Direct current transmission converter valve abnormal sound detection method and device based on deep learning

Abstract

The invention relates to a direct current transmission converter valve abnormal sound detection method and device based on deep learning, and the method comprises the following steps: collecting a normal acoustic signal and an abnormal sound signal of a direct current transmission converter valve, and obtaining a normal sample and an abnormal sound sample after data processing and sample division; dividing normal samples into a training set, a verification set and a test set in proportion, and dividing all abnormal sound samples into the test set; the verification set is input into a trained abnormal sound detection model, an abnormal sound detection threshold value is obtained through calculation, and the abnormal sound detection model is a convolution attention mechanism module-multi-scale convolution auto-encoder; and inputting a real-time acoustic signal of the direct-current power transmission converter valve into the trained abnormal sound detection model, automatically outputting an abnormal detection result in real time in combination with an abnormal sound detection threshold obtained by calculation, and respectively storing a normal acoustic signal and an abnormal sound signal. Compared with the prior art, the method has the advantages that the abnormal sound of the converter valve can be automatically detected in real time, the detection accuracy is high, and stability is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of high-voltage direct current transmission, and in particular to a method and device for detecting abnormal noise of a direct current transmission converter valve based on deep learning. Background Art

[0002] Ultra-high voltage direct current transmission technology is widely used because of its advantages in long-distance transmission, such as low transmission line cost, fast regulation speed and no operational stability problems. The converter valve is the key core equipment of high voltage direct current transmission, which is used to realize the AC-DC conversion of electric energy. However, in actual operation, it will be affected by abnormal operating conditions such as overload and overvoltage and harsh electromagnetic environment. These abnormal operating conditions and natural aging of components can easily lead to the occurrence of converter valve device failure.

[0003] The existing online monitoring systems and manual inspection methods have blind spots, which makes it difficult to detect some converter valve faults in time. The acoustic signals generated by the converter valve during operation contain a large amount of status information. When a fault occurs, the converter valve makes an abnormal sound, and the occurrence of the fault can be captured in time through its abnormal sound data, but the abnormal sound of the converter valve is difficult to capture in time and accurately through the human ear. The abnormal sound detection method based on deep learning only needs to provide a normal data training model without manual participation, and can automatically realize the above abnormal sound detection process. Based on the above problems, it is urgent to provide a new abnormal sound detection method for DC transmission converter valves based on deep learning.

[0004] After searching, Chinese invention patent publication number CN114997210A discloses a method for identifying and detecting abnormal machine noises based on deep learning, which simplifies the process of identifying abnormal machine noises, including: 1) data preprocessing, feature extraction, data cleaning, reconstruction and denoising, and division of training data and test data; 2) building and training a neural network, using a data set to train the model; 3) using the test data in step 1 to test the neural network model; 4) using the model to identify the sound and determine whether it is an abnormal sound. The existing patent has the problem of inaccurate abnormal sound detection.

[0005] How to achieve automatic and accurate detection of abnormal noise from DC transmission converter valves has become a technical problem that needs to be solved. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method and device for detecting abnormal noise of a DC transmission converter valve based on deep learning.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] According to one aspect of the present invention, a method for detecting abnormal noise of a DC transmission converter valve based on deep learning is provided, and the method comprises the following steps:

[0009] 1. A method for detecting abnormal noise of a DC power transmission converter valve based on deep learning, characterized in that the method comprises the following steps:

[0010] Step S1: respectively collecting normal acoustic signals and abnormal sound signals of a DC transmission converter valve, obtaining normal samples and abnormal sound samples after data processing and sample division, dividing the normal samples into a training set, a validation set and a test set in proportion, and dividing all abnormal sound samples into a test set;

[0011] Step S2: input the verification set into the trained abnormal sound detection model to calculate the abnormal sound detection threshold, wherein the abnormal sound detection model is a convolutional attention mechanism module based on an encoder-decoder architecture - a multi-scale convolutional autoencoder;

[0012] Step S3: input the real-time acoustic signal of the DC transmission converter valve into the trained abnormal sound detection model, combine it with the abnormal sound detection threshold calculated in step S2, automatically output the abnormal detection result in real time, and store the normal acoustic signal and the abnormal sound signal separately.

[0013] Preferably, the encoder and decoder respectively include multiple parallel convolutional neural networks and transposed convolutional neural networks of different scales. The encoder is used to learn the multi-scale characteristics of normal acoustic signals of DC transmission converter valves. The decoder reconstructs data of multiple scales based on the characteristics output by the encoder, and adaptively and parallelly fuses the reconstructed data of each scale through a one-dimensional convolution operation.

[0014] Preferably, the encoder and the decoder respectively include n series-connected convolutional layers or transposed convolutional layers, and a convolutional attention mechanism module is connected in series after each of the first n-1 convolutional layers or transposed convolutional layers.

[0015] More preferably, the convolutional attention mechanism module includes a channel attention submodule and a spatial attention submodule connected in series, which are used to learn the channel information and spatial information of the normal acoustic signal of the DC transmission converter valve respectively.

[0016] More preferably, the channel attention submodule first obtains different features from the input features using maximum pooling and average pooling operations respectively, then the two features are passed through a shared multi-layer perceptron and element-by-element summed, and a channel attention vector is calculated using a Sigmoid function, and finally multiplied element-by-element with the input feature to assign the channel attention to the original feature;

[0017] The spatial attention submodule first uses pooling operations to extract average pooling features and maximum pooling features across channels respectively; then the pooling features are cascaded and convolved through one-dimensional convolution, and finally the spatial attention vector is calculated using the Sigmoid function and then assigned to the input features.

[0018] Preferably, the reconstruction loss of the abnormal sound detection model Including mean square error loss and cosine similarity, specifically:

[0019] For input data of length n and reconstructing data Reconstruction loss The calculation formula is as follows:

[0020]

[0021] Among them, λ is an adjustable weight hyperparameter used to balance the distance and angle difference between data; MSE is the mean square error loss between input data and reconstructed data, and CS∈[-1,1] is the cosine similarity between input data and reconstructed data.

[0022] Preferably, the abnormal sound detection model is trained using a training set in an offline mode.

[0023] Preferably, the test set is input into a trained abnormal sound detection model, and the health status of the DC transmission converter valve is obtained by comparing the data reconstruction loss with the abnormal sound detection threshold, and compared with the actual status to verify the performance of the abnormal sound detection model.

[0024] Preferably, the method further comprises updating the abnormal sound detection model, specifically, periodically and randomly extracting a portion of normal signals stored during online abnormal sound detection to retrain the abnormal sound detection model, and calculating a new abnormal sound detection threshold.

[0025] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the method described above is implemented when the processor executes the program.

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

[0027] 1) The present invention collects normal acoustic signals of converter valves to train abnormal sound detection models and calculates abnormal sound detection thresholds, and then uses the trained model for online detection of converter valves. The multi-scale convolutional neural network can learn multi-scale different features from the time series acoustic signals of converter valves, thereby improving the feature extraction capability of the model. After the decoder of the model outputs multi-scale reconstructed data, it adaptively and parallelly fuses data of each scale through a one-dimensional convolution operation to reduce the amount of calculation when calculating reconstruction loss, thereby realizing automatic real-time and accurate detection of abnormal sound of DC transmission converter valves.

[0028] 2) The abnormal sound detection model of the present invention integrates a convolutional autoencoder, a multi-scale convolutional neural network, a convolutional attention mechanism module and an adaptive parallel data fusion mechanism, which improves the feature extraction and data reconstruction capabilities of the model; after the abnormal sound detection model is put into use, normal data is still regularly collected for updating, thereby ensuring the accuracy and robustness of the abnormal sound detection model.

[0029] 3) In the process of selecting the abnormal sound detection threshold, the present invention does not select the maximum value of the validation sample loss value as the threshold, but selects 97% of the value of the reconstruction loss of all validation set samples as the abnormal sound detection threshold, avoiding the influence of individual data on the threshold selection, thereby ensuring the robustness and accuracy of the model.

[0030] 4) In the abnormal sound detection method of the present invention, the abnormal sound signal of the converter valve collected is only used to verify the model and does not participate in the model training process. Therefore, there is no need to provide a large amount of abnormal sound data, which is in line with the situation in actual applications where there is a large amount of normal data but a lack of abnormal sound data, and is easy to put into actual engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flowchart of abnormal noise detection of DC transmission converter valve based on deep learning in the present invention;

[0032] Figure 2 Schematic diagram of the multi-scale data adaptive parallel fusion mechanism in the present invention;

[0033] Figure 3 It is a schematic diagram of the CBAM module structure in the present invention.

[0034] Figure 4 It is a structural schematic diagram of the CBAM-MCAE model in the present invention;

[0035] Figure 5 It is a flow chart of the abnormal noise detection method of the DC transmission converter valve based on deep learning in the present invention;

[0036] Figure 6 It is a comparison chart of abnormal sound detection results of various models in the present invention. DETAILED DESCRIPTION

[0037] 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 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 should fall within the scope of protection of the present invention.

[0038] In view of the blind spots in existing online monitoring systems and manual inspection methods, which make it difficult to detect some converter valve faults in time and require manual intervention, the present invention provides a DC power transmission converter valve abnormal noise detection method based on deep learning. This method determines the health status of the converter valve by comparing the reconstruction loss of the converter valve acoustic signal with the threshold using only the abnormal noise detection model based on deep learning trained with the normal acoustic signal of the converter valve. It has the advantages of no need for abnormal noise data training model, easy to put into application, automatic real-time detection, and high detection accuracy.

[0039] Example 1

[0040] This embodiment relates to a method for detecting abnormal noise of a DC transmission converter valve based on deep learning, such as Figure 1 and Figure 5 , including the following steps:

[0041] Step S1: Data collection and processing, collect the normal acoustic signal and abnormal sound signal of the converter valve respectively, with a sampling frequency of 51.2kHz. Take 10240 sampling points as one sample, and after data processing and sample division, get 1560 normal samples and 330 abnormal sound samples. 600 samples of the normal samples are divided into training sets, 600 samples are divided into verification sets, and the remaining 360 samples are divided into test sets; all abnormal sound samples are divided into test sets. That is, the test set contains 360 normal samples and 330 abnormal sound samples.

[0042] Step S2: offline model training: use the training set to train the abnormal sound detection model, i.e., the Convolutional Block Attention Module-Multi-scale Convolutional Auto-Encoder (CBAM-MCAE), and input the validation set into the trained model to calculate the abnormal sound detection threshold;

[0043] Step S3: offline model testing, input the test set into the trained model, obtain the health status of the converter valve by comparing the data reconstruction loss with the abnormal sound detection threshold, and compare it with the actual status to obtain the abnormal sound detection accuracy, so as to verify the abnormal sound detection performance of the model;

[0044] Step S4: Online abnormal sound detection, using the trained abnormal sound detection model and the calculated abnormal sound detection threshold to detect the real-time acoustic signal of the converter valve, obtain its health status, and store the normal signal and the abnormal sound signal respectively;

[0045] Step S5: Model update, regularly randomly extracting some normal signals stored during online abnormal sound detection to retrain the abnormal sound detection model, and calculate a new abnormal sound detection threshold.

[0046] In step S2, the abnormal sound detection model, namely the Convolutional Block Attention Module-Multi-scale Convolutional Auto-Encoder (CBAM-MCAE), is developed on the basis of the Convolutional Auto-Encoder (CAE). Its basic architecture is encoder-decoder. The encoder reduces the dimension of the data to obtain its hidden representation, and then uses the decoder to increase the dimension of the hidden representation to reconstruct the data. The error between the reconstructed data and the input data, namely the data reconstruction loss, is calculated and back-propagated to optimize the model parameters to complete the model training process.

[0047] When training the model, the batch size is set to 128, the number of training rounds is 300, the learning rate is 0.01, and the Adaptive Moment Estimation algorithm (Adam) is used as the parameter optimization algorithm.

[0048] The encoder receives the input data X (normal acoustic signal of the converter valve) and transforms it into a latent representation H through the following nonlinear mapping:

[0049] H=σ(WX+b)

[0050] Where: σ is the nonlinear activation function of the encoder network. Commonly used nonlinear activation functions include Softmax function, ReLU function, LeakyReLU function, Tanh function, Sigmoid function, etc. This embodiment adopts LeakyReLU function. W is the weight matrix in the encoder network, and b is the bias vector in the encoder network.

[0051] The decoder then maps the hidden representation H back to the original representation in a similar way:

[0052]

[0053] Where σ is the nonlinear activation function of the decoder network, W' is the weight matrix of the decoder network, and b' is the bias vector of the decoder network.

[0054] The autoencoder reconstructs the data by minimizing The reconstruction loss between the input data X To optimize the model parameters θ = [W, b, W', b'], the optimization function of the autoencoder is:

[0055]

[0056] The encoder in CAE is composed of multiple convolutional layers, and each convolutional layer is composed of a one-dimensional convolutional layer, a batch normalization layer, and a nonlinear activation function. The one-dimensional convolutional layer convolves the local area of ​​the input with the convolution kernel and extracts the local features of the input by sliding horizontally or vertically. There are three important parameters in the convolution process: the convolution kernel size, the convolution stride, and the padding number. Among them, the stride determines the distance between each sliding interval, and the padding number can increase the size of the input and avoid the loss of edge information. The size of these three numbers together determines the dimension of the output feature. Taking one-dimensional convolution as an example, assuming that the length of the signal is L, the convolution kernel size is F, the stride is S, and the padding number is P, then the feature length after convolution is (L+2P-F) / S+1. The encoder in CAE reduces the data dimension and extracts data features through the convolution layer. Correspondingly, the decoder in CAE implements upsampling operations through the transposed convolution layer to restore the data dimension to achieve data reconstruction. The feature length after transposed convolution is S×(L-1)+F-2×P.

[0057] Considering the converter valve acoustic signal as a time series, it has multi-time scale characteristics. Therefore, a multi-scale convolutional neural network is introduced to learn the multi-scale characteristics of the converter valve acoustic signal by setting convolution kernels F of different sizes, so as to better reconstruct the converter valve acoustic signal. Three convolution layers with different convolution kernel sizes are set to extract the multi-scale characteristics of the converter valve acoustic signal. The convolution kernel sizes of the first convolution layer of scales one, two, and three in the model are 48, 96, and 192, respectively. After that, the three parallel scales learn different characteristics of the converter valve acoustic data. Figure 2 As shown in the figure, after the decoder of the final model outputs the reconstructed data of three scales, a one-dimensional convolution operation is used to adaptively and parallelly fuse the data of each scale to reduce the amount of computation when calculating the reconstruction loss.

[0058] like Figure 3As shown in the figure, in order to improve the feature extraction and data reconstruction capabilities of CAE, an attention module, namely the Convolutional Block Attention Module (CBAM), is introduced into CAE. This attention module connects the channel attention submodule and the spatial attention submodule in series to help the model learn the channel information and spatial information of the converter valve acoustic signal respectively, thereby improving the model performance. Assume that the feature output of the convolutional layer is Where W and C are the length and number (number of channels) of the features respectively.

[0059] The channel attention submodule first uses the maximum pooling (MaxPool) and average pooling (AvgPool) operations to obtain different features, and then the two features are passed through a shared multilayer perceptron (Multilayer Perceptron, MLP) and element-by-element summation and Sigmoid function to obtain the channel attention vector Finally, it is element-wise multiplied with the input feature to assign channel attention to the original feature Z. The channel attention is calculated as follows:

[0060] W ca =Sigmoid(MLP(AvgPool(Z))+MLP(MaxPool(Z)))

[0061] MLP(·)=W1(Sigmoid(W0(·)))

[0062] Where W0 and W1 are the weights in MLP, and the MLP weights are shared for two pooling features; AvgPool is the average pooling operation, and MaxPool is the maximum pooling operation.

[0063] The spatial attention submodule first uses the pooling operation to extract the average pooling features and the maximum pooling features across channels respectively; then, the pooling features are cascaded and convolved through one-dimensional convolution, and finally the spatial attention vector W is calculated using the Sigmoid function. sa , and then assigned to the input feature Z. The spatial attention is calculated as follows:

[0064] W sa =Sigmoid(Conv1D k ([AvgPool(Z); MaxPool(Z)]))

[0065] Where Conv1D k It is a one-dimensional convolution operation with a convolution kernel size of k, where k = 7.

[0066] like Figure 4 The figure shows the final CBAM-MCAE model structure. The encoder and decoder in the model are composed of five convolutional layers / transposed convolutional layers of three parallel scales connected in series, and a convolutional attention mechanism module is connected in series after the first four convolutional layers / transposed convolutional layers. Finally, the reconstructed data of the three scales output by the decoder are fused by a one-dimensional convolutional layer.

[0067] Reconstruction loss in CBAM-MCAE model It consists of the mean square error (MSE) loss function and the cosine similarity (CS) to simultaneously consider the distance and angle differences between data. For input data of length n and reconstructing data The MSE loss function calculates the square distance between the input data and the reconstructed data, that is:

[0068]

[0069] Cosine similarity CS is also used to measure the difference between data, namely:

[0070]

[0071] Among them: CS∈[-1,1] is the cosine similarity between data, and the smaller the value, the greater the directional difference.

[0072] The final reconstruction loss is The calculation is as follows:

[0073]

[0074] Among them, λ is an adjustable weight hyperparameter used to balance the distance and angle difference between data, and here λ = 0.2.

[0075] In step S2, the process of calculating the abnormal sound detection threshold value specifically includes:

[0076] Input the validation set into the trained CBAM-MCAE model and calculate the reconstruction loss of all normal samples in the validation set. In the process of selecting the abnormal sound detection threshold, in order to ensure the robustness of the model and avoid the influence of individual data on the threshold selection, the maximum value of the validation sample loss value was not selected as the threshold. Instead, the value of 97% of the reconstruction loss of all validation set samples was selected as the abnormal sound detection threshold.

[0077] In the subsequent abnormal noise detection process, the reconstruction loss calculated by the CBAM-MCAE model of the converter valve acoustic data is compared with the selected abnormal noise detection threshold. The data with reconstruction loss greater than the abnormal noise detection threshold are regarded as abnormal noise data, otherwise they are regarded as normal data.

[0078] In step S3, in order to verify the abnormal sound detection performance of the model, the CBAM-MCAE model is compared with other related models, including: CBAM-MCAE-MSE, MCAE, CAE and CAE-MSE models. Among them, the CBAM-MCAE-MSE model is a CBAM-MCAE model that only uses MSE in the loss function; the MCAE model is a CBAM-MCAE model without the CBAM module; the CAE model is the first branch of the MCAE model; and the CAE-MSE model is a CAE model that only uses MSE in the loss function.

[0079] In order to fully compare the abnormal sound detection performance of each model, four evaluation indicators are set, namely accuracy, precision, recall and F1 value. The definitions of these four evaluation indicators are as follows:

[0080]

[0081] Among them, TP, FP, TN, and FN represent the number of correctly detected abnormal sound samples, the number of incorrectly detected normal samples, the number of correctly detected normal samples, and the number of incorrectly detected abnormal sound samples, respectively.

[0082] The abnormal noise detection results of all models are as follows Figure 6 As shown in the figure, it can be seen that the CAE model has a significant improvement in the four evaluation indicators compared with the CAE-MSE model, and the CBAM-MCAE model has a significant improvement in the four evaluation indicators compared with the CBAM-MCAE-MSE model, which verifies the help of using MSE and CS in the loss function to improve the abnormal sound detection model. In addition, the CBAM-MCAE model has a significant increase in the four evaluation indicators compared with the MCAE model, reflecting that the CBAM module enhances the feature extraction ability of the model, thereby improving the abnormal sound detection performance. Comparing the MCAE model and the CAE model, the MCAE model is higher than the CAE model in all four evaluation indicators, reflecting the help of multi-scale convolution in improving the abnormal sound detection performance of the model. In summary, the CBAM-MCAE model has the highest values ​​in all four evaluation indicators, reflecting the superior abnormal sound detection performance of the model.

[0083] Example 2

[0084] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0085] Multiple components in the device are connected to the I / O interface, including: input units, such as keyboards, mice, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as disks, optical disks, etc.; and communication units, such as network cards, modems, wireless communication transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunication networks.

[0086] The processing unit performs the various methods and processes described above, such as methods S1 to S5. For example, in some embodiments, methods S1 to S5 may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more steps of methods S1 to S5 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute methods S1 to S5 in any other appropriate manner (e.g., by means of firmware).

[0087] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0088] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0089] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0090] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for detecting abnormal noise of a DC power transmission converter valve based on deep learning, characterized in that: The method comprises the following steps: Step S1: respectively collecting normal acoustic signals and abnormal sound signals of a DC transmission converter valve, obtaining normal samples and abnormal sound samples after data processing and sample division, dividing the normal samples into a training set, a validation set and a test set in proportion, and dividing all abnormal sound samples into a test set; Step S2: input the verification set into the trained abnormal sound detection model to calculate the abnormal sound detection threshold, wherein the abnormal sound detection model is a convolutional attention mechanism module based on an encoder-decoder architecture - a multi-scale convolutional autoencoder; Step S3: input the real-time acoustic signal of the DC transmission converter valve into the trained abnormal sound detection model, combine it with the abnormal sound detection threshold calculated in step S2, automatically output the abnormal detection result in real time, and store the normal acoustic signal and the abnormal sound signal separately.

2. According to the method for detecting abnormal noise of DC power transmission converter valve based on deep learning in claim 1, it is characterized in that: The encoder and decoder respectively include multiple parallel convolutional neural networks and transposed convolutional neural networks of different scales. The encoder is used to learn the multi-scale features of normal acoustic signals of DC transmission converter valves. The decoder reconstructs data of multiple scales based on the features output by the encoder, and adaptively and parallelly fuses the reconstructed data of each scale through a one-dimensional convolution operation.

3. The method for detecting abnormal noise of a DC power transmission converter valve based on deep learning according to claim 1, characterized in that: The encoder and the decoder respectively include n serially connected convolutional layers or transposed convolutional layers, and a convolutional attention mechanism module is serially connected after each of the first n-1 convolutional layers or transposed convolutional layers.

4. The method for detecting abnormal noise of a DC power transmission converter valve based on deep learning according to claim 2, characterized in that: The convolutional attention mechanism module includes a channel attention submodule and a spatial attention submodule connected in series, which are used to learn the channel information and spatial information of the normal acoustic signal of the DC transmission converter valve respectively.

5. The method for detecting abnormal noise of a DC power transmission converter valve based on deep learning according to claim 4, characterized in that: The channel attention submodule first uses the maximum pooling and average pooling operations to obtain different features from the input features, then the two features are passed through a shared multi-layer perceptron and summed element by element, and the channel attention vector is calculated by the Sigmoid function, and finally multiplied element by element with the input feature to assign the channel attention to the original feature; The spatial attention submodule first uses pooling operations to extract average pooling features and maximum pooling features across channels respectively; then the pooling features are cascaded and convolved through one-dimensional convolution, and finally the spatial attention vector is calculated using the Sigmoid function and then assigned to the input features.

6. The method for detecting abnormal noise of a DC power transmission converter valve based on deep learning according to claim 1, characterized in that: The reconstruction loss of the abnormal sound detection model Including mean square error loss and cosine similarity, specifically: For input data of length n and reconstructing data Reconstruction loss The calculation formula is as follows: Among them, λ is an adjustable weight hyperparameter used to balance the distance and angle difference between data; MSE is the mean square error loss between input data and reconstructed data, and CS∈[-1,1] is the cosine similarity between input data and reconstructed data.

7. The method for detecting abnormal noise of a DC power transmission converter valve based on deep learning according to claim 1, characterized in that: The abnormal sound detection model is trained using the training set in an offline mode.

8. The method for detecting abnormal noise of a DC power transmission converter valve based on deep learning according to claim 1, characterized in that: The test set is input into the trained abnormal sound detection model, and the health status of the DC transmission converter valve is obtained by comparing the data reconstruction loss with the abnormal sound detection threshold, and compared with the actual status to verify the performance of the abnormal sound detection model.

9. The method for detecting abnormal noise of a DC power transmission converter valve based on deep learning according to claim 1, characterized in that: The method also includes updating the abnormal sound detection model, specifically: regularly randomly extracting a portion of normal signals stored during online abnormal sound detection to retrain the abnormal sound detection model, and calculating a new abnormal sound detection threshold.

10. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

Images