Transformer fault detection method and device based on automatic encoder and multi-scale feature fusion

Abstract

The invention discloses a transformer fault detection method and device based on an automatic encoder and multi-scale feature fusion, and belongs to the technical field of power transformer fault detection.The method comprises the steps that sound signals during operation of a transformer are collected, and data frame division and normalization preprocessing are conducted on the collected sound signals; reconstructing the data of the collected sound signals by using an automatic encoder trained by normal sound data, and extracting the data coded by the automatic encoder as the implicit features of the sound signals; according to the error before and after reconstruction, sound abnormity judgment is carried out, and fault type detection is further carried out on abnormal data by using a classifier; and the classifier receives the collected sound data and the characteristic signal extracted by the automatic encoder as input, and classifies fault types. According to the method, the classifier is arranged, fault features are mined through the multi-scale features of the transformer sound signals, and the accuracy of transformer fault classification is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power transformer fault detection, and in particular relates to a transformer fault detection method and device based on automatic encoder and multi-scale feature fusion. Background Art

[0002] Power transformers are one of the most critical pieces of equipment in power systems, and fault detection technology for these transformers is crucial for their safe and stable operation. Currently, the mainstream methods for transformer fault diagnosis include oil chromatography, vibration diagnosis, infrared thermal imaging, acoustic diagnosis, and spectroscopy. Transformer acoustic signals, however, have become a research hotspot in recent years due to their ease of acquisition, rich information about equipment operation, and contactless detection.

[0003] Traditional sound signal analysis methods primarily rely on time-frequency domain analysis based on manual experience. However, the information contained in sound signals is extensive and complex, and the conclusions drawn from this method cannot accurately determine the fault type. In recent years, deep learning has been widely used in the field of fault detection. Deep learning neural network classifiers, such as convolutional neural networks and long short-term memory networks, are used to identify fault types in fault detection tasks. However, since fault sounds rarely occur during transformer operation, the model is mostly used to detect normal sounds. Using a complex deep learning neural network classifier with a single detection target wastes computing resources. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] A transformer fault detection method based on autoencoder and multi-scale feature fusion, comprising:

[0006] Step 1: collecting sound signals when the transformer is running, and performing data frame division and normalization preprocessing on the collected sound signals;

[0007] Step 2: reconstruct the collected sound signal data using an autoencoder that has been trained with normal sound data, and extract the data encoded by the autoencoder as the implicit features of the sound signal;

[0008] Step 3: Determine if the sound is abnormal based on the errors before and after reconstruction. The data that is determined to be abnormal is further classified using a classifier to detect the type of fault.

[0009] In step 4, the classifier receives the collected sound data and the feature signal extracted by the autoencoder as input and classifies the fault type.

[0010] A transformer fault detection device based on autoencoder and multi-scale feature fusion, comprising:

[0011] The data acquisition and preprocessing module collects the sound signals during transformer operation, and performs data frame division and normalization preprocessing on the collected sound signals;

[0012] The data reconstruction and latent feature acquisition module uses an autoencoder that has been trained with normal sound data to reconstruct the collected sound signal data, and extracts the data encoded by the autoencoder as the latent features of the sound signal;

[0013] The judgment module determines whether the sound is abnormal based on the errors before and after reconstruction. The data judged as abnormal is further detected by the classifier to determine the type of fault;

[0014] Fault classification module,the classifier receives the collected sound data and the feature signals extracted by the,autoencoder as input and classifies the fault types.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the transformer fault detection method based on the fusion of an autoencoder and multi-scale features are implemented.

[0016] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the transformer fault detection method based on the fusion of an autoencoder and multi-scale features.

[0017] The present invention has the following beneficial effects:

[0018] The present invention uses an autoencoder to detect whether the transformer sound signal is normal. This eliminates the need to run a classifier when detecting normal sound, thus saving computing resources. When the autoencoder detects that the transformer sound signal is abnormal, the implicit features of the sound signal extracted by the autoencoder are used as input to the classifier along with the transformer sound signal to determine the transformer fault category. By configuring the classifier to exploit the multi-scale features of the transformer sound signal to identify fault characteristics, the present invention improves the accuracy of transformer fault classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the transformer fault detection method based on autoencoder and multi-scale feature fusion proposed in the present invention;

[0020] Figure 2 is a structural diagram of an automatic encoder according to an embodiment of the present invention;

[0021] Figure 3 It is a structural diagram of a classifier model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0023] This paper proposes a transformer fault detection method based on an autoencoder and multi-scale feature fusion. This method can effectively process the collected transformer sound data and detect whether the transformer sound is abnormal and the fault type when an abnormality occurs. The specific steps include:

[0024] Step 1: collecting sound signals when the transformer is running, and performing data frame division and normalization preprocessing on the collected sound signals;

[0025] Step 2: reconstruct the collected sound signal data using an autoencoder that has been trained with normal sound data, and extract the data encoded by the autoencoder as the implicit features of the sound signal;

[0026] Step 3: Determine if the sound is abnormal based on the errors before and after reconstruction. The data that is determined to be abnormal is further classified using a classifier to detect the type of fault.

[0027] In step 4, the classifier receives the collected sound data and the feature signal extracted by the autoencoder as input and classifies the fault type.

[0028] Wherein, step 1 includes:

[0029] Step 1.1, transformer signal collection: Install a microphone near the transformer and use the microphone to collect the sound signal generated during the operation of the transformer :

[0030] ;

[0031] in, is the signal amplitude of the first sampling point, corresponding to the instantaneous value of the sound signal at the start of acquisition, is the signal amplitude at the second sampling point, corresponding to the instantaneous value after the first sampling interval; For the The signal amplitude of the sampling point, is the number of sampling points.

[0032] Step 1.2, sound signal preprocessing: the collected sound signal is processed as M continuous data as a data frame unit and normalized to :

[0033] ;

[0034] in, Respectively 、 、 The signal amplitude of the sampling point, Indicates the division of the sound signal into data frames, which are It is the maximum value function, which is used to get the maximum value of the elements in the sequence.

[0035] Wherein, step 2 includes:

[0036] Step 2.1, autoencoder sound signal reconstruction; first use the collected normal sound signal of the transformer to train the autoencoder neural network. The autoencoder selects 2h+1 layers, where h is a positive integer and the input data is , the output data is , the training process uses Constraints, the loss function uses the mean square error function:

[0037] ;

[0038] in, The first values, Output reconstructed data values;

[0039] The collected sound data Input into the trained autoencoder to get the reconstructed data .

[0040] Step 2.2, extract the data of the hth hidden layer of the autoencoder As implicit features of sound signals.

[0041] Wherein, step 3 includes:

[0042] Step 3.1: After the autoencoder is trained with normal sound, the mean μ and variance σ of the mean square error function before and after the reconstruction of the normal sound signal are calculated during the last several rounds of training. The threshold is set to Q=μ+3. σ, calculation and The mean square error of .

[0043] Step 3.2, if and If the mean square error is greater than the threshold Q, it is determined to be an abnormal signal. and It is fed into the classifier as input; otherwise it is judged as a normal signal, and the automatic encoder is used to classify the next frame of sound data. to be processed.

[0044] Wherein, step 4 includes:

[0045] In step 4.1, a multi-column convolutional neural network (MCNN-LSTM) model is selected as the classifier: a one-dimensional CNN feature extraction module with three channels is constructed, and the information from different channels is weightedly fused using the attention mechanism. The weighted fused feature vector is sent as input to the LSTM network for classification and recognition. The output of the LSTM network is sent to the fully connected layer to output the fault category. The training set and test set are input to the classifier model for training, the model parameters are updated, and the model parameters with the best results are saved as the parameters of the fault detection model.

[0046] Step 4.2, Sound Signal First, downsampling is performed to reduce the sampling rate, and then it is simultaneously input into two one-dimensional CNNs to extract the features of low-frequency signals and high-frequency signals respectively;

[0047] Feature data The features are input into the remaining one-dimensional CNN to extract the feature data;

[0048] The features extracted by the three one-dimensional CNNs are weighted and fused through the attention mechanism, and the obtained feature vector is used as the input vector of the LSTM network;

[0049] The LSTM network consists of layered LSTMs. Each LSTM unit includes three gates: input, forget, and output, and a memory unit. Each gate and memory unit has its own weight and bias. The fully connected layer uses the softmax function to convert neuron output into a probability distribution of fault types.

[0050] Finally, according to the probability distribution, the fault type with the maximum probability is selected as the fault type identification result.

[0051] The technical solution of the present invention is described in detail below with reference to specific embodiments.

[0052] like Figure 1 As shown, the transformer fault detection method based on the fusion of automatic encoder and multi-scale features of the present invention includes the following steps:

[0053] Step 1: Use a microphone to collect sound signals during transformer operation, and perform data frame division and normalization preprocessing on the collected sound signals; specifically, the following steps are included:

[0054] Step 1.1, transformer signal collection: Install a microphone near the transformer. In this embodiment, the microphone is used to collect the sound signal generated during the operation of the transformer. , the sampling rate is 40kHz, expressed as:

[0055] ;

[0056] in, is the signal amplitude of the first sampling point, corresponding to the instantaneous value of the sound signal at the start of acquisition, is the signal amplitude at the second sampling point, corresponding to the instantaneous value after the first sampling interval; For the The signal amplitude of the sampling point, is the number of sampling points.

[0057] Step 1.2, sound signal preprocessing: In this embodiment, the collected sound signal is processed as a data frame unit with M (in this embodiment, M=2000) continuous data, and normalized to , expressed as:

[0058] ;

[0059] in, Respectively 、 、 The signal amplitude of the sampling point, Indicates the division of the sound signal into data frames, which are It is the maximum value function, which is used to get the maximum value of the elements in the sequence.

[0060] Step 2: reconstruct the collected sound signal data using an autoencoder that has been trained with normal sound data, and extract the data encoded by the autoencoder as the implicit features of the sound signal; specifically, the following steps are performed:

[0061] Step 2.1, autoencoder sound signal reconstruction: First, the autoencoder neural network is trained using the collected normal sound signal of the transformer. In this embodiment, the autoencoder selects 2h+1 layers (h is a positive integer, and a 7-layer structure is selected in this embodiment. Usually, too many layers are not selected. The more layers, the slower the running speed and the lower the efficiency). Figure 2As shown, the dimensions of the input layer and output layer are both M (M=2000), and the dimensions of the five hidden layers are 1024, 512, 256, 512, and 1024 respectively. , the output data is , the training process uses Constraints, the loss function uses the mean square error function:

[0062] ;

[0063] in, The first values, Output reconstructed data values;

[0064] The collected sound data Input into the trained autoencoder to get the reconstructed data ;

[0065] Step 2.2, extract the data of the third hidden layer of the autoencoder As implicit features of sound signals.

[0066] Step 3: Determine if the sound is abnormal based on the errors before and after reconstruction. A classifier is then used to detect the type of fault for the data that is considered abnormal. This specifically includes:

[0067] Step 3.1: After the autoencoder is trained with normal sound, the mean μ and variance σ of the mean square error function before and after the reconstruction of the normal sound signal are calculated during the last 50 rounds of training. The threshold is set to Q=μ+3. σ, calculation and The mean square error of

[0068] Step 3.2, if and If the mean square error is greater than the threshold Q, it is determined to be an abnormal signal. and It is fed into the classifier as input; otherwise it is judged as a normal signal, and the automatic encoder is used to classify the next frame of sound data. to be processed.

[0069] Step 4: The classifier receives the original sound signal (collected sound data) Feature signals extracted with autoencoder As input, the fault type is classified; specifically:

[0070] Step 4.1: In this embodiment, a multi-column convolutional neural network (MCNN)-long short-term memory network (LSTM) model is selected as the classifier. The classifier model is as follows: Figure 3 As shown in the figure, a one-dimensional convolutional neural network (CNN) feature extraction module with three channels is constructed. The attention mechanism is used to perform weighted fusion of information from different channels. The weighted fusion feature vector is sent as input to the LSTM network for classification and recognition. The output of the LSTM network is sent to the fully connected layer (softmax) to output the fault category. The training set and test set are input to the classifier model for training, the model parameters are updated, and the model parameters with the best results are saved as the parameters of the fault detection model.

[0071] The fault sound uses the sound data in the transformer fault sound dataset. There are nine types of fault sounds, namely heavy overload, loose clamps, DC bias magnetization, short-circuit impact, surface discharge, corona discharge, internal discharge, suspended potential discharge, and abnormal cooler sound.

[0072] 200 seconds of sound data for each type were selected, for a total of 1800 seconds of sound, with a sampling rate of 40kHz; 70% of the sound data was divided into a training set for MCNN-LSTM network model training; 30% was used as a test set to verify the MCNN-LSTM network's ability to identify and classify transformer faults; a learning rate of 0.0001 was designed to train the MCNN-LSTM network model. After training, the MCNN-LSTM network was used as a fault detection model to detect the sound data determined to be abnormal in step 3.2.

[0073] In the three one-dimensional CNNs: CNN_1, CNN_2, and CNN_3, each CNN hidden layer includes a convolution layer and a pooling layer. The convolution layer uses filters with shared weights to extract features from the input signal, and one-dimensional convolution is selected. The pooling layer selects and filters the extracted features, and maximum pooling is selected.

[0074] Step 4.2, Sound Signal First, the sampling rate is downsampled by 8 times, and then it is input into two one-dimensional CNNs (CNN_1 and CNN_2) at the same time. CNN_1 uses 30 30 convolution kernels to automatically extract the features of low-frequency signals, while CNN_2 uses 10 A convolution kernel of 10 is used to extract the features of high-frequency signals.

[0075] Feature data It is input into one of the one-dimensional CNN: CNN_3 to extract features, using 20 20 convolution kernels are used to extract the features of the feature data.

[0076] The three one-dimensional CNNs all use the stochastic gradient descent method to update the network parameters; the features extracted by CNN_1, CNN_2 and CNN_3 are weightedly fused through the attention mechanism, and the obtained feature vector is used as the input vector of the LSTM network.

[0077] The LSTM network consists of layered LSTMs. Each LSTM unit includes three gates: input, forget, and output, and a memory unit. Each gate and memory unit has its own weight and bias. The fully connected layer uses the softmax function (normalized exponential function) to convert the neuron output into the probability distribution of nine fault types. The softmax function is as follows:

[0078] .

[0079] in, Indicates the The output of a neuron, The current sample belongs to The predicted probability of the class, is the softmax function, To sum the sequence values in the function, For the Finally, according to the probability distribution, the fault type with the highest probability is selected as the fault type recognition result.

[0080] The present invention is not only applicable to the detection method of transformer sound signals, but can also be extended to various fault data such as transformer shell vibration signals and transformer infrared temperature measurement.

[0081] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention may be implemented using various computer languages.

[0082] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0083] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0085] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0086] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

[0087] The above descriptions are merely embodiments of the present invention and are not intended to limit the scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied to other related system fields, are also included in the scope of protection of the present invention.

[0088] The contents not described in detail in the specification of the present invention belong to the prior art known to those skilled in the art.

Claims

1. A transformer fault detection method based on autoencoder and multi-scale feature fusion, characterized in that: include: Step 1: collecting sound signals when the transformer is running, and performing data frame division and normalization preprocessing on the collected sound signals; Step 2: reconstruct the collected sound signal data using an autoencoder that has been trained with normal sound data, and extract the data encoded by the autoencoder as the implicit features of the sound signal; Step 3: Determine if the sound is abnormal based on the errors before and after reconstruction. The data that is determined to be abnormal is further classified using a classifier to detect the type of fault. In step 4, the classifier receives the collected sound data and the feature signal extracted by the autoencoder as input and classifies the fault type.

2. The transformer fault detection method based on autoencoder and multi-scale feature fusion according to claim 1 is characterized in that: Step 1 includes: Step 1.1, transformer signal collection: Install a microphone near the transformer and use the microphone to collect the sound signal generated during the operation of the transformer : ； in, is the signal amplitude of the first sampling point, corresponding to the instantaneous value of the sound signal at the start of acquisition, is the signal amplitude at the second sampling point, corresponding to the instantaneous value after the first sampling interval; For the The signal amplitude of the sampling point, is the number of sampling points; Step 1.2, sound signal preprocessing: the collected sound signal is processed as M continuous data as a data frame unit and normalized to : ； in, Respectively 、 、 The signal amplitude of the sampling point, Indicates the division of the sound signal into data frames, which are It is the maximum value function, which is used to get the maximum value of the elements in the sequence.

3. The transformer fault detection method based on autoencoder and multi-scale feature fusion according to claim 2 is characterized in that: Step 2 includes: Step 2.1, autoencoder sound signal reconstruction; first use the collected normal sound signal of the transformer to train the autoencoder neural network. The autoencoder selects 2h+1 layers, where h is a positive integer and the input data is , the output data is , the training process uses Constraints, the loss function uses the mean square error function: ； in, The first values, Output reconstructed data values; The collected sound data Input into the trained autoencoder to get the reconstructed data ; Step 2.2, extract the data of the hth hidden layer of the autoencoder As implicit features of sound signals.

4. The transformer fault detection method based on autoencoder and multi-scale feature fusion according to claim 3 is characterized in that: Step 3 includes: Step 3.1: After the autoencoder is trained with normal sound, the mean μ and variance σ of the mean square error function before and after the reconstruction of the normal sound signal are calculated during the last several rounds of training. The threshold is set to Q=μ+3. σ, calculation and The mean square error of Step 3.2, if and If the mean square error is greater than the threshold Q, it is determined to be an abnormal signal. and It is fed into the classifier as input; otherwise it is judged as a normal signal, and the automatic encoder is used to classify the next frame of sound data. to be processed.

5. The transformer fault detection method based on autoencoder and multi-scale feature fusion according to claim 4 is characterized in that: Step 4 includes: Step 4.1: Select the multi-column convolutional neural network (MCNN-LSTM) model as the classifier. Construct a one-dimensional CNN feature extraction module with three channels. Use the attention mechanism to weightedly fuse information from different channels. Send the weighted fused feature vector as input to the LSTM network for classification and recognition. Send the output of the LSTM network to the fully connected layer to output the fault category. Input the training set and test set to the classifier model for training, update the model parameters, and save the model parameters with the best results as the parameters of the fault detection model. Step 4.2, Sound Signal First, downsampling is performed to reduce the sampling rate, and then it is simultaneously input into two one-dimensional CNNs to extract the features of low-frequency signals and high-frequency signals respectively; Feature data The features are input into the remaining one-dimensional CNN to extract the feature data; The features extracted by the three one-dimensional CNNs are weighted and fused through the attention mechanism, and the obtained feature vector is used as the input vector of the LSTM network; The LSTM network consists of layered LSTMs. Each LSTM unit includes three gates: input, forget, and output, and a memory unit. Each gate and memory unit has its own weight and bias. The fully connected layer uses the softmax function to convert neuron output into a probability distribution of fault types. Finally, according to the probability distribution of the fault types, the fault type with the highest probability is selected as the fault type identification result.

6. The transformer fault detection method based on autoencoder and multi-scale feature fusion according to claim 5 is characterized in that: In step 4.1, the fault sound uses the sound data in the transformer fault sound dataset. There are nine types of fault sounds: heavy overload, loose clamps, DC bias magnetization, short circuit impact, surface discharge, corona discharge, internal discharge, floating potential discharge, and abnormal cooler sound. Each hidden layer of the three one-dimensional CNNs includes a convolution layer and a pooling layer. The convolution layer uses filters with shared weights to extract features from the input signal, and one-dimensional convolution is selected. The pooling layer selects and filters the extracted features, and maximum pooling is selected.

7. The transformer fault detection method based on autoencoder and multi-scale feature fusion according to claim 5 is characterized in that: In step 4.2, the softmax function is as follows: ； in, Indicates the The output of a neuron, The current sample belongs to The predicted probability of the class, is the softmax function, To sum the sequence values in the function, For the The output of a neuron.

8. A transformer fault detection device based on autoencoder and multi-scale feature fusion, characterized in that: include: The data acquisition and preprocessing module collects the sound signals during transformer operation, and performs data frame division and normalization preprocessing on the collected sound signals; The data reconstruction and latent feature acquisition module uses an autoencoder that has been trained with normal sound data to reconstruct the collected sound signal data, and extracts the data encoded by the autoencoder as the latent features of the sound signal; The judgment module determines whether the sound is abnormal based on the errors before and after reconstruction. The data judged as abnormal is further detected by the classifier to determine the type of fault; Fault classification module,the classifier receives the collected sound data and the feature signals extracted by the,autoencoder as input and classifies the fault types.

9. An electronic device comprising 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 steps of the transformer fault detection method based on autoencoder and multi-scale feature fusion according to any one of claims 1 to 7 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the transformer fault detection method based on autoencoder and multi-scale feature fusion as described in any one of claims 1 to 7 are implemented.

Images