Transformer fault detection method and device, medium and equipment

Abstract

The invention discloses a transformer fault detection method and device, a medium and equipment, and relates to the technical field of computers. Historical voiceprint data of a transformer are firstly acquired as sample data, fault types of the sample data are marked, then composite multi-scale dispersion entropies of the sample data are extracted, the composite multi-scale dispersion entropies are input to an encoder of a Transformer model to obtain predicted fault types, and the predicted fault types are optimized by taking minimization of deviation between a real fault type and the predicted fault types as an optimization target. And training an encoder of the Transform model to obtain a fault detection model, and finally extracting a composite multi-scale dispersion entropy corresponding to voiceprint data of a to-be-detected transformer, and inputting the composite multi-scale dispersion entropy into the fault detection model to perform fault feature extraction so as to determine the fault type of the to-be-detected transformer. According to the method, the time domain features of the voiceprint data of the transformer are reflected by adopting the composite multi-scale dispersion entropy, so that the fault features can be extracted from the voiceprint data more comprehensively, and the fault detection accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and particularly to a transformer fault detection method, device, medium and equipment. Background Art

[0002] Generally, the transformer fault acoustic signal is a non-linear time series. Traditional time-domain feature extraction methods, such as mean, root mean square, kurtosis and skewness, and frequency-domain feature extraction methods such as power spectrum, envelope spectrum analysis and fast Fourier transform, are difficult to analyze such non-linear signals.

[0003] In the prior art, the transformer fault acoustic detection method is essentially a multi-classification problem. Generally, a multi-classification machine learning algorithm is required to learn the internal association rules between acoustic features and fault labels from the training data, and then the effect of the multi-classification machine learning algorithm learning the internal rules is tested through the test data. Finally, the multi-classifier is capable of making an intelligent diagnosis of the transformer fault acoustic signal. From the existing technical paths, one type is to use static data-driven supervised classical machine learning algorithms, such as decision trees, ensemble learning, support vector machines, etc. Since the transformer fault acoustic signal is a non-linear time series information, the above methods destroy the potential rules contained in the time series, resulting in only learning the data association rules of the previous layer and ignoring the semantic information of the sequence context. Another type of method discovers the above problems, so a recurrent neural network (RNN) is used to learn the time series features of the acoustic signal. However, RNN is restricted by its own algorithm structure and has a poor learning effect on long sequence data and short-term memory. And RNN is prone to gradient disappearance and gradient explosion during the training process.

[0004] In summary, the existing transformer fault detection methods have poor analysis and utilization of the acoustic signals of transformers and low fault detection accuracy. Summary of the Invention

[0005] Based on this, it is necessary to provide a transformer fault detection method, device, medium and equipment for the above technical problems.

[0006] This specification adopts the following technical solutions:

[0007] This specification provides a transformer fault detection method, including:

[0008] Obtain the historical acoustic data of the transformer as sample data, and label the true fault type corresponding to the sample data;

[0009] Slide and decompose the sample data into multiple time-domain scales according to the time series, and calculate the dispersion entropy at each scale to obtain the composite multi-scale dispersion entropy corresponding to the sample data;

[0010] Input the composite multi-scale dispersion entropy corresponding to the sample data into the encoder of the Transformer model to obtain the predicted fault type; and train the encoder of the Transformer model with the optimization goal of minimizing the deviation between the true fault type and the predicted fault type to obtain a fault detection model;

[0011] Extract the composite multi-scale dispersion entropy of the voiceprint data of the transformer to be detected, and input it into the fault detection model. The fault detection model extracts the fault features of the composite multi-scale dispersion entropy of the voiceprint data of the transformer to be detected, and determines the fault type of the transformer to be detected according to the fault features.

[0012] Optionally, the step of sliding and decomposing the sample data into multiple time domain scales according to the time series and calculating the dispersion entropy at each scale to obtain the composite multi-scale dispersion entropy corresponding to the sample data specifically includes:

[0013] Filter the sample data to eliminate background noise, and normalize the filtered sample data to obtain preprocessed sample data;

[0014] Generate multiple coarse-grained sequences at multiple different time domain scales according to the following formula based on the preprocessed sample data:

[0015]

[0016] For each different time domain scale, determine the dispersion entropy of each coarse-grained sequence at this time domain scale, and determine the mean value of the dispersion entropy of each coarse-grained sequence;

[0017] Take the mean value of the dispersion entropy of each coarse-grained sequence at different time domain scales as the composite multi-scale dispersion entropy corresponding to the sample data;

[0018] Wherein, is the j-th element in the k-th coarse-grained sequence corresponding to the time domain scale τ, x is the preprocessed sample data sequence, and τ max is the preset maximum time domain scale, and N is the number of elements in the preprocessed sample data sequence.

[0019] Optionally, the step of training the encoder of the Transformer model with the optimization goal of minimizing the deviation between the true fault type and the predicted fault type specifically includes:

[0020] Determine the deviation loss between the true fault type and the predicted fault type according to the cross-entropy loss function;

[0021] Train the encoder of the Transformer model with the optimization goal of minimizing the total deviation loss of each sample data.

[0022] Optionally, the true fault types include one or more of winding deformation, core looseness, accessory looseness, harmonic current, DC bias magnetization, fan aging, core looseness, heavy load, abnormal fan noise, and partial discharge.

[0023] This specification provides a transformer fault detection device, including:

[0024] An acquisition module, configured to acquire historical acoustic fingerprint data of a transformer as sample data, and label the true fault type corresponding to the sample data;

[0025] An extraction module, configured to slide and decompose the sample data into multiple time-domain scales according to a time series, and calculate the dispersion entropy at each scale to obtain the composite multi-scale dispersion entropy corresponding to the sample data;

[0026] A training module, configured to input the composite multi-scale dispersion entropy corresponding to the sample data into the encoder of the Transformer model to obtain a predicted fault type; and taking minimizing the deviation between the true fault type and the predicted fault type as an optimization objective, train the encoder of the Transformer model to obtain a fault detection model;

[0027] A detection module, configured to extract the composite multi-scale dispersion entropy of the acoustic fingerprint data of the transformer to be detected, input it into the fault detection model, extract the fault features of the composite multi-scale dispersion entropy of the acoustic fingerprint data of the transformer to be detected through the fault detection model, and determine the fault type of the transformer to be detected according to the fault features.

[0028] This specification provides a computer-readable storage medium, where the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned transformer fault detection method is implemented.

[0029] This specification provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above-mentioned transformer fault detection method is implemented.

[0030] At least one of the above technical solutions adopted in this specification can achieve the following beneficial effects:

[0031] First, obtain the historical acoustic fingerprint data of the transformer as sample data, and label the fault types of each sample data. Then, extract the composite multi-scale dispersion entropy of each sample data and input it into the encoder of the Transformer model to obtain the predicted fault type. Taking the minimization of the deviation between the true fault type and the predicted fault type as the optimization goal, train the encoder of the Transformer model to obtain a fault detection model. Finally, for the acoustic fingerprint data of the transformer to be detected, extract its corresponding composite multi-scale dispersion entropy and input it into the fault detection model for fault feature extraction to determine the fault type of the transformer to be detected. The present invention uses the composite multi-scale dispersion entropy to reflect the time-domain characteristics of the acoustic fingerprint data of the transformer, and through coarse-graining and time shift, fuses the characteristics such as multi-time-domain scales and amplitude relationships of the acoustic fingerprint data of the transformer, which is convenient to more comprehensively extract fault features from it and improves the accuracy of fault detection. Description of the Drawings

[0032] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation to the present application. In the drawings:

[0033] Figure 1 It is a schematic flow chart of a transformer fault detection method provided by this specification;

[0034] Figure 2 It is a schematic flow chart of a transformer fault classification provided by this specification;

[0035] Figure 3 It is a schematic diagram of a Transformer model architecture provided by this specification;

[0036] Figure 4 It is a schematic diagram of a Transformer multi-classifier model architecture provided by this specification;

[0037] Figure 5 It is a schematic diagram of a transformer fault detection device provided by this specification;

[0038] Figure 6 It is a schematic diagram of a computer device for implementing the transformer fault detection method provided by this specification. Detailed Embodiments

[0039] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of them. Based on the embodiments in the specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this application.

[0040] Currently, traditional time- and frequency-domain feature extraction methods often cannot effectively extract hidden fault features, resulting in a reduced fault recognition rate. Due to different system responses caused by fault shocks, the degree of chaos in the monitoring data varies. Analysis methods based on entropy theory can directly measure the complexity of data without data decomposition or transformation and can complete the recognition of different faults. Sample entropy, permutation entropy, dispersion entropy, etc. have been successively proposed and used for fault diagnosis. Traditional entropy methods can all be used as means to measure the complexity of time series. When performing feature extraction, single-scale entropy methods can only analyze signals at a single time scale and ignore useful information on other time scales of the time series.

[0041] In the selection of classifiers, most still use algorithms of the recurrent neural network family, such as long short-term memory networks (LSTM) and gated recurrent units (GRU), to replace the traditional RNN model. Since LSTM and GRU still use the recurrent network structure, there are still problems of gradient vanishing and gradient explosion. On the other hand, LSTM and GRU lack an attention mechanism, making it impossible to globally focus on the input sequence and resulting in no substantial improvement in the memory of the algorithm.

[0042] Sample entropy is time-consuming for long data calculations and is susceptible to mutation signals. Although permutation entropy has a relatively fast calculation speed, the relationship between signal amplitudes is not considered in the calculation process. Dispersion entropy overcomes some of the defects of the above two, considers the relationship between amplitudes, and has advantages such as fast calculation speed and less influence by mutation signals. The dispersion entropy can be combined with the multi-scale entropy theory to obtain multi-scale dispersion entropy, which is a computer-based nonlinear time series analysis method. It decomposes the time series into multiple time domain scales, calculates the dispersion entropy at each scale, and then combines them to obtain the complexity characteristics of the entire time series. However, the original time series decreases as the coarse-graining scale factor increases, and the change in the length of the time series will cause a mutation in the entropy value and lead to a large deviation in the entropy value calculation. This problem can be solved by the composite multi-scale dispersion entropy that combines time-shift operations on the basis of the coarse-graining process. Therefore, the present invention uses composite multi-scale dispersion entropy for feature extraction of voiceprint data.

[0043] Meanwhile, the present invention will adopt the Transformer model as the classifier, which has many advantages over LSTM and GRU:

[0044] Global view: When calculating the encoder and decoder, the Transformer model can consider all positions in the input sequence simultaneously, without being restricted by distance, and can better process global information.

[0045] Context-free encoding: The position encoding in the Transformer model provides position information for each position in the input sequence, which helps the model distinguish words or tokens at different positions and maintains position sensitivity.

[0046] Modeling of long-range dependencies: Due to the self-attention mechanism, the Transformer can effectively capture long-range dependencies in the input sequence, which is particularly important for processing long sequences and avoids the problems of gradient vanishing or explosion in traditional recurrent neural networks and convolutional neural networks when dealing with long sequences.

[0047] Parallel computing ability: The Transformer model adopts the self-attention mechanism, allowing the model to concurrently focus on information at different positions when processing sequence data, which significantly improves the computing efficiency and training and inference speed.

[0048] Flexibility and adaptability: The structure of the Transformer is more flexible and can adapt to different task and dataset requirements by increasing or decreasing the number of layers, adjusting the number of attention heads, etc.

[0049] The following will, in conjunction with the accompanying drawings, elaborate on the technical solutions provided by each embodiment of the present application.

[0050] Figure 1 It is a schematic flowchart of a transformer fault detection method in this specification, specifically including the following steps:

[0051] S101: Obtain the historical acoustic fingerprint data of the transformer as sample data and label the true fault types corresponding to each sample data.

[0052] Generally, transformer mechanical defect faults are generally caused by overcurrent and fault current due to short circuits or surges, which will generate stress between wires, leads, and coils, resulting in faults such as winding deformation, core loosening, and accessory loosening. Factors such as design defects, insulation aging, and overloading will also cause winding deformation. The above problems are generally detected using acoustic fingerprint technology, and their fault types are generally divided into harmonic current, DC bias, fan aging, core loosening, heavy load, abnormal fan noise, partial discharge, etc.

[0053] Based on this, in one or more embodiments of the present invention, the server of the service platform may first obtain the historical acoustic fingerprint data of the transformer of the type to be detected during the historical operation process as sample data. The historical acoustic fingerprint data may include the acoustic fingerprint data under various types of faults described above, and may also include the acoustic fingerprint data under fault-free conditions.

[0054] Furthermore, the server may also pre-determine the true fault types corresponding to each sample data and use them as labels for application in model training.

[0055] The server mentioned in this specification may be a server set up in the service platform, or a device such as a desktop computer or a laptop computer that can execute the solution of this specification. For the convenience of description, only the server is used as the execution subject for description below.

[0056] S102: Slide and decompose each sample data into multiple time domain scales according to the time series, and calculate the dispersion entropy at each scale to obtain the composite multi-scale dispersion entropy corresponding to each sample data.

[0057] After obtaining the sample data through the above method, further, the server may first perform time domain feature extraction on each sample data, and use the extracted time domain features as the model input to train the model.

[0058] The present invention mainly uses the composite multi-scale dispersion entropy to comprehensively extract the features of the acoustic fingerprint data, and then inputs the obtained feature sequence into the Transformer model (only the encoder) for fault multi-classification processing, as Figure 2 shown, Figure 2 which is a schematic diagram of a transformer fault classification process in this specification.

[0059] Among them, the multi-scale dispersion entropy is a computer-based non-linear time series analysis method. It decomposes the time series into multiple time domain scales, calculates the dispersion entropy at each scale, and then combines them to obtain the complexity characteristics of the entire time series. However, the original time series decreases as the coarse-graining scale factor increases, and the change in the length of the time series will cause a sudden change in the entropy value, and will also lead to a large deviation in the entropy value calculation. This problem can be solved by combining the time shift operation on the basis of the coarse-graining process. This method directly measures the complexity of the data without decomposing or transforming the data. Thus, it overcomes the disadvantage that the traditional time and frequency domain feature extraction methods often cannot effectively extract the hidden fault features, and extracts more comprehensive acoustic fingerprint features.

[0060] Specifically, in one or more embodiments of the present invention, the server may first filter the sample data to eliminate background noise, perform normalization processing on the filtered sample data to obtain the preprocessed sample data, and then generate multiple coarse-grained sequences at multiple different time domain scales according to the preprocessed sample data through the following formula:

[0061]

[0062] wherein, is the j-th element in the k-th coarse-grained sequence corresponding to the time-domain scale τ, x is the preprocessed sample data sequence, and τ max is the preset maximum time-domain scale, and N is the number of elements in the preprocessed sample data sequence.

[0063] After that, for each different time-domain scale, determine the dispersion entropy of each coarse-grained sequence at that time-domain scale, and determine the mean value of the dispersion entropy of each coarse-grained sequence. Finally, use the mean value of the dispersion entropy of each coarse-grained sequence at different time-domain scales as the composite multi-scale dispersion entropy corresponding to the sample data.

[0064] For example, noise reduction processing can be performed first: filter the voiceprint signal (such as wavelet denoising, band-pass filtering) to eliminate background noise. Then perform normalization: standardize the voiceprint signal to zero mean and unit variance to avoid the influence of amplitude differences on the analysis.

[0065] Then perform multi-scale coarse-graining, with the goal of generating multiple coarse-grained sequences at different time scales (τ) to capture the multi-level complexity of the signal.

[0066] Select the scale range: set the maximum scale factor τ_max = 10.

[0067] Generate coarse-grained sequences: for each scale τ, generate τ coarse-grained sequences.

[0068] Each sequence is segmented and averaged through different starting points (k = 1, 2,..., τ):

[0069]

[0070] After that, calculate the dispersion entropy (DE), and set parameters: embedding dimension, number of classes, time delay.

[0071] Symbolic mapping: use the cumulative distribution function to map the values of the coarse-grained sequences to c discrete classes.

[0072] Construct the dispersion pattern: generate an m-dimensional pattern vector (for example, the symbol sequence [3, 5, 2] corresponds to the embedding dimension m = 3).

[0073] Statistical pattern probability: calculate the occurrence frequency of each pattern in the sequence to obtain the probability distribution P.

[0074] Calculate the entropy value:

[0075]

[0076] Finally, the Composite Multiscale Dispersion Entropy (CMDE) is determined. For each scale τ, the mean value of DE of its τ coarse-grained sequences is calculated:

[0077]

[0078] The final feature is a vector composed of CMDE values corresponding to each scale τ (τ = 1, 2,..., τ_max). CMDE synthesizes the entropy values of coarse-grained sequences with different starting points at different time domain scales, improving the feature stability.

[0079] S103: Input the Composite Multiscale Dispersion Entropy corresponding to each sample data into the encoder of the Transformer model to obtain the predicted fault type; and take minimizing the deviation between the true fault type and the predicted fault type as the optimization goal to train the encoder of the Transformer model to obtain a fault detection model.

[0080] S104: Extract the Composite Multiscale Dispersion Entropy of the voiceprint data of the transformer to be detected and input it into the fault detection model. Extract the fault features of the Composite Multiscale Dispersion Entropy of the voiceprint data of the transformer to be detected through the fault detection model, and determine the fault type of the transformer to be detected according to the fault features.

[0081] After extracting the Composite Multiscale Dispersion Entropy corresponding to each sample data, the server can input it into the encoder of the Transformer model to train the encoder of the Transformer model.

[0082] Transformer is a neural network architecture based on the self-attention mechanism, aiming to solve the sequence-to-sequence (seq2seq) problems in natural language processing. This model has achieved great success in machine translation tasks. It adopts a classic encoder-decoder structure, including an encoder and a decoder. In the encoder part, each word is processed through a self-attention mechanism, which allows the model to process all positions in the input sequence in parallel during the calculation process, improving the computational efficiency. Such a mechanism helps capture long-range dependencies in the input sequence. The core ideas of Transformer include the self-attention mechanism, positional encoding, multi-head attention, and feed-forward neural networks. The self-attention mechanism allows the model to consider all positions in the input sequence simultaneously instead of processing them step by step. Positional encoding is to preserve the position information of the elements in the input sequence because the model itself has no explicit order information. The multi-head attention mechanism captures different relationships through multiple attention heads, while the feed-forward neural network is used to learn non-linear relationships. The encoder of the Transformer model is stacked by six identical layers, and each layer contains two sub-layers: the first sub-layer is the multi-head self-attention mechanism layer, and the second sub-layer is a fully connected feed-forward neural network. These two sub-layers are connected through a residual network structure and followed by a layer normalization layer. The decoder structure is similar and also consists of six layers. In addition to the two sub-layers, there is an additional sub-layer for implementing the multi-head attention mechanism on the output of the encoder. The masked self-attention sub-layer in the decoder is specially designed to prevent the leakage of information from future positions, such as Figure 3 shown Figure 3 is a schematic diagram of a Transformer model architecture in this specification.

[0083] Essentially, the encoder is used for the efficient extraction of feature relationships, and the decoder uses the key feature relationships extracted by the encoder for sequence generation. Since this problem is not a sequence generation problem but a classification problem, the present invention uses a Transformer model (only the encoder) to strengthen the fault features extracted by the composite multi-scale dispersion entropy in the temporal semantic analysis. The attention focuses on the information that is more critical to the current task, reduces the attention to other information, and even filters out irrelevant information, thereby improving the efficiency and accuracy of task processing. In the encoder part, each word is processed through a self-attention mechanism, which allows the model to process all positions in the input sequence in parallel during the calculation process, improving the calculation efficiency. The self-attention mechanism creates query, key, and value vectors for each feature data, enabling the model to focus on all other features in the input sequence. Such a mechanism helps to capture the long-range dependencies in the input sequence. Before passing it to the Softmax function to achieve fault classification, the Transformer multi-classifier used in the present invention is as Figure 4 shown Figure 4 a schematic diagram of a Transformer multi-classifier model architecture in this specification.

[0084] In one or more embodiments of the present invention, the cross-entropy loss function can be used to determine the loss. That is, the server can determine the deviation loss between the true fault type and the predicted fault type according to the cross-entropy loss function, and take minimizing the total deviation loss of each sample data as the optimization goal to train the encoder of the Transformer model.

[0085] For example, if the fault type is a discrete category y ∈ {1, 2,..., K}, the cross-entropy loss is as follows:

[0086]

[0087] where N is the number of samples, K is the total number of fault types, y i,k ∈ {0, 1} is the true fault type (One-Hot encoding) of the i-th sample data, and p i,k (θ) is the probability that the i-th sample data belongs to the fault type k (output by Softmax), and θ is the Transformer model parameter.

[0088] Based on Figure 1The transformer fault detection method shown first obtains the historical acoustic fingerprint data of the transformer as sample data, and labels the fault types of each sample data. Then, the composite multi-scale dispersion entropy of each sample data is extracted and input into the encoder of the Transformer model to obtain the predicted fault type. Taking the minimization of the deviation between the true fault type and the predicted fault type as the optimization goal, the encoder of the Transformer model is trained to obtain a fault detection model. Finally, for the acoustic fingerprint data of the transformer to be detected, its corresponding composite multi-scale dispersion entropy is extracted and input into the fault detection model for fault feature extraction to determine the fault type of the transformer to be detected. The present invention uses the composite multi-scale dispersion entropy to reflect the time-domain characteristics of the acoustic fingerprint data of the transformer, and fuses the characteristics of the acoustic fingerprint data of the transformer in multiple time-domain scales and amplitude relationships through coarse-graining and time shift, which is convenient for more comprehensively extracting fault features from it and improving the accuracy of fault detection.

[0089] The present invention uses the composite multi-scale dispersion entropy to extract features from the acoustic fingerprint data without decomposing or transforming the data, and directly measures the complexity of the data. Thus, it overcomes the disadvantage that the traditional time-domain and frequency-domain feature extraction methods often cannot effectively extract hidden fault features.

[0090] In addition, the encoder of the Transformer model is used for efficient extraction of feature relationships, and the decoder uses the key feature relationships extracted by the encoder for sequence generation. Since this problem is not a sequence generation problem but a classification problem. Therefore, the present invention uses the Transformer model (only the encoder) to strengthen the fault features extracted by the composite multi-scale dispersion entropy in the time-series semantic analysis, focusing the attention on the information that is more critical to the current task, reducing the attention to other information, and even filtering out irrelevant information, thereby improving the efficiency and accuracy of task processing. Extract the correlation relationship between the key features and the entire sequence globally, and then transfer it to the Softmax function to achieve fault classification. This makes the overall model have the great advantage of being able to process long-sequence acoustic fingerprint data and achieve accurate classification.

[0091] When applying the transformer fault detection method provided in this specification, it is not necessary to execute according to Figure 1 the order of the steps shown. The specific execution order of each step can be determined as needed, and this specification does not limit this.

[0092] The above is the transformer fault detection method provided by one or more embodiments of this specification. Based on the same idea, this specification also provides a corresponding transformer fault detection device, as Figure 5 shown.

[0093] Figure 5 It is a schematic diagram of a transformer fault detection device provided by this specification, including:

[0094] An acquisition module 201, configured to acquire historical acoustic fingerprint data of a transformer as sample data, and label the true fault type corresponding to the sample data;

[0095] An extraction module 202, configured to slide and decompose the sample data into multiple time domain scales according to a time series, and calculate the dispersion entropy at each scale to obtain the composite multi-scale dispersion entropy corresponding to the sample data;

[0096] A training module 203, configured to input the composite multi-scale dispersion entropy corresponding to the sample data into the encoder of the Transformer model to obtain a predicted fault type; and taking minimizing the deviation between the true fault type and the predicted fault type as an optimization objective, training the encoder of the Transformer model to obtain a fault detection model;

[0097] A detection module 204, configured to extract the composite multi-scale dispersion entropy of the acoustic fingerprint data of the transformer to be detected, input it into the fault detection model, extract the fault features of the composite multi-scale dispersion entropy of the acoustic fingerprint data of the transformer to be detected through the fault detection model, and determine the fault type of the transformer to be detected according to the fault features.

[0098] For the specific limitations of the transformer fault detection device, reference can be made to the limitations on the transformer fault detection method in the above text, which will not be elaborated here. Each module in the above transformer fault detection device can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0099] This specification also provides a computer-readable storage medium, which stores a computer program, and the computer program can be used to execute the above Figure 1 provided transformer fault detection method.

[0100] This specification also provides Figure 6 a schematic structural diagram of the computer device shown in Figure 6 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, there may also be other hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above Figure 1 provided transformer fault detection method.

[0101] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant 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 embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0102] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

Claims

1. A transformer fault detection method, characterized in that: include: Obtain the transformer's historical voiceprint data as sample data, and mark the actual fault type corresponding to the sample data; Decompose the sample data into multiple time domain scales according to the time series sliding, and calculate the scatter entropy at each scale to obtain the composite multi-scale scatter entropy corresponding to the sample data; The composite multi-scale scattered entropy corresponding to the sample data is input into the encoder of the Transformer model to obtain the predicted fault type; and the encoder of the Transformer model is trained with minimizing the deviation between the real fault type and the predicted fault type as the optimization goal to obtain the fault detection model; The composite multi-scale scatter entropy of the voiceprint data of the transformer to be detected is extracted and input into the fault detection model. The fault characteristics of the composite multi-scale scatter entropy of the voiceprint data of the transformer to be detected are extracted through the fault detection model, and the fault type of the transformer to be detected is determined according to the fault characteristics.

2. The method according to claim 1, characterized in that The sample data is decomposed into multiple time domain scales according to the time series sliding, and the scatter entropy at each scale is calculated to obtain the composite multi-scale scatter entropy corresponding to the sample data, specifically including: Filtering the sample data to eliminate background noise, and normalizing the filtered sample data to obtain preprocessed sample data; Generate multiple coarse-grained sequences at different time domain scales based on the preprocessed sample data: For each different time domain scale, determine the spread entropy of each coarse-grained sequence at the time domain scale, and determine the mean value of the spread entropy of each coarse-grained sequence; The mean of the scatter entropy of each coarse-grained sequence at different time domain scales is used as the composite multi-scale scatter entropy corresponding to the sample data; in, is the jth element in the kth coarse-grained sequence corresponding to the time domain scale τ, x is the preprocessed sample data sequence, τ max is the preset maximum time domain scale, and N is the number of elements in the preprocessed sample data sequence.

3. The method according to claim 1, characterized in that The encoder of the Transformer model is trained with minimizing the deviation between the real fault type and the predicted fault type as the optimization goal, specifically including: Determine the deviation loss between the true fault type and the predicted fault type according to the cross entropy loss function; The encoder of the Transformer model is trained with the optimization goal of minimizing the total deviation loss of each sample data.

4. The method according to claim 1, characterized in that The real fault types include: one or more of: winding deformation, loose core, loose accessories, harmonic current, DC bias magnetism, fan aging, loose core, overload, abnormal fan noise and partial discharge.

5. A transformer fault detection device, characterized in that: include: An acquisition module is used to acquire the historical voiceprint data of the transformer as sample data and mark the real fault type corresponding to the sample data; The extraction module is used to decompose the sample data into multiple time domain scales according to the time series sliding, and calculate the scatter entropy at each scale to obtain the composite multi-scale scatter entropy corresponding to the sample data; The training module is used to input the composite multi-scale scattered entropy corresponding to the sample data into the encoder of the Transformer model to obtain the predicted fault type; and the encoder of the Transformer model is trained with minimizing the deviation between the real fault type and the predicted fault type as the optimization goal to obtain the fault detection model; The detection module is used to extract the composite multi-scale scatter entropy of the voiceprint data of the transformer to be detected, and input it into the fault detection model, extract the fault characteristics of the composite multi-scale scatter entropy of the voiceprint data of the transformer to be detected through the fault detection model, and determine the fault type of the transformer to be detected according to the fault characteristics.

6. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

7. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method described in any one of claims 1 to 4 is implemented.

Images