Transformer fault detection method and device based on multi-modal data fusion

Abstract

The invention relates to a transformer fault detection method and device based on multi-modal data fusion. The method comprises the steps that operation state sensing data of multiple modes of different data sources of a transformer are acquired, initial operation state features are extracted from the operation state sensing data of each mode, each mode has a corresponding operation state type, and the initial operation state features are extracted from the operation state sensing data of each mode; aligning each initial operation state feature to a target dimension, carrying out fusion based on an attention weight, carrying out fusion based on a data source, obtaining target operation state features, determining an initial fault detection result, corresponding to each operation state type, of the transformer according to each target operation state feature, and obtaining a fault detection result of the transformer according to probability distribution of each initial fault detection result. And fusing the initial fault detection results corresponding to the operation state types to obtain target fusion data, and outputting a target fault detection result of the transformer according to the target fusion data. By adopting the method, high-precision, high-applicability and high-reliability transformer fault detection can be provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment detection, in particular to a transformer fault detection method and device based on multi-modal data fusion, computer equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] With the continuous expansion of the power grid scale and the sustained growth of power demand, the operation stability of the transformer as the core equipment of the power system directly affects the safe and reliable operation of the power grid. If the early fault of the transformer is not discovered in time, it may lead to winding short circuit, iron core overheating, insulation breakdown and other serious consequences, and even cause large-scale power outages and equipment damage. Therefore, it is of great significance to realize accurate monitoring and fault warning of the operation state of the transformer to ensure the reliability of the power grid and reduce operation and maintenance costs.

[0003] In related technologies, three independent detection systems of partial discharge detection, temperature detection and overvoltage detection are used to detect transformer faults. However, this method is not only limited by the singleness of data, but also leads to one-sided fault detection results due to the independence of each system and the lack of information interaction, which makes it difficult to reflect the actual fault condition of the transformer. SUMMARY

[0004] Therefore, it is necessary to provide a transformer fault detection method, device, computer equipment, computer readable storage medium and computer program product based on multi-modal data fusion, which can realize high-precision transformer fault detection, in view of the above technical problems.

[0005] In a first aspect, the present application provides a transformer fault detection method based on multi-modal data fusion, comprising:

[0006] Obtaining running state sensing data of multiple modalities of different data sources of a transformer, and extracting initial running state features of the running state sensing data of each modality; wherein each modality has a corresponding running state type;

[0007] Aligning each initial running state feature to a target dimension to obtain a first aligned running state feature;

[0008] Determining the attention weights between all first aligned running state features, updating each first aligned running state feature according to the attention weights, and obtaining a second aligned running state feature;

[0009] Fusing the second aligned running state features belonging to the same data source to obtain a target running state feature;

[0010] According to each of the target operating state feature, determine the initial fault detection result corresponding to each of the operating state type of the transformer;

[0011] According to the probability distribution of each of the initial fault detection result, fuse the initial fault detection result corresponding to each of the operating state type to obtain target fusion data;

[0012] According to the target fusion data, output the target fault detection result of the transformer.

[0013] In one of the embodiments, the determination of the attention weight between each of the first aligned operating state features, updating each of the first aligned operating state features according to the attention weight, comprises:

[0014] Determine the feature similarity between each of the first aligned operating state features, and determine the similarity sum of all the feature similarities;

[0015] Determine the ratio of each of the feature similarities to the similarity sum to obtain the attention weight between each of the first aligned operating state features;

[0016] Traverse each of the first aligned operating state features;

[0017] In each round of the traversal, obtain other aligned operating state features except the first aligned operating state feature of the current round, and the target attention weight associated with the other aligned operating state features, and use the target attention weight to weight and fuse the other aligned operating state features to update the first aligned operating state feature of the current round.

[0018] In one of the embodiments, the fusion of the second aligned operating state features belonging to the same data source to obtain the target operating state feature comprises:

[0019] Determine the information metric value of each of the second aligned operating state features belonging to the same data source;

[0020] Determine the sum of all the information metric values belonging to the same data source to obtain the information amount sum;

[0021] Determine the ratio of each of the information metric values to the corresponding information amount sum to obtain the second information weight of each of the second aligned operating state features belonging to the same data source;

[0022] Fuse all the second aligned operating state features belonging to the same data source according to the corresponding second information weight to obtain the target operating state feature.

[0023] In one of the embodiments, the fusing the initial fault detection results corresponding to each of the operation state types according to the probability distribution of each of the initial fault detection results to obtain target fusion data comprises:

[0024] converting each of the initial fault detection results into a corresponding evidence body, wherein the evidence body is used to represent the probability distribution of the initial fault detection result;

[0025] determining the evidence body similarity between each of the evidence bodies, and determining the credibility weight of each of the evidence bodies according to each of the evidence body similarities;

[0026] determining the first information amount weight of each of the evidence bodies, wherein the first information amount weight is used to represent the uncertainty of the evidence body;

[0027] fusing the credibility weight and the first information amount weight of each of the evidence bodies to obtain the fusion weight of each of the evidence bodies;

[0028] fusing each of the evidence bodies according to each of the fusion weights to obtain the target fusion data.

[0029] In one of the embodiments, the determining the evidence body similarity between each of the evidence bodies, and determining the credibility weight of each of the evidence bodies according to each of the evidence body similarities comprises:

[0030] determining the triangular divergence between each of the evidence bodies, and determining the similarity matrix of each of the evidence bodies according to each of the triangular divergences, wherein the similarity matrix is used to represent the evidence body similarity;

[0031] normalizing each of the similarity matrices to obtain the credibility weight of each of the evidence bodies.

[0032] In one of the embodiments, the aligning each of the initial operation state features to a target dimension to obtain aligned operation state features comprises:

[0033] determining the data dimension of each of the initial operation state features, and determining the dimension weight of each of the initial operation state features according to each of the data dimensions and the target dimension;

[0034] inputting each of the initial operation state features and the dimension weight corresponding to each of the initial operation state features into a preset linear transformation function to obtain the aligned operation state features.

[0035] In a second aspect, the present application further provides a transformer fault detection device based on multi-modal data fusion, comprising:

[0036] The acquisition module is configured to acquire operation state sensing data of multiple modalities of different data sources of a transformer, and extract initial operation state features from the operation state sensing data of each modality; each modality has a corresponding operation state type;

[0037] The first alignment module is configured to align the initial operation state features to a target dimension to obtain first aligned operation state features;

[0038] The second alignment module is configured to determine attention weights between all the first aligned operation state features, update the first aligned operation state features according to the attention weights, and obtain second aligned operation state features;

[0039] The first fusion module is configured to fuse the second aligned operation state features belonging to the same data source to obtain target operation state features;

[0040] The first fault detection module is configured to determine initial fault detection results corresponding to each operation state type of the transformer according to the target operation state features;

[0041] The second fusion module is configured to fuse the initial fault detection results corresponding to each operation state type according to probability distribution of the initial fault detection results, and obtain target fusion data;

[0042] The second fault detection module is configured to output target fault detection results of the transformer according to the target fusion data.

[0043] In a third aspect, the present application also provides a computer device, including a memory and a processor, the memory stores a computer program, and the processor implements the transformer fault detection method based on multi-modal data fusion as above when executing the computer program.

[0044] In a fourth aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the transformer fault detection method based on multi-modal data fusion as above.

[0045] In a fifth aspect, the present application also provides a computer program product, including a computer program, and the computer program is executed by a processor to implement the transformer fault detection method based on multi-modal data fusion as above.

[0046] The transformer fault detection method, device, computer device, computer readable storage medium and computer program product provided by the method can preliminarily align each initial operation state feature in the data dimension by aligning each initial operation state feature to a target dimension to obtain first aligned operation state features, solve the difference problem of modal data of different dimensions, determine the attention weight between each of the first aligned operation state features, update each of the first aligned operation state features according to the attention weight to obtain second aligned operation state features, enhance the expression ability of different modal data features to key state information through the second aligned operation state features, and realize multi-level information integration from local to global through data fusion of merging second aligned operation state features of the same data source and cross-modal decision fusion based on probability distribution. Then, a target fault diagnosis result with high confidence and strong interpretability is output, the limitation of single modal data is effectively overcome, and the reliability of transformer fault detection in a complex operation environment is improved. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other related drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 The application environment schematic diagram of the transformer fault detection method based on multi-modal data fusion provided by an embodiment of the present application;

[0049] Figure 2 The step flowchart of the transformer fault detection method based on multi-modal data fusion provided by an embodiment of the present application;

[0050] Figure 3 The model structure schematic diagram of the transformer fault detection model based on multi-modal data fusion provided by an embodiment of the present application;

[0051] Figure 4 The structure schematic diagram of the hierarchical multi-modal learning edge intelligent processing framework provided by an embodiment of the present application.

[0052] Figure 5 The structure block diagram of the transformer fault detection device provided by an embodiment of the present application;

[0053] Figure 6 The structure schematic diagram of the computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application.

[0055] It should be noted that the terms "first", "second" and the like used in the present application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "include" and "have" and any variations thereof used in the present application are intended to cover non-exclusive inclusion. The term "a plurality of" used in the present application means two or more. The term "and / or" used in the present application means one of the options or any combination of multiple options.

[0056] The transformer fault detection method based on multi-modal data fusion provided by the embodiments of the present application can be applied to the application environment as shown in the figure. Figure 1 As shown in the figure, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data required to be processed by the server 104. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers. The terminal 102 initiates a transformer fault detection request to the server 104, and the server 104 responds to the request to obtain the running state sensing data of multiple modes of different data sources of the transformer, and extracts the initial running state features of each mode of the running state sensing data. Each mode has a corresponding running state type. The server 104 aligns each initial running state feature to the target dimension to obtain the first aligned running state feature. The server 104 determines the attention weight between all first aligned running state features, updates each first aligned running state feature according to each attention weight, and obtains the second aligned running state feature. The server 104 fuses the second aligned running state features belonging to the same data source to obtain the target running state feature. The server 104 determines the initial fault detection result corresponding to each running state type of the transformer according to each target running state feature. The server 104 fuses the initial fault detection results corresponding to each running state type according to the probability distribution of each initial fault detection result to obtain the target fusion data. The server 104 outputs the target fault detection result of the transformer according to the target fusion data, and returns the target fault detection result to the terminal 102.

[0057] The terminal 102 can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, unmanned aerial vehicles, low-altitude aerial vehicles, Internet of Things devices, and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle-mounted device, a projection device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The head-mounted device can be a virtual reality (VR) device, an augmented reality (AR) device, smart glasses, etc. The server 104 can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0058] In an exemplary embodiment, as shown in Figure 2 , a transformer fault detection method based on multi-modal data fusion is provided. The method is applied to the server 104 in Figure 1 for example, and the method comprises the following steps.

[0059] In step 202, the operating state sensing data of multiple modalities of different data sources of the transformer is obtained, and the initial operating state features of each modality of the operating state sensing data are extracted. Each modality has a corresponding operating state type.

[0060] Different data sources can be different signal sources, such as different sensors. The operating state sensing data of multiple modalities can include one or more of core grounding current sensing data, temperature and humidity sensing data, noise sensing data, oil temperature and oil level sensing data, oil chromatography sensing data, load current sensing data, ultrasonic sensor data, ultra-high frequency sensing data, and high-frequency current sensing data. Different operating state types of modalities can reflect the corresponding operating state of the transformer.

[0061] In a specific implementation, the initial operating state features can be extracted by using corresponding feature extraction algorithms according to the data characteristics of different operating state sensing data, such as data type, data distribution, etc.

[0062] In some examples, the time domain feature extraction and frequency domain feature extraction are performed on the operating state sensing data of the vibration signal type, the time-frequency feature extraction is performed on the operating state sensing data of the ultrasonic signal type, and the image feature extraction is performed on the operating state sensing data of the infrared temperature signal type.

[0063] Specifically, the time domain feature extraction adopts a bidirectional gated recurrent unit (BGRU), which can consider the context information before and after the vibration signal and effectively capture the time sequence dependence in the signal. Let the vibration time domain signal sequence be X = {x1, x2,..., xT}, where T is the time sequence length, and the BGRU forward propagation process is implemented by the following formula:

[0064]

[0065] wherein, is the update gate vector at the current time, and σ is the Sigmoid activation function, is the weight matrix of the update gate, is the hidden state at the previous time, is the input vector at the current time, is the reset gate vector at the current time, is the weight matrix of the reset gate. is the candidate hidden state at the current time, tanh is the hyperbolic tangent activation function, and W is the weight matrix of the candidate hidden state, is the final hidden state at the current time.

[0066] The backward propagation process adopts a similar mechanism, and the final time sequence feature is obtained by splicing the bidirectional hidden states:

[0067]

[0068] This bidirectional processing mechanism ensures that the model can fully understand the time sequence characteristics of the vibration signal, laying a foundation for subsequent analysis.

[0069] For frequency domain feature extraction, a multi-scale convolutional neural network can be used, which can analyze the frequency domain features of the vibration signal from different scales. First, the vibration signal is subjected to fast Fourier transform to obtain the frequency domain representation Then, the multi-scale CNN is used for feature extraction:

[0070]

[0071] wherein, represents the convolution feature map obtained by the kth filter. ReLU is a linear rectification activation function that sets negative values to zero. is a one-dimensional convolution operation The input is one-dimensional time sequence data. is the weight parameter of the kth filter (convolution kernel). is the bias parameter corresponding to the kth filter. is the result of pooling the k-th convolutional feature map. is a one-dimensional max-pooling operation for downsampling. is the spliced result of features extracted from multiple different filter paths, representing frequency domain (or multi-scale) features. is the operation of concatenating multiple tensors along the specified dimension, respectively represent the convolution kernel size of different scales.

[0072] This multi-scale design enables the model to capture both local details and global features, enhancing the expressiveness of the features.

[0073] For time-frequency feature extraction, since the ultrasonic signal contains key information of internal partial discharge and mechanical vibration of the transformer, its time-frequency characteristics are crucial for fault diagnosis. Wavelet neural network can be used for in-depth time-frequency analysis, represented as:

[0074]

[0075] where a is the scale parameter controlling the stretching of the wavelet, b is the translation parameter controlling the position of the wavelet, and ψ(t) is the wavelet basis function. This transformation can provide joint information of the signal in time and frequency domains.

[0076] Wavelet packet decomposition further enhances the precision of feature extraction, and its energy feature extraction formula is:

[0077]

[0078] where d j,k [n] is the wavelet packet coefficient, j is the decomposition level, and k is the node index. By analyzing the energy distribution of different frequency bands, the fault features in the ultrasonic signal can be accurately identified.

[0079] For image feature extraction, the infrared thermal image reflects the temperature distribution on the surface of the transformer and is an important means for detecting overheating faults. An improved ResNet-50 network model can be used for feature extraction, and the core residual block of this network model is represented as:

[0080]

[0081]

[0082] where, is the identity mapping, is the residual function, and f is the ReLU activation function. This residual structure effectively solves the gradient vanishing problem of deep networks, ensuring the depth and accuracy of feature extraction.

[0083] Step 204, aligning each of the initial operation state features to a target dimension to obtain a first aligned operation state feature;

[0084] In a specific implementation, the initial operation state features can be aligned to the unified target dimension through a full connection layer of a pre-trained model to obtain the first aligned operation state features.

[0085] In some embodiments, the aligning each of the initial operation state features to the target dimension to obtain an aligned operation state feature comprises:

[0086] determining a data dimension of each of the initial operation state features, and determining a dimension weight of each of the initial operation state features according to each of the data dimensions and the target dimension;

[0087] inputting each of the initial operation state features and the dimension weight corresponding to each of the initial operation state features into a preset linear transformation function to obtain the aligned operation state feature.

[0088] In a specific implementation, each initial operation state feature is represented as a vector F i (i∈1,2...M), the data dimension of each initial operation state feature is d i (i∈1,2...M), and the target dimension is D, then the dimension weight w i =[D,d i ](i∈1,2...M) of each initial operation state feature represented by a matrix is constructed. Each initial operation state feature and the dimension weight corresponding to each initial operation state feature are input into a linear transformation function shown in the following formula to obtain an aligned operation state feature :

[0089] (i∈1,2...M)

[0090] wherein b i is a bias of the linear transformation function.

[0091] Step 206, determining an attention weight between each of the first aligned operation state features, and updating each of the first aligned operation state features according to each of the attention weights to obtain a second aligned operation state feature;

[0092] Exemplarily, the similarity between each pair of the first alignment running state features can be determined by a standard self-attention calculation, normalized by a normalization function into attention weights, and each first alignment running state feature can be updated by weighted summation of the attention weights to obtain a second alignment running state feature; or the first alignment running state features can be projected into multiple subspaces to perform self-attention in parallel to obtain attention weights, and each first alignment running state feature can be updated by the attention weights to obtain a second alignment running state feature.

[0093] In some embodiments, the determining of the attention weights between each pair of the first alignment running state features and the updating of each first alignment running state feature according to the attention weights comprise:

[0094] determining the feature similarity between each pair of the first alignment running state features, and determining a similarity sum of all the feature similarities;

[0095] determining the ratio of each feature similarity to the similarity sum to obtain the attention weights between each pair of the first alignment running state features;

[0096] traversing each first alignment running state feature;

[0097] In each round of the traversal, obtaining other alignment running state features except the first alignment running state feature of the current round, and target attention weights associated with the other alignment running state features, and performing weighted fusion on the other alignment running state features by using the target attention weights to update the first alignment running state feature of the current round.

[0098] In a specific implementation, there are M first alignment running state features, the i-th first alignment running state feature is , the j-th first alignment running state feature is , and The attention weight of can be represented as:

[0099]

[0100] wherein, represents the feature similarity between and , i.e., the feature similarity between each pair of the first alignment running state features, and the exp() function is used to amplify the difference between each pair of the first alignment running state features, represents the similarity sum of all the feature similarities.

[0101] ​The first alignment running state feature of each iteration is traversed, and the first alignment running state feature of the current iteration is The other alignment running state features are , The target attention weight associated with is The attention weight of is The other alignment running state features are fused by weighted summation using target attention according to the following formula, and the result after fusion is used to update the first alignment running state feature of the current iteration to obtain the second alignment running state feature :

[0102]

[0103] Step 208: The second alignment running state features belonging to the same data source are fused to obtain a target running state feature.

[0104] Exemplarily, the target running state feature can be obtained by weighted summation after calculating the attention or similarity weight of the second alignment running state features belonging to the same data source based on a weighted average aggregation algorithm; or the target running state feature can be obtained by extracting a low-dimensional shared representation through decomposition by regarding the second alignment running state features belonging to the same data source as a high-order tensor based on tensor decomposition or factor decomposition fusion.

[0105] In some embodiments, the fusion of the second alignment running state features belonging to the same data source to obtain a target running state feature comprises:

[0106] Determining the information quantity metric value of each second alignment running state feature belonging to the same data source;

[0107] Determining the sum of all information quantity metric values belonging to the same data source to obtain an information quantity sum;

[0108] Determining the ratio of each information quantity metric value to the corresponding information quantity sum to obtain a second information quantity weight of each second alignment running state feature belonging to the same data source;

[0109] Fusing all second alignment running state features belonging to the same data source according to the corresponding second information quantity weight to obtain the target running state feature.

[0110] The information quantity metric value is used to determine the certainty of the information contained in each second alignment running state feature belonging to the same data source.

[0111] In a specific implementation, each second alignment running state feature belonging to the same data source is , and the information metric value thereof is represented as: (F i ), the second information weight of each second alignment running state feature belonging to the same data source can be determined by the following formula :

[0112]

[0113] wherein the exp() function is used to perform nonlinear amplification and positive numberization on (F i ), and represents the total information sum.

[0114] Further, all second alignment running state features belonging to the same data source are fused according to the corresponding second information weights by the following formula to obtain the target running state feature :

[0115]

[0116] Step 210, determining the initial fault detection result corresponding to each running state type of the transformer according to each target running state feature;

[0117] In a specific implementation, according to each target running state feature, the transformer can be preliminarily fault detected by a David triangle method, a three-ratio method, a PRPD, a PRPS, a time-frequency analysis, an ultrasonic positioning, a trend monitoring and baseline comparison, a deep neural network, a vector machine, a random forest algorithm, etc., to obtain the initial fault detection result corresponding to the running state type of the transformer, such as the fault detection result representing the normal, overheat and discharge of the transformer.

[0118] Step 212, fusing the initial fault detection results corresponding to each running state type according to the probability distribution of each initial fault detection result to obtain target fusion data.

[0119] In a specific implementation, based on the D-S evidence theory algorithm, each initial fault detection result can be converted into an evidence body, each evidence body representing the probability distribution of the initial fault detection result, and then the initial fault detection results are fused based on the Dempster combination rule to obtain the target fusion data; the target fusion data can include multiple single faults, composite faults and confidence quality.

[0120] In some embodiments, the fusing of the initial fault detection results corresponding to each running state type according to the probability distribution of each initial fault detection result to obtain the target fusion data comprises:

[0121] convert each of the initial fault detection results into a corresponding evidence body; wherein the evidence body is used to represent a probability distribution of the initial fault detection result;

[0122] determine an evidence body similarity between each pair of the evidence bodies, and determine a credibility weight of each of the evidence bodies according to each of the evidence body similarities;

[0123] determine a first information quantity weight of each of the evidence bodies; wherein the first information quantity weight is used to represent an uncertainty of the evidence body;

[0124] fuse the credibility weight and the first information quantity weight of each of the evidence bodies to obtain a fusion weight of each of the evidence bodies;

[0125] fuse each of the evidence bodies according to each of the fusion weights to obtain the target fusion data.

[0126] In a specific implementation, each of the initial fault detection results can be converted into a corresponding evidence body m i :

[0127]

[0128] wherein, is a recognition framework, represents a possible combination of each fault in each of the initial fault detection results, A is a focal element, represents a subset of . For example, = {insulation aging, winding fault, partial discharge (PD), core fault, overload, normal}, A is a subset of .

[0129] Further, a distance, a cosine similarity, etc. between each pair of the evidence bodies can be determined, and then normalized to obtain an evidence body similarity sim(m i ,m j ), and then the average similarity of each evidence body to other evidence bodies is converted into a credibility weight of each evidence body; or the difference between each pair of the evidence bodies can be measured based on information divergence, and the reciprocal or negative correlation of the difference is mapped as the evidence body similarity, and then aggregated to obtain the credibility weight of each evidence body;

[0130] In a specific implementation, a belief entropy can be used to represent the first information quantity weight of each evidence body, and the belief entropy is calculated as shown in the following formula:

[0131]

[0132] wherein, A and B are focal elements, A, B , | is the potential of the focal element;

[0133] Further, the belief entropy is normalized to obtain the first information quantity weight

[0134] Further, the fusion weights of each evidence body are obtained by fusing

[0135] and

[0136]

[0137] Further, the corresponding fusion weights are added to each evidence body to obtain the weighted evidence body

[0138] Finally, the Dempster combination rule is used to fuse

[0139] N-1 times to obtain the target fusion data:

[0140] wherein A, B, C are focal elements, i.e. subsets of

[0141] .

[0142] In some embodiments, the determination of the evidence body similarity between each two of all the evidence bodies, and the determination of the credibility weight of each evidence body according to the evidence body similarity, comprises:

[0143] determining the triangular divergence between each two of all the evidence bodies, and determining the similarity matrix of each evidence body according to the triangular divergence; wherein the similarity matrix is used to represent the evidence body similarity;

[0144] normalizing each similarity matrix to obtain the credibility weight of each evidence body.

[0145] In a specific implementation, the triangular divergence between the ith evidence body and the mth evidence body is denoted as STD(mi, mj), which can be determined by the following formula:

[0146]

[0147] wherein represents the probability mass of the evidence body mi assigned to the focal element A k , represents the probability mass of the evidence body m​​​​j the probability mass assigned to the focus cell A k .

[0148] Further, a similarity matrix is constructed and normalized to obtain a confidence weight :

[0149]

[0150] At step 214, a target fault detection result of the transformer is output according to the target fusion data.

[0151] Exemplarily, the target fusion data can include a plurality of single faults, composite faults and confidence qualities, and can be sorted according to the confidence qualities, and the top K (such as K=3) and the confidence values thereof are output as the target fault detection result, such as outputting “insulation fault (0.7); insulation or winding fault (0.2); overload (0.1)”; or the single fault or the composite fault with the highest confidence quality can be output as the target fault detection result.

[0152] In some embodiments, as shown in Figure 3 , a transformer fault detection model based on multi-modal data fusion is also provided, which comprises:

[0153] a multi-modal data feature extraction layer, configured to perform feature extraction on the operating state sensing data of the transformer in multiple modes according to a plurality of preset feature extraction algorithms, such as time sequence features and frequency domain features, to obtain initial operating state features;

[0154] a feature alignment and multi-modal analysis layer, configured to align the initial operating state features to a target dimension to obtain first aligned operating state features, and then determine attention weights between all pairs of the first aligned operating state features, update each first aligned operating state feature according to the attention weights to obtain second aligned operating state features; and then perform multi-modal analysis on the second aligned operating state features, fuse the second aligned operating state features belonging to the same data source to obtain target operating state features;

[0155] an information fusion decision layer, configured to determine initial fault detection results of the transformer corresponding to each operating state type according to the target operating state features, convert each initial fault detection result into an evidence body, calculate a confidence weight and an information weight (information amount weight) of each evidence body respectively, fuse the confidence weight and the information weight to obtain a fusion weight, and fuse each evidence body according to the fusion weight to obtain target fusion data, and output a target fault detection result of the transformer, i.e., a fault diagnosis analysis result of the transformer, according to the target fusion data.

[0156] In some embodiments, as shown in Figure 4As shown, a hierarchical multi-modal learning edge intelligent processing framework for embedded devices is also provided, comprising:

[0157] The first layer: single-modal feature parallel extraction, using a convolutional neural network to extract local features from vibration, ultra-high frequency, current waveform signals, respectively; using a small fully connected network to process gas, temperature scalar data; each kind of data is extracted in parallel, and the overhead is dispersed.

[0158] The second layer: single-modal feature fusion and diagnosis, splicing the feature vectors extracted from each kind of signal, inputting a light gradient boosting machine (LightGBM) or a micro Transformer model or the above transformer fault detection model or other models for fusion diagnosis, and outputting target fusion data (such as normal, overheating, discharge), fault location and confidence.

[0159] The third layer: fusion decision, using the conclusions obtained by the second layer single-modal data diagnosis to obtain the final target fault detection result.

[0160] In this embodiment, the framework breaks through the computing power limit of the edge device, realizes high-precision real-time diagnosis, and divides the complex multi-modal learning task into parallel single-modal feature extraction and centralized cross-modal decision-making.

[0161] The embodiments of the application have the following advantages: by aligning each initial operating state feature to the target dimension, the initial operating state features are preliminarily aligned in the data dimension to obtain first aligned operating state features, the difference problem of modal data of different dimensions is solved, the attention weights between all the first aligned operating state features are determined, each first aligned operating state feature is updated according to the attention weight to obtain a second aligned operating state feature, the expression ability of different modal data features to key state information can be enhanced through the second aligned operating state feature, and through data fusion of the second aligned operating state features of the same data source and cross-modal decision fusion based on probability distribution, multi-level information integration from local to global is realized, and then a target fault diagnosis result with high confidence and strong interpretability is output, the limitations of single modal data are effectively overcome, and the reliability of transformer fault detection in complex operating environments is improved.

[0162] It should be understood that although each step in the flowchart involved in the above-described embodiments is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in the above-described embodiments can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or steps or stages in other steps. It can be understood that the steps in different embodiments can be freely combined as needed, and various non-contradictory schemes formed by the combination are within the scope of protection of the present application.

[0163] Based on the same inventive concept, the embodiments of the present application also provide a transformer fault detection device for implementing the above-mentioned transformer fault detection method based on multi-modal data fusion. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more transformer fault detection device embodiments provided below can refer to the limitations of the transformer fault detection method based on multi-modal data fusion in the above text, which will not be repeated here.

[0164] In one exemplary embodiment, as shown in Figure 5 A transformer fault detection device 50 is provided, comprising:

[0165] The acquisition module 501 is configured to acquire operating state sensing data of multiple modalities of different data sources of a transformer, and extract initial operating state features for the operating state sensing data of each modality. Each modality has a corresponding operating state type.

[0166] The first alignment module 502 is configured to align each of the initial operating state features to a target dimension to obtain first aligned operating state features.

[0167] The second alignment module 503 is configured to determine attention weights between each of the first aligned operating state features, update each of the first aligned operating state features according to the attention weights, and obtain second aligned operating state features.

[0168] The first fusion module 504 is configured to fuse the second aligned operating state features belonging to the same data source to obtain target operating state features.

[0169] The first fault detection module 505 is configured to determine an initial fault detection result corresponding to each operation state type of the transformer according to each target operation state feature.

[0170] The second fusion module 506 is configured to fuse the initial fault detection results corresponding to each operation state type according to probability distribution of each initial fault detection result to obtain target fusion data.

[0171] The second fault detection module 507 is configured to output a target fault detection result of the transformer according to the target fusion data.

[0172] In one of the embodiments, the determining of the attention weight between each pair of the first aligned operation state features comprises:

[0173] determining a feature similarity between each pair of the first aligned operation state features, and determining a similarity sum of all the feature similarities;

[0174] determining a ratio of each feature similarity to the similarity sum to obtain the attention weight between each pair of the first aligned operation state features;

[0175] traversing each first aligned operation state feature;

[0176] In each round of the traversal, obtaining other aligned operation state features except the first aligned operation state feature of the current round, and a target attention weight associated with the other aligned operation state features, and performing weighted fusion on the other aligned operation state features by using the target attention weight to update the first aligned operation state feature of the current round.

[0177] In one of the embodiments, the fusing of the second aligned operation state features belonging to the same data source comprises:

[0178] determining an information quantity measurement value of each second aligned operation state feature belonging to the same data source;

[0179] determining a sum of all the information quantity measurement values belonging to the same data source to obtain an information quantity sum;

[0180] determining a ratio of each information quantity measurement value to the corresponding information quantity sum to obtain a second information quantity weight of each second aligned operation state feature belonging to the same data source;

[0181] Fusing all the second aligned running state features belonging to the same data source according to corresponding second information amount weights to obtain the target running state feature.

[0182] In one of the embodiments, the fusing of the initial fault detection results corresponding to each running state type according to the probability distribution of each initial fault detection result to obtain target fusion data comprises:

[0183] Converting each initial fault detection result into a corresponding evidence body; wherein the evidence body is used to represent the probability distribution of the initial fault detection result;

[0184] Determining the evidence body similarity between each pair of the evidence bodies, and determining the credibility weight of each evidence body according to each evidence body similarity;

[0185] Determining the first information amount weight of each evidence body; wherein the first information amount weight is used to represent the uncertainty of the evidence body;

[0186] Fusing the credibility weight and the first information amount weight of each evidence body to obtain the fusion weight of each evidence body;

[0187] Fusing each evidence body according to each fusion weight to obtain the target fusion data.

[0188] In one of the embodiments, the determining of the evidence body similarity between each pair of the evidence bodies, and the determining of the credibility weight of each evidence body according to each evidence body similarity comprises:

[0189] Determining the triangular divergence between each pair of the evidence bodies, and determining the similarity matrix of each evidence body according to each triangular divergence; wherein the similarity matrix is used to represent the evidence body similarity;

[0190] Normalizing each similarity matrix to obtain the credibility weight of each evidence body.

[0191] In one of the embodiments, the aligning of each initial running state feature to a target dimension to obtain an aligned running state feature comprises:

[0192] Determining the data dimension of each initial running state feature, and determining the dimension weight of each initial running state feature according to each data dimension and the target dimension;

[0193] Inputting each initial running state feature and the dimension weight corresponding to each initial running state feature into a preset linear transformation function to obtain the aligned running state feature.

[0194] The various modules in the transformer fault detection device 50 described above can be implemented in whole or in part by software, hardware, and combinations thereof. The various modules described above can be embedded in the processor in the computer device in hardware form or independent of the processor in the computer device, or can be stored in the memory in the computer device in software form so as to be called and executed by the processor to perform the operations corresponding to the various modules.

[0195] In an exemplary embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 6 The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store relevant data including, but not limited to, running state sensor data. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to communicate with terminals outside through a network connection. The computer program is executed by the processor to implement a transformer fault detection method based on multi-modal data fusion.

[0196] Those skilled in the art can understand that Figure 6 The structure shown in the above

[0197] In an exemplary embodiment, a computer device is provided, which includes a memory and a processor, and the memory stores a computer program. The processor executes the computer program to implement the transformer fault detection method based on multi-modal data fusion as described above.

[0198] In an embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the transformer fault detection method based on multi-modal data fusion as described above.

[0199] In an embodiment, a computer program product is provided, which includes a computer program. The computer program is executed by a processor to implement the transformer fault detection method based on multi-modal data fusion as described above.

[0200] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0201] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments of each method. In the embodiments provided in the present application, any reference to memory, database or other medium can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (Resistive Random Access Memory, ReRAM), magnetoresistive random access memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (Artificial Intelligence, AI) processor, etc., without being limited thereto.

[0202] Any technical features in the above embodiments can be combined, and for the sake of brevity, not all possible combinations are described above, however, any combination of these technical features is deemed to be within the scope of the present application.

[0203] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be pointed out that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A transformer fault detection method based on multimodal data fusion, characterized in that, The method includes: The transformer's operating status sensing data from different data sources is acquired for multiple modes, and initial operating status features are extracted from the operating status sensing data for each mode; wherein, each mode has a corresponding operating status type. Align each of the initial running state features to the target dimension to obtain the first aligned running state feature; Determine the attention weights between all pairs of the first alignment running state features, and update each of the first alignment running state features according to the attention weights to obtain the second alignment running state features; The second aligned running state features belonging to the same data source are fused to obtain the target running state features; Determine the initial fault detection result of the transformer for each of the target operating state types based on the characteristics of each target operating state; Based on the probability allocation of each initial fault detection result, the initial fault detection results corresponding to each operating state type are fused to obtain the target fused data; The target fault detection result of the transformer is output based on the target fusion data.

2. The method according to claim 1, characterized in that, The step of determining the attention weights between all pairs of the first alignment running state features, and updating each of the first alignment running state features according to each attention weight, includes: Determine the feature similarity between each pair of all the first aligned running state features, and determine the sum of the similarities of all the features; Determine the ratio of each feature similarity to the sum of the similarities to obtain the attention weights between all pairs of the first aligned running state features; Iterate through each of the first aligned running state features; In each round of the traversal, other alignment running state features besides the first alignment running state feature of the current round are obtained, as well as the target attention weight associated with the other alignment running state features. The target attention weight is used to perform weighted fusion on the other alignment running state features to update the first alignment running state feature of the current round.

3. The method according to claim 2, characterized in that, The step of fusing the second aligned running state features belonging to the same data source to obtain the target running state features includes: Determine the information content metric for each second aligned running state feature belonging to the same data source; The sum of all the information quantity measures belonging to the same data source is determined to obtain the total information quantity. Determine the ratio of each information quantity metric value belonging to the same data source to the corresponding sum of information quantities, and obtain the second information quantity weight of each second aligned running state feature belonging to the same data source; All second-aligned running state features belonging to the same data source are fused according to their corresponding second information content weights to obtain the target running state feature.

4. The method according to claim 1, characterized in that, The step of fusing the initial fault detection results corresponding to each of the aforementioned operating state types based on the probability allocation of each initial fault detection result to obtain target fused data includes: Each of the initial fault detection results is converted into a corresponding evidence body; wherein the evidence body is used to characterize the probability allocation of the initial fault detection results; Determine the evidence similarity between each pair of all the evidence pieces, and determine the credibility weight of each piece of evidence based on the evidence similarity. A first information weight is determined for each of the aforementioned pieces of evidence; wherein the first information weight is used to characterize the uncertainty of the aforementioned pieces of evidence. The credibility weight and the first information weight of each piece of evidence are fused together to obtain the fused weight of each piece of evidence. The evidence bodies are fused according to the fusion weights to obtain the target fused data.

5. The method according to claim 4, characterized in that, The step of determining the evidence similarity between all pairs of evidence, and determining the credibility weight of each piece of evidence based on the evidence similarity, includes: Determine the triangular divergence between each pair of all the evidence pieces, and determine the similarity matrix for each piece of evidence based on each triangular divergence; wherein, the similarity matrix is ​​used to characterize the similarity of the evidence pieces; Normalize each of the similarity matrices to obtain the credibility weight of each piece of evidence.

6. The method according to any one of claims 1 to 5, characterized in that, Aligning each of the initial running state features to the target dimension to obtain aligned running state features includes: Determine the data dimensions of each of the initial operating state features, and determine the dimension weights of each of the initial operating state features based on each of the data dimensions and the target dimension; The initial running state features and their corresponding dimension weights are input into a preset linear transformation function to obtain the aligned running state features.

7. A transformer fault detection device based on multimodal data fusion, characterized in that, The device includes: The acquisition module is used to acquire operating status sensing data of the transformer from different data sources for multiple modes, and to extract initial operating status features from the operating status sensing data of each mode; wherein, each mode has a corresponding operating status type; The first alignment module is used to align each of the initial running state features to the target dimension to obtain the first aligned running state features; The second alignment module is used to determine the attention weights between all pairs of the first alignment running state features, and update each of the first alignment running state features according to each attention weight to obtain the second alignment running state features. The first fusion module is used to fuse the second aligned running state features belonging to the same data source to obtain the target running state features; The first fault detection module is used to determine the initial fault detection result of the transformer and each type of operating state based on the characteristics of each target operating state. The second fusion module is used to fuse the initial fault detection results corresponding to each of the operating state types according to the probability allocation of each initial fault detection result to obtain target fused data; The second fault detection module is used to output the target fault detection result of the transformer based on the target fusion data.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images