Transformer Fault Diagnosis Technology Based on Knowledge Mapping

Abstract

The present invention discloses a transformer fault diagnosis technology based on knowledge mapping, which relates to the field of fault diagnosis technology and includes the following steps: S1, obtaining transformer multimodal data and performing data preprocessing; S2, establishing a memory-enhanced early warning model to predict transformer operation data; S3, transformer fault diagnosis. The present invention comprehensively considers online monitoring data or offline test data such as transformer oil dissolved gas data, oil temperature, partial discharge signal, core or clamp grounding current, DC current, infrared spectrum, winding deformation curve, ambient temperature and humidity, etc., performs fault diagnosis on the predicted multimodal data, and predicts possible future transformer faults.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis, and in particular to a transformer fault diagnosis technology based on knowledge mapping. Background Art

[0002] In recent years, time series prediction and transformer fault identification technology based on artificial neural networks have been widely used and achieved good results. Models suitable for long-sequence prediction and multimodal data fusion technology that can perform more comprehensive analysis have received widespread attention. Knowledge graphs have been valued in various fields and have become an important technical means in the power field. Transformer fault knowledge graphs contain rich and reliable transformer fault-related knowledge, including transformer fault types, fault causes, fault handling methods, etc. It also includes transformer online and offline monitoring data, experimental data and fault mapping relationships. It can more comprehensively and comprehensively judge transformer fault conditions based on transformer multimodal data and provide further analysis.

[0003] The Informer model is a Transformer-based model, which is divided into two parts: the encoding component and the decoding component, which are composed of an encoder and a decoder respectively. It avoids the problem that traditional neural networks such as long short-term memory networks have poor performance as the sequence becomes longer and it is difficult to converge. It improves the problems of the quadratic time complexity, high memory consumption, and the inherent limitations of the encoder-decoder structure of the Transformer. At the same time, the installation of the transformer online monitoring device provides a large amount of comprehensive real-time monitoring data of the transformer, such as dissolved gas in oil, oil temperature, partial discharge signals, etc., combined with multi-modal data such as field test data, historical data, equipment attribute data, etc., and can jointly analyze from multiple angles to diagnose transformer faults more comprehensively and accurately. Summary of the invention

[0004] The purpose of the present invention is to provide a transformer fault diagnosis technology based on knowledge mapping to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a transformer fault diagnosis technology based on knowledge mapping, comprising the following steps:

[0006] S1. Obtain transformer multimodal data and data preprocessing: Obtain online monitoring data or offline test data of the transformer during fault, extract features and preprocess the acquired data to obtain multimodal data, and normalize them as input variables of the Informer model;

[0007] S2. Establish a memory-enhanced early warning model to predict transformer operation data: positionally encode the input variables of the Informer model, and obtain the self-attention operation results through the self-attention mechanism calculation formula. According to the self-attention operation results, a focused self-attention feature map is constructed to finally obtain the final hidden representation of the encoder. According to the final hidden representation of the encoder, the most relevant items in the memory are retrieved and reconstructed. The reconstruction error is used as the judgment basis for fault diagnosis, and it is embedded in the Informer model for prediction to obtain multimodal data of transformers in the future.

[0008] S3. Transformer fault diagnosis: The different modal data features predicted and output by the Informer model are fused, and the fault type of the equipment at the future time node is obtained by building a fault warning model.

[0009] Preferably, in step S1, firstly, the online monitoring data or offline test data of the dissolved gas in the oil of the transformer during the fault, the oil temperature, the partial discharge signal, the core or clamp grounding current, the DC current, the infrared spectrum, the winding deformation curve, and the ambient temperature and humidity are obtained, and the geographical location of the substation is used as an external static covariate. At the same time, the time and space relationship are considered, and the acquired data are tested to obtain the infrared spectrum, the oil chromatogram and the high-frequency partial discharge detection spectrum.

[0010] Preferably, in step S1, the feature extraction and data preprocessing refer to extracting key information from infrared spectra, oil chromatograms and high-frequency partial discharge detection spectra through image feature extraction technology and converting them into data in numerical form, performing data cleaning, differential denoising and data anomaly processing on the data to obtain multimodal data, and normalizing the data to use it as an input variable of the Informer model.

[0011] Preferably, step S2 comprises the following steps:

[0012] S201, position encoding: Add a position encoding vector to the traditional Informer model input vector to represent the order of input data;

[0013] S202, self-attention mechanism module: multiply each input vector of the Informer model encoder by three trained matrices to obtain three matrices, namely, query matrix Query, key matrix Key and value matrix Value, wherein the query vector query ranked top u in the difference M between attention and uniform distribution is selected as the query matrix Query, and the selection process is to first randomly sample the key value key, do a dot product of each query vector query with respect to the sampled key value key, and then calculate the difference M of attention and uniform distribution of the dot product corresponding to each query vector query, and then select the query vector query with the largest difference top u as the processed query matrix Query to calculate the attention probability;

[0014] Since the probability distribution of the self-attention mechanism is sparse, only a few dot products contribute most of the attention. Therefore, the query vector query ranked top u in the difference M between the attention and the uniform distribution is selected as the query matrix Query.

[0015] Preferably, the step S2 further comprises the following steps:

[0016] S203, attention distillation mechanism: extract the better feature map from the result obtained after the self-attention mechanism operation and give it the dominant feature, construct the focused self-attention feature map in the next layer, greatly reduce the time dimension of the input, and reduce the number of self-attention distillation layers by establishing a master sequence copy with half the input. Finally, construct multiple feature maps with a length of 1 / 4 of the total length and splice them together to obtain the final hidden representation of the encoder, which can enhance the robustness of the attention distillation mechanism;

[0017] The result obtained after the self-attention mechanism operation in the encoder has redundant combinations of the value matrix Value. It is necessary to extract the better feature map and give it the dominant feature, and construct a focused self-attention feature map in the next layer to greatly reduce the time dimension of the input.

[0018] Preferably, the step S2 further comprises the following steps:

[0019] S204, knowledge memory enhancement module: According to the final hidden representation of the encoder, the most relevant items in the memory are retrieved through the attention-based addressing operator, and then these items are delivered to the decoder for reconstruction. The reconstruction is close to the normal sample of the device. The reconstruction error of the normal sample is small, while the reconstruction error of the abnormal device failure is large. Therefore, the reconstruction error is used as the judgment basis for fault diagnosis. The sparse representation model is used to introduce a sparse representation of the mixed norm constraint, so that the spatiotemporal similarity between the different modal data of the device is maximized in the transform domain.

[0020] Preferably, the step S2 further comprises the following steps:

[0021] S205, generative decoder: When entering the decoder, part of the true value and the value to be predicted are concatenated, and the values ​​to be predicted are all set to 0 first, so as to obtain the input sequence of the decoder, and directly output the prediction result in one step. The Informer model prediction is embedded through the knowledge memory enhancement module to obtain the transformer multimodal data for a period of time in the future.

[0022] Preferably, step S3 comprises the following steps:

[0023] S301, feature fusion representation: After the knowledge memory enhancement module is embedded in the prediction output of the Informer model, the final fusion feature representation is obtained by simply splicing the features of different modal data;

[0024] S302, constructing a fault warning model: adding a linear connection layer to map the feature dimension, and finally combining the manually annotated fault type labels, using the cross entropy loss function to train the entire network model, and finally obtaining a transformer fault warning model;

[0025] S303, predicting fault type: During the test, different modal data of the transformer are input, and the fault type of the equipment at a future time node is obtained through the fault warning model.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. The present invention comprehensively considers online monitoring data or offline test data such as the transformer's oil dissolved gas data, oil temperature, partial discharge signal, core or clamp grounding current, DC current, infrared spectrum, winding deformation curve, ambient temperature and humidity, etc., to perform fault diagnosis on the predicted multimodal data and predict possible future transformer faults;

[0028] 2. The present invention predicts transformer multimodal data by combining a knowledge memory enhancement module and an embedded Informer model. The sparse representation model of the mixed norm constraint fully considers the spatiotemporal correlation between modal data and is limited to using only a small number of related items for reconstruction. Therefore, the reconstruction supervision forces the memory to record the most representative prototype patterns, and performs a limited number of normal mode reconstructions in the knowledge memory module, which helps to produce larger reconstruction errors in abnormal situations, making subsequent fault anomaly diagnosis more advantageous. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of the overall method flow provided by an embodiment of the present invention;

[0030] Figure 2A schematic flow chart of a method for establishing a memory-enhanced early warning model to predict transformer operation data provided by an embodiment of the present invention;

[0031] Figure 3 A schematic diagram of a transformer fault diagnosis method flow chart provided by an embodiment of the present invention;

[0032] Figure 4 A schematic diagram of the Informer model framework provided by an embodiment of the present invention;

[0033] Figure 5 A schematic diagram of the calculation flow of the self-attention mechanism provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] See also Figures 1 to 5 The present invention provides a technical solution: a transformer fault diagnosis technology based on knowledge mapping, comprising the following steps:

[0036] S1. Obtain transformer multimodal data and data preprocessing: Obtain online monitoring data or offline test data of the transformer during fault, extract features and preprocess the acquired data to obtain multimodal data, and normalize them as input variables of the Informer model;

[0037] S2. Establish a memory-enhanced early warning model to predict transformer operation data: positionally encode the input variables of the Informer model, and obtain the self-attention operation results through the self-attention mechanism calculation formula. According to the self-attention operation results, a focused self-attention feature map is constructed to finally obtain the final hidden representation of the encoder. According to the final hidden representation of the encoder, the most relevant items in the memory are retrieved and reconstructed. The reconstruction error is used as the judgment basis for fault diagnosis, and it is embedded in the Informer model for prediction to obtain multimodal data of transformers in the future.

[0038] S3. Transformer fault diagnosis: The different modal data features predicted and output by the Informer model are fused, and the fault type of the equipment at the future time node is obtained by building a fault warning model.

[0039] In step S1, firstly, the online monitoring data or offline test data of the transformer oil dissolved gas data, oil temperature, partial discharge signal, core or clamp grounding current, DC current, infrared spectrum, winding deformation curve, ambient temperature and humidity are obtained when the transformer fails, and the geographical location of the substation is used as an external static covariate. At the same time, the time and space relationship is considered, and the acquired data is tested to obtain the infrared spectrum, oil chromatogram and high-frequency partial discharge detection spectrum;

[0040] In step S1, feature extraction and data preprocessing refer to extracting key information from infrared spectra, oil chromatograms and high-frequency partial discharge detection spectra through image feature extraction technology and converting them into numerical data, performing data cleaning, differential denoising and data anomaly processing on the data to obtain multimodal data, and normalizing the data to use it as an input variable of the Informer model;

[0041] Step S2 includes the following steps:

[0042] S201, position encoding: Add a position encoding vector to the traditional Informer model input vector to represent the order of input data;

[0043] The position encoding follows the formula:

[0044]

[0045] In the formula, pos represents the position, i represents the dimension, and d model represents the dimension of positional encoding embedding, sin represents sine function, cos represents cosine function, and PE represents positional encoding;

[0046] S202, self-attention mechanism module: multiply each input vector of the Informer model encoder by three trained matrices to obtain three matrices, namely, query matrix Query, key matrix Key and value matrix Value, wherein the query vector query ranked top u in the difference M between attention and uniform distribution is selected as the query matrix Query, and the selection process is to first randomly sample the key value key, do a dot product of each query vector query with respect to the sampled key value key, and then calculate the difference M of attention and uniform distribution of the dot product corresponding to each query vector query, and then select the query vector query with the largest difference top u as the processed query matrix Query to calculate the attention probability;

[0047] The calculation formula of the self-attention mechanism is:

[0048]

[0049] In the formula, softmax represents the function, Q, K, and V represent the query matrix Query, the key matrix Key, and the value matrix Value, respectively. k Represents the dimension of each column feature in the key matrix Key, K T Respectively represent the transposed matrices of the processed query matrix Query and key matrix Key;

[0050] S203, attention distillation mechanism: extract the better feature map from the result obtained after the self-attention mechanism operation and give it the dominant feature, construct the focused self-attention feature map in the next layer, greatly reduce the time dimension of the input, reduce the number of self-attention distillation layers by establishing a master sequence copy with half the input, and finally construct multiple feature maps with a length of 1 / 4 of the total length and splice them together to obtain the final hidden representation of the encoder;

[0051] S204, knowledge memory enhancement module: according to the final hidden representation of the encoder, the most relevant items in the memory are retrieved through the attention-based addressing operator, and then these items are delivered to the decoder for reconstruction. The reconstruction is close to the normal sample of the device. The reconstruction error of the normal sample is small, while the reconstruction error of the abnormal device fault is large. Therefore, the reconstruction error is used as the judgment basis for fault diagnosis. The sparse representation model is used to introduce the sparse representation of the mixed norm constraint, so that the spatiotemporal similarity between the different modal data of the device is maximized in the transform domain.

[0052] The sparse representation model is expressed as:

[0053]

[0054] W=||(X k ) T (D k )|| 12

[0055] In the formula, K is the number of spatial samples, X k is the set of data corresponding to the kth data in each spatial sampling data, D k is the corresponding dictionary, A k is the sparse representation corresponding to the kth data under this dictionary, ||·|| 12 represents the sum of all column elements of the matrix and then normalizes them by the bi-norm, W represents the intermediate variable, R represents the sparse representation model, diag represents the diagonal matrix, λ represents the constant value, Denotes the definition of F norm, (X k ) T Represents X k The transposed matrix of

[0056] S205, generative decoder: when entering the decoder, part of the real value and the value to be predicted are concatenated, and the values ​​to be predicted are all set to 0 first, so as to obtain the input sequence of the decoder, and directly output the prediction result in one step, and embed the Informer model prediction through the knowledge memory enhancement module to obtain the transformer multimodal data for a period of time in the future;

[0057] Step S3 includes the following steps:

[0058] S301, feature fusion representation: After the knowledge memory enhancement module is embedded in the prediction output of the Informer model, the final fusion feature representation is obtained by simply splicing the features of different modal data;

[0059] S302, constructing a fault warning model: adding a linear connection layer to map the feature dimension, and finally combining the manually annotated fault type labels, using the cross entropy loss function to train the entire network model, and finally obtaining a transformer fault warning model;

[0060] S303, predicting fault type: During the test, different modal data of the transformer are input, and the fault type of the equipment at a future time node is obtained through the fault warning model.

[0061] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0062] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Transformer fault diagnosis technology based on knowledge mapping, characterized in that: The method comprises the following steps: S1. Obtain transformer multimodal data and data preprocessing: Obtain online monitoring data or offline test data of the transformer during fault, extract features and preprocess the acquired data to obtain multimodal data, and normalize them as input variables of the Informer model; S2. Establish a memory-enhanced early warning model to predict transformer operation data: positionally encode the input variables of the Informer model, and obtain the self-attention operation results through the self-attention mechanism calculation formula. According to the self-attention operation results, a focused self-attention feature map is constructed to finally obtain the final hidden representation of the encoder. According to the final hidden representation of the encoder, the most relevant items in the memory are retrieved and reconstructed. The reconstruction error is used as the judgment basis for fault diagnosis, and it is embedded in the Informer model for prediction to obtain multimodal data of transformers in the future. S3. Transformer fault diagnosis: The different modal data features predicted and output by the Informer model are fused, and the fault type of the equipment at the future time node is obtained by building a fault warning model.

2. The transformer fault diagnosis technology based on knowledge mapping according to claim 1 is characterized in that: In the step S1, firstly, the online monitoring data or offline test data of the transformer oil dissolved gas data, oil temperature, partial discharge signal, core or clamp grounding current, DC current, infrared spectrum, winding deformation curve, ambient temperature and humidity are obtained when the transformer fails. The geographical location of the substation is used as an external static covariate, and the time and space relationship are considered at the same time. The acquired data is tested to obtain infrared spectrum, oil chromatogram and high-frequency partial discharge detection spectrum.

3. The transformer fault diagnosis technology based on knowledge mapping according to claim 2 is characterized in that: In step S1, the feature extraction and data preprocessing refer to extracting key information from infrared spectra, oil chromatograms and high-frequency partial discharge detection spectra through image feature extraction technology and converting them into numerical data, performing data cleaning, differential denoising and data anomaly processing on the data to obtain multimodal data, and normalizing the data to use it as an input variable of the Informer model.

4. The transformer fault diagnosis technology based on knowledge mapping according to claim 1 is characterized in that: The step S2 comprises the following steps: S201, position encoding: Add a position encoding vector to the traditional Informer model input vector to represent the order of input data; S202, self-attention mechanism module: multiply each input vector of the Informer model encoder by three trained matrices to obtain three matrices, namely, the query matrix Query, the key matrix Key and the value matrix Value, wherein the query vector query ranked in the top u in the difference M between attention and uniform distribution is selected as the query matrix Query. The selection process is to first randomly sample the key value key, dot product each query vector query with the sampled key value key, and then calculate the attention and uniform distribution difference M of the dot product corresponding to each query vector query, and then select the query vector query with the largest difference in the top u as the processed query matrix Query to calculate the attention probability.

5. The transformer fault diagnosis technology based on knowledge mapping according to claim 4 is characterized in that: The step S2 further comprises the following steps: S203, attention distillation mechanism: extract the better feature map from the result obtained after the self-attention mechanism operation and give it the dominant feature, construct a focused self-attention feature map in the next layer, greatly reduce the time dimension of the input, and reduce the number of self-attention distillation layers by establishing a main sequence copy with half the input. Finally, construct multiple feature maps with a length of 1 / 4 of the total length and splice them together to obtain the final hidden representation of the encoder.

6. The transformer fault diagnosis technology based on knowledge mapping according to claim 5 is characterized in that: The step S2 further comprises the following steps: S204, knowledge memory enhancement module: According to the final hidden representation of the encoder, the most relevant items in the memory are retrieved through the attention-based addressing operator, and then these items are delivered to the decoder for reconstruction. The reconstruction is close to the normal sample of the device. The reconstruction error of the normal sample is small, while the reconstruction error of the abnormal device failure is large. Therefore, the reconstruction error is used as the judgment basis for fault diagnosis. The sparse representation model is used to introduce a sparse representation of the mixed norm constraint, so that the spatiotemporal similarity between the different modal data of the device is maximized in the transform domain.

7. The transformer fault diagnosis technology based on knowledge mapping according to claim 6 is characterized in that: The step S2 further comprises the following steps: S205, generative decoder: When entering the decoder, part of the true value and the value to be predicted are concatenated, and the values ​​to be predicted are all set to 0 first, so as to obtain the input sequence of the decoder, and directly output the prediction result in one step. The Informer model prediction is embedded through the knowledge memory enhancement module to obtain the transformer multimodal data for a period of time in the future.

8. The transformer fault diagnosis technology based on knowledge mapping according to claim 7 is characterized in that: The step S3 comprises the following steps: S301, feature fusion representation: After the knowledge memory enhancement module is embedded in the prediction output of the Informer model, the final fusion feature representation is obtained by simply splicing the features of different modal data; S302, constructing a fault warning model: adding a linear connection layer to map the feature dimension, and finally combining the manually annotated fault type labels, using the cross entropy loss function to train the entire network model, and finally obtaining a transformer fault warning model; S303, predicting fault type: During the test, different modal data of the transformer are input, and the fault type of the equipment at a future time node is obtained through the fault warning model.

Images