoid function is adopted to... With normalization By combining the results, a comprehensive importance score is obtained. The formula is:

[0088] ,

[0089] in, Potential fault area The overall importance score, The larger the value, the higher the probability that the region is the root cause; The sigmoid activation function is used for adaptive weighting. , The weighting coefficients can be dynamically optimized during the training process of the graph neural network; for The normalization operation will Mapped to the [0,1] interval, and Dimensions are consistent.

[0090] In one embodiment of the present invention, the graph neural network adopts a graph attention network (GAT) architecture, which enhances the expressive ability of neighborhood information aggregation during message passing through a multi-head attention mechanism.

[0091] Propagating the fault network With overall importance score Combined, the input to the graph neural network is constructed. Each node (corresponding to potential fault areas) The initial eigenvectors of ) It is composed of the regional average time series characteristics and the comprehensive importance score, that is:

[0092] ,

[0093] in, For the corresponding potential fault region in the graph neural network nodes The initial feature vector (superscript (0) indicates the 0th layer, i.e., the initial input) has a dimension of (D represents the number of feature channels, and +1 corresponds to the overall importance score). Potential fault area The average temporal feature on the d-th feature channel is derived from the temporal feature vector of that region. The average over the time dimension is obtained. Where T is the total number of times. Potential fault area The temporal characteristics of the d-th channel at time t; Potential fault area The comprehensive importance score integrates the anomalous salience (attention weight) and causal correlation (causal relationship strength) of the potential fault region, directly reflecting the potential probability that the potential fault region is the root cause.

[0094] The adjacency relationship follows the set of directed edges in the fault propagation network. The weights of directed edges are determined by the strength of causality. Quantify the impact of neighboring nodes to ensure that information transmission conforms to the fault propagation logic.

[0095] This invention uses a message passing mechanism to allow each node to aggregate the feature information of its neighboring nodes and update the node representation by combining it with its own features, thus highlighting the unique attributes of the root cause region.

[0096] Let the number of attention heads be K ( ), through a learnable weight matrix Projecting the initial features of the nodes, calculate the nodes. With neighboring nodes The attention coefficient is calculated using the following formula:

[0097] ,

[0098] in, For activation function, By introducing nonlinear transformations, the feature representation capability is enhanced; is the learnable attention vector of the k-th attention head, used to weight and score the concatenated node features; For neighboring nodes The feature vector obtained by mapping the features through the projection matrix of the k-th attention head. , Let be the learnable projection matrix of the k-th attention head, with the number of rows equal to the initial feature dimension (D+1, where D is the number of feature channels and +1 corresponds to the overall importance score), and the number of columns equal to the output dimension of each attention head. Its function is to linearly map high-dimensional initial features to a low-dimensional space, while simultaneously learning the weight assignments of the features, highlighting the features that are more important for attention calculation. Feature dimensions for each attention head; For nodes The initial feature vector; For neighboring nodes The feature vector obtained by mapping the features through the projection matrix of the k-th attention head. ; For vector concatenation operations, and Concatenating along the feature dimensions yields a result with dimension . The joint feature vector is used to capture the interaction information between the two nodes.

[0099] Next, for the nodes The attention coefficients of all neighboring nodes are normalized using softmax to obtain the normalized attention coefficients. The formula is:

[0100] ,

[0101] In the formula, For the k-th attention head, node For neighboring nodes The normalized attention coefficient has a value range of [0,1]. For nodes The Softmax function operation is performed on all neighboring nodes j; For the k-th attention head, node right Attention coefficient; For nodes The set of neighboring nodes, containing Self and all Nodes with a direct causal relationship, i.e., those in the fault propagation network... Connected nodes; For neighborhood set Other node indices are used to traverse all neighboring nodes to calculate the normalized denominator.

[0102] Then, based on the normalized attention coefficients, the projection features of neighboring nodes are aggregated to obtain the node update features of the k-th attention head, and the corresponding formula is:

[0103] ,

[0104] In the formula, It is the sigmoid function;

[0105] Finally, the node update features of the K attention heads are averaged to obtain the node... Final update indicates The corresponding formula is:

[0106] ,

[0107] In the formula, For nodes The final update representation after processing by the Lth layer of the Graph Neural Network (GAT) is where L is the number of network layers, which can be 1 to 2 layers.

[0108] The determination of root cause regions is based on the final updated representation of nodes. The probability of each node being a root cause region is calculated through classification, and the region with the highest probability is selected as the root cause region. The calculation formula is as follows:

[0109] ,

[0110] In the formula, Potential fault area The probability of being the root cause. ,and , Larger values ​​indicate a larger node The higher the confidence level of the corresponding region as the root cause; This is the classification weight vector; Let be the final representation of node i in the graph neural network output; L is the number of layers in the graph neural network.

[0111] This invention selects the region with the highest root cause probability as the root cause region of the fault origin. ,Right now: ,

[0112] S5: Based on the root cause region and the fault propagation path extracted from the fault propagation network, generate and output the spatial location information of the root cause region, the visualized fault propagation path map, and the diagnostic results of the risk level assessment of the future deterioration of the root cause region, thereby realizing fault prediction and health management.

[0113] In this embodiment of the invention, the spatial localization of the root cause region is based on the size mapping relationship between the attention-weighted feature map in step S2 and the original infrared image in step S1. The coordinates of the root cause region on the attention-weighted feature map are converted inversely to the physical pixel coordinates of the original image, ensuring that maintenance personnel can quickly locate the actual position.

[0114] Specifically, based on root cause regions Set of pixel locations on the attention-weighted feature map By combining the size ratio of the original infrared image in step S1 with that of the attention-weighted feature map in step S2, the coordinates of the attention-weighted feature map are reverse-mapped back to the original infrared image.

[0115] The spatial positioning information specifically includes the area identifier, bounding box coordinates, area center coordinates, and the current average temperature of the area. The area identifier corresponds to the name of the potential fault area pre-marked in step S1, such as winding terminal A.

[0116] The bounding box coordinates are ,in,

[0117] ,

[0118] ,

[0119] ,

[0120] ;

[0121] In the formula, , This is the space scaling factor. , , The height and width of the original infrared image. Add height and width to the attention-weighted feature map.

[0122] The formula for calculating the coordinates of the region center is:

[0123] ,

[0124] ;

[0125] The current average temperature of this area is calculated based on the temperature field data after preprocessing in step S1. The calculation formula is as follows:

[0126] ,

[0127] In the formula, The current average temperature (°C) of the root cause region; The root cause region is the set of pixels in the original image; The true temperature of pixel (x,y) (after S1 temperature field calibration).

[0128] A visualized fault propagation path diagram is an abstract representation of the fault propagation network. Transformed into intuitive graphical representations, clearly presenting the propagation logic of root cause, intermediate derivative areas, and terminal impact areas.

[0129] Specifically, from the fault propagation network in step S3 In the middle, with root cause nodes Starting from the causal relationship strength of directed edges, Sort the nodes and select the neighboring node with the strongest causal relationship on each branch as the next hop to form one or more main propagation paths. When there are multiple parallel propagation branches, retain the top 3 branches with the highest sum of causal relationship strength. The corresponding fault propagation path expression is:

[0130] ,

[0131] in, For a fault propagation path, it must satisfy the following conditions: , , Root cause region For the region The strength of the causal relationship; Root cause region For any other region ( The strength of the causal relationship; For the middle area For the region The strength of the causal relationship; For the middle area For any other region The strength of the causal relationship; The root cause region; For the intermediate derivative region in the fault propagation process ( , ) and end-effect area ( All of these originate from the set of potential fault regions pre-marked in step S1.

[0132] For example, if the root cause region right The causal strength is 0.8 ( ),right The strength is 0.5 ( According to the corresponding conditions, the first step of the path must be... (Because 0.8 ≥ 0.5), not .

[0133] In the specific visualization, root cause nodes can be marked with solid red circles, intermediate derived nodes with hollow yellow circles, and terminal nodes with hollow blue circles. The node size can be correlated with the overall importance score. Positive correlation; the thickness of directed edges correlates with the strength of the causal relationship. Positive correlation, arrows point in the direction of propagation, and specific intensity values ​​are marked on the side; a thumbnail of the original infrared image is superimposed on the background layer, and nodes are arranged according to their actual spatial locations to achieve a dual correspondence between logical propagation and physical spatial location.

[0134] In one embodiment of the present invention, risk level assessment of future deterioration of the root cause region is a core part of fault prediction and health management, including: establishing a mapping relationship between temperature change rate and fault development speed based on a historical fault case library, and predicting the probability of deterioration within a specified time window in the future by combining the current temperature trend of the root cause region, so as to provide forward-looking guidance for maintenance decisions.

[0135] First, the rate of temperature change in the root cause region is calculated. Specifically, the most recent temperature change rate in the time series image sequence from step S1 is selected. A time datum is used to ensure trend stability; the formula is:

[0136] ,

[0137] In the formula, The rate of temperature change (°C / hour); The initial average temperature (°C) within the time window. The final average temperature (°C) within the time window; The number of time points included in the time window (≥30); The time interval (in hours) for time-series infrared images; where,

[0138] ,

[0139] ,

[0140] Subsequently, a mapping relationship was established based on a historical fault case database. This database must contain at least 1,000 fault records of similar power transformers, with each record including the temperature change rate. Fault development speed (Unit: days / fault level, i.e., the time required for the current state to deteriorate to a higher level of fault), Fault type The following was obtained by fitting the data using a linear regression model:

[0141] ,

[0142] in , , The regression coefficients are obtained using the least squares method. ; Encode the current fault type, such as poor contact as 1, insulation aging as 2, and core overheating as 3.

[0143] Then predict the future within a specified time window. The probability of deterioration within 12 hours can be configured for 12 hours, 24 hours, and 72 hours, as shown in the formula:

[0144] ,

[0145] In the formula, As a risk threshold, It can be set to 1.2. .

[0146] Finally, the risk levels are classified, among which... The risk level is low, no emergency repairs are needed, continuous monitoring is sufficient. The risk level is medium; repairs will be arranged within 24 hours. This is considered a high-risk situation. Immediately shut down the machine for repairs to prevent the malfunction from escalating.

[0147] The final diagnostic output is a structured report, which may include three parts: root cause area details, a visualized fault propagation path diagram, and a risk assessment report, to support fault prediction and health management decisions. The root cause area details include the area identifier, spatial coordinates, and current average temperature; the visualized fault propagation path diagram is a graphic file with nodes, directed edges, and intensity labels; the risk assessment report includes the temperature change rate, fault development speed prediction, probability of deterioration in a specified future time window, risk level, and corresponding maintenance recommendations.

[0148] According to the image analysis-based power transformer fault diagnosis method of the present invention, the present invention achieves accurate, dynamic, and root cause diagnosis of power transformer faults through image analysis technology. Specifically, by acquiring high-frequency temporal infrared image sequences and utilizing a spatiotemporal attention 3D convolutional neural network to extract initial spatiotemporal features and generate an attention weight map, key abnormal regions are effectively focused, improving the efficiency and targeting of fault feature extraction. A fault propagation network is constructed by combining a causal discovery module to quantify the causal relationship strength between potential fault regions, thereby distinguishing the root cause from derivative phenomena and accurately locating the root cause region. Furthermore, by aggregating neighborhood information and updating node representations through a graph neural network, the ability to analyze fault propagation paths is enhanced. Finally, diagnostic results including root cause spatial location, visualized propagation paths, and risk level assessment are output, overcoming the limitations of traditional static image analysis methods that cannot capture dynamic evolution, lack focus on key regions, and fail to analyze fault propagation paths, thus improving the accuracy of fault diagnosis.

[0149] Corresponding to the image analysis-based power transformer fault diagnosis method in the above embodiments, the present invention also proposes an image analysis-based power transformer fault diagnosis system.

[0150] An image analysis-based power transformer fault diagnosis system according to an embodiment of the present invention includes an image acquisition and preprocessing module, a spatiotemporal feature extraction and attention focusing module, a causal discovery and fault propagation network construction module, a graph learning and root cause localization module, and a diagnostic decision and health management output module.

[0151] The module comprises several components: an image acquisition and preprocessing module for acquiring high-frequency temporal infrared image sequences of power transformers, which include infrared images acquired at multiple consecutive time points covering multiple potential fault areas on the transformer surface; a spatiotemporal feature extraction and attention focusing module for extracting initial spatiotemporal features of the high-frequency temporal infrared image sequences using a spatiotemporal attention 3D convolutional neural network, and generating an attention weight map reflecting the anomalous significance of each spatial location at different time points based on the initial spatiotemporal features using an attention weight calculation layer; and a weighted fusion of the attention weight map and the initial spatiotemporal features to output an attention-weighted feature map. Finally, a causal discovery and fault propagation network construction module is used to calculate the causal relationships between the temporal feature vectors of each potential fault area derived from the attention-weighted feature map using a causal discovery module. The system calculates the relationship strength and constructs a fault propagation network with potential fault regions as nodes and causal relationship strength as directed edges based on the causal relationship strength. The graph learning and root cause localization module is used to weight and combine the attention weights representing regional saliency in the attention weight graph with the causal relationship strength to obtain the comprehensive importance score of each potential fault region. The fault propagation network containing the comprehensive importance score is input into the graph neural network. The neighborhood information of each node in the fault propagation network is aggregated through a message passing mechanism and the node representation is updated. Based on the updated node representation, the root cause region as the origin of the fault is identified. The diagnostic decision and health management output module is used to generate and output diagnostic results, including the spatial location information of the root cause region, a visualized fault propagation path map, and a risk level assessment of the future deterioration of the root cause region, based on the root cause region and the fault propagation path extracted from the fault propagation network.

[0152] According to embodiments of the present invention, the image analysis-based power transformer fault diagnosis system achieves accurate, dynamic, and root cause diagnosis of power transformer faults through image analysis technology. Specifically, by acquiring high-frequency temporal infrared image sequences and utilizing a spatiotemporal attention 3D convolutional neural network to extract initial spatiotemporal features and generate an attention weight map, it effectively focuses on key abnormal regions, improving the efficiency and targeting of fault feature extraction. Combined with a causal discovery module, a fault propagation network is constructed to quantify the causal relationship strength between potential fault regions, thereby distinguishing between the root cause and derivative phenomena of the fault and accurately locating the root cause region. Furthermore, by aggregating neighborhood information and updating node representations through a graph neural network, the system enhances the ability to analyze fault propagation paths. Finally, it outputs diagnostic results including root cause spatial location, visualized propagation paths, and risk level assessment. This overcomes the limitations of traditional static image analysis methods, such as their inability to capture dynamic evolution, lack of focus on key regions, and inadequate fault propagation path analysis, thus improving the accuracy of fault diagnosis.

[0153] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0154] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0155] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0156] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0157] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0158] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0159] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A power transformer fault diagnosis method based on image analysis, characterized in that, Includes the following steps: S1: Acquire a high-frequency time-series infrared image sequence of a power transformer, the high-frequency time-series infrared image sequence containing infrared images collected at multiple consecutive time points, covering multiple potential fault areas on the surface of the power transformer; S2: Based on the spatiotemporal attention 3D convolutional neural network, the initial spatiotemporal features of the high-frequency temporal infrared image sequence are extracted through its spatiotemporal feature extraction layer. Then, the attention weight calculation layer generates an attention weight map reflecting the anomalous significance of each spatial location at different time points based on the initial spatiotemporal features. The attention weight map is then weighted and fused with the initial spatiotemporal features to output an attention weighted feature map. S3: Based on the attention-weighted feature map, the causal relationship strength between the temporal feature vectors of each potential fault region derived from the attention-weighted feature map is calculated by the causal discovery module, and a fault propagation network with potential fault regions as nodes and causal relationship strength as directed edges is constructed according to the causal relationship strength. S4: The attention weights representing regional saliency in the attention weight map are weighted and combined with the causal relationship strength to obtain a comprehensive importance score for each potential fault region; the fault propagation network containing the comprehensive importance score is input into the graph neural network; The neighborhood information of each node in the fault propagation network is aggregated and the node representation is updated through a message passing mechanism. Based on the updated node representation, the root cause region that is the origin of the fault is identified. S5: Based on the root cause region and the fault propagation path extracted from the fault propagation network, generate and output a diagnostic result including the spatial location information of the root cause region, a visualized fault propagation path map, and a risk level assessment of the future deterioration of the root cause region.

2. The image analysis-based power transformer fault diagnosis method according to claim 1, characterized in that, In step S2, the attention weight calculation layer of the spatiotemporal attention 3D convolutional neural network adopts a soft attention mechanism, which calculates the attention weight of each spatial location at each time point through learnable query vectors, key vectors and value vectors to generate the attention weight map.

3. The image analysis-based power transformer fault diagnosis method according to claim 2, characterized in that, In step S3, the fault propagation network is a directed weighted graph; the nodes of the directed weighted graph correspond to the spatial location of each potential fault region, and the weight of the directed edge connecting two nodes is defined by the causal relationship strength.

4. The image analysis-based power transformer fault diagnosis method according to claim 3, characterized in that, In step S3, the causal discovery module uses Granger causality analysis to quantify the strength of the causal relationship through a vector autoregression model and an F-test.

5. The image analysis-based power transformer fault diagnosis method according to claim 4, characterized in that, In step S4, the attention weights and the causal relationship strengths are weighted and combined using an adaptive weighting method based on the sigmoid function.

6. The image analysis-based power transformer fault diagnosis method according to claim 5, characterized in that, The graph neural network adopts a graph attention network architecture, which enhances the expressive ability of neighborhood information aggregation during message passing through a multi-head attention mechanism.

7. The image analysis-based power transformer fault diagnosis method according to claim 6, characterized in that, The risk assessment of the future deterioration of the root cause region includes: Based on a historical failure case library, a mapping relationship between temperature change rate and failure development speed is established, and the probability of deterioration within a specified time window is predicted by combining the current temperature trend of the root cause region.

Images