Electrical equipment production fault diagnosis method for small sample learning
Through unsupervised migration training of the combined model of the network and the prototype network, the multi-scale attention mechanism and distance penalty loss function are used to solve the problem of low accuracy in the fault diagnosis of small samples in electrical equipment production, and high-precision fault recognition and rapid detection are achieved. Improve the security of power information systems.
Patent Information
- Application Number
- CN202411008513.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-12-06
AI Technical Summary
In the production fault diagnosis of electrical equipment, the deep transfer learning method has low diagnostic accuracy and is difficult to achieve small high-precision due to the semantic gap between the training fault sample and the target fault sample and the domain offset in domain adaptation. Sample fault diagnosis.
The combined model of unsupervised migration training network and prototype network is adopted, and the block and distance penalty loss function are extracted through the multi-scale attention mechanism to build a small sample fault diagnosis model, and the labeled data of the source domain and labelless data of the target domain are used for features. Extraction and fault identification reduce dependence on large number of fault samples.
It realizes high-precision fault identification based on small samples, reduces the time and cost of fault sample collection, improves the speed and accuracy of fault detection, and enhances the safety and reliability of power information systems.
Smart Images

Figure CN119089258A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of electrical equipment production fault diagnosis, and in particular to an electrical equipment production fault diagnosis method for small sample learning. Background Art
[0002] Fault diagnosis of electrical equipment production is an important part of the electrical equipment production process.
[0003] It is time-consuming, labor-intensive and costly to obtain enough fault samples during the degradation process of industrial equipment from healthy to faulty. In order to solve the problem of difficulty in collecting fault samples, the existing technology combines the deep neural network method with the maximum mean difference term to propose a deep transfer learning method for fault diagnosis; sparse autoencoders can also be used to perform domain adaptive learning on training fault samples and target fault samples. This type of method solves the problem of missing target fault samples to a certain extent, but the semantic gap between training fault samples and target fault samples and the domain shift problem in domain adaptation lead to generally low diagnostic accuracy.
[0004] Therefore, it is necessary to develop a fault diagnosis method in the production process of electrical equipment to achieve high-precision fault diagnosis based on small samples. Summary of the invention
[0005] The object of the present invention is to provide an electrical equipment production fault diagnosis method for small sample learning, which can achieve high-precision fault diagnosis based on small samples.
[0006] The embodiments of the present invention are implemented by the following technical solutions:
[0007] A method for diagnosing electrical equipment production faults for small sample learning comprises the following steps:
[0008] Install multiple sensors and collect fault signals through the multiple sensors respectively;
[0009] Perform data preprocessing to obtain unsupervised transfer training sets, meta-training tasks, and meta-testing tasks;
[0010] Identify production faults of electrical equipment through trained small sample fault diagnosis models;
[0011] The small sample fault diagnosis model includes an unsupervised transfer training network and a prototype network, the unsupervised transfer training set is connected to the input end of the unsupervised transfer training network, the meta-training task and the meta-test task are connected to the input end of the prototype network, and the feature extractor of the unsupervised transfer training network is connected to the second feature extractor of the prototype network.
[0012] Preferably, the data preprocessing is performed by:
[0013] Obtaining labeled data in a source domain and unlabeled data in a target domain as the unsupervised transfer training set;
[0014] Obtaining labeled data in the source domain as the meta-training task;
[0015] Obtain labeled data of the target domain as the meta-testing task.
[0016] Preferably, the unsupervised transfer training network includes a first feature extractor, a source domain classifier and a domain discriminator; the source domain classifier outputs the fault category, and the domain discriminator is used to identify the source of the feature;
[0017] The unsupervised transfer training set is connected to an input end of the first feature extractor, and an output end of the first feature extractor is connected to an input end of the source domain classifier and an input end of the domain discriminator.
[0018] Preferably, the first feature extractor adopts a feature extraction block based on a multi-scale attention mechanism, including multiple branch networks and a fusion network; the output ends of the multiple branch networks are respectively connected to the input end of the fusion network;
[0019] Each of the branch networks includes a one-dimensional convolution module and an attention module connected in sequence;
[0020] The one-dimensional convolution module uses different void rates to extract features with different resolutions.
[0021] Preferably, the attention module includes a dimensionality reduction module, a matrix transposition module, a first matrix multiplier, a compression module, a correction module and a first matrix dot multiplier;
[0022] The dimension reduction module is used to reduce the dimension of the output of the one-dimensional convolution module. The output end of the dimension reduction module is connected to the input end of the matrix transposition module and the first input end of the first matrix multiplier. The output end of the matrix transposition module is connected to the second input end of the first matrix multiplier. The output end of the first matrix multiplier is connected to the input end of the compression module. The output end of the compression module is connected to the input end of the correction module. The output end of the one-dimensional convolution module and the output end of the correction module are respectively connected to the first input end and the second input end of the first matrix dot multiplier.
[0023] Preferably, the fusion network includes a first splicing module, a first transformation module, a second transformation module, a second matrix multiplier, a compression module, a fully connected layer, a second matrix point multiplier and a second splicing module;
[0024] The first concatenation module is used to concatenate the outputs of the branch networks together; the first transformation module is used for dimensionality reduction, and the second transformation module is used for dimensionality reduction and then matrix transposition;
[0025] The output end of the first concatenation module is connected to the input end of the first transformation module and the input end of the second transformation module, the output end of the first transformation module and the output end of the second transformation module are respectively connected to the first input end and the second input end of the second matrix multiplier; the output end of the second matrix multiplier is connected to the input end of the fully connected layer, the output end of the compression module is connected to the input end of the fully connected layer, and the output end of the fully connected layer and the output end of the branch network are respectively connected to the two input ends of the second matrix dot multiplier;
[0026] The second concatenation module is used to concatenate the dot product results of the output of each branch network and the output of the fully connected layer.
[0027] Preferably, the prototype network comprises a second feature extractor and a feature encoder connected in sequence; the feature encoder maps the basic fault features extracted by the second feature extractor into a metric feature vector.
[0028] Preferably, the feature encoder works as follows:
[0029] Using two convolutional layers and two maximum pooling layers to perform nonlinear transformation and dimension reduction on the fault features output by the second feature extractor;
[0030] The auxiliary convolution layer is used to adjust the number of feature map channels to d;
[0031] Use global average pooling to convert the feature map into a d-dimensional feature vector.
[0032] Preferably, the prototype network adopts a distance penalty loss function L pd (θ):
[0033]
[0034] Among them, m1 and m2 are constants, and m1 <m2, To support the k-th type of fault samples in the set, c i 、c j 、c kare the prototype centers calculated from the features of the i-th, j-th, and k-th samples in the support set, respectively. d(.,.) represents the function of finding the distance between two parameters. for Medium Sample Eigenvectors in metric space.
[0035] The technical solution of the embodiment of the present invention has at least the following advantages and beneficial effects:
[0036] The present invention can build a fault model without acquiring a large number of fault samples, which helps to save the time and computing power costs caused by sample collection;
[0037] Based on a small sample, the present invention is not affected by the difference between the training fault sample and the target fault sample, can achieve high-precision fault identification, and improve the accuracy of the model in small sample fault diagnosis tasks;
[0038] The present invention helps to speed up the fault detection, making the power information system safer and more reliable, and has great application value;
[0039] The present invention improves the loss function of the prototype network and can better improve the classification performance of the prototype network. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 A schematic diagram of the structure of a small sample fault diagnosis model provided by an embodiment of the present invention;
[0041] Figure 2 A schematic diagram of the working principle of the first feature extractor provided in an embodiment of the present invention;
[0042] Figure 3 A schematic diagram of the working principle of the fusion network provided by an embodiment of the present invention;
[0043] Icon: 1-. DETAILED DESCRIPTION
[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0045] Example 1
[0046] This embodiment provides an electrical equipment production fault diagnosis method for small sample learning, comprising the following steps:
[0047] During the production and operation of electrical equipment, multiple sensors are installed to collect fault signals respectively through the multiple sensors;
[0048] Perform data preprocessing to obtain unsupervised transfer training sets, meta-training tasks, and meta-testing tasks;
[0049] Identify production faults of electrical equipment through trained small sample fault diagnosis models;
[0050] See also Figure 1 The small sample fault diagnosis model includes an unsupervised transfer training network and a prototype network, the unsupervised transfer training set is connected to the input end of the unsupervised transfer training network, the meta-training task and the meta-test task are connected to the input end of the prototype network, and the feature extractor of the unsupervised transfer training network is connected to the second feature extractor of the prototype network to realize parameter migration. Specifically, the source domain is a working condition with a large amount of fault data, and on the other hand, the special working condition is used as the target domain.
[0051] In the model of this embodiment, the unsupervised transfer training network uses the labeled data of the source domain and the unlabeled diagnostic data of the target domain as the unsupervised transfer training set, and performs unsupervised transfer training on the feature extractor of the unsupervised transfer training network, so that the feature extractor can extract domain-invariant features. On the other hand, the feature extractor in the prototype network can extract basic fault features from the original vibration signal, and the prototype network can also be used to map the basic fault features into a metric feature space suitable for distance measurement. The prototype network can adopt a meta-training method, taking a large number of meta-training tasks sampled and divided from a large number of source domain labeled fault data as the training set. After multiple rounds of training, the model learns a large amount of small sample fault diagnosis knowledge, and then performs fault diagnosis on several meta-test tasks sampled and divided from the target domain labeled fault data.
[0052] Example 2
[0053] This embodiment further illustrates data preprocessing based on the technical solution of Embodiment 1.
[0054] In this embodiment, the data preprocessing is performed by:
[0055] Obtaining labeled data in a source domain and unlabeled data in a target domain as the unsupervised transfer training set;
[0056] Obtaining labeled data in the source domain as the meta-training task;
[0057] Obtain labeled data of the target domain as the meta-testing task.
[0058] Example 3
[0059] This embodiment is based on the technical solution of Embodiment 1 and further explains the unsupervised transfer training network.
[0060] In this embodiment, the unsupervised transfer training network includes a first feature extractor, a source domain classifier and a domain discriminator; the source domain classifier outputs the fault category, and the domain discriminator is used to identify the source of the feature;
[0061] The unsupervised transfer training set is connected to an input end of the first feature extractor, and an output end of the first feature extractor is connected to an input end of the source domain classifier and an input end of the domain discriminator.
[0062] Furthermore, the first feature extractor adopts a feature extraction block based on a multi-scale attention mechanism, including multiple branch networks and a fusion network; the output ends of the multiple branch networks are respectively connected to the input end of the fusion network;
[0063] Each of the branch networks includes a one-dimensional convolution module and an attention module connected in sequence;
[0064] The one-dimensional convolution module uses different void rates to extract features with different resolutions.
[0065] In addition, the attention module includes a dimensionality reduction module, a matrix transposition module, a first matrix multiplier, a compression module, a correction module and a first matrix dot multiplier;
[0066] The dimension reduction module is used to reduce the dimension of the output of the one-dimensional convolution module. The output end of the dimension reduction module is connected to the input end of the matrix transposition module and the first input end of the first matrix multiplier. The output end of the matrix transposition module is connected to the second input end of the first matrix multiplier. The output end of the first matrix multiplier is connected to the input end of the compression module. The output end of the compression module is connected to the input end of the correction module. The output end of the one-dimensional convolution module and the output end of the correction module are respectively connected to the first input end and the second input end of the first matrix dot multiplier.
[0067] On this basis, the fusion network includes a first splicing module, a first transformation module, a second transformation module, a second matrix multiplier, a compression module, a fully connected layer, a second matrix point multiplier, and a second splicing module;
[0068] The first concatenation module is used to concatenate the outputs of the branch networks together; the first transformation module is used for dimensionality reduction, and the second transformation module is used for dimensionality reduction and then matrix transposition;
[0069] The output end of the first concatenation module is connected to the input end of the first transformation module and the input end of the second transformation module, the output end of the first transformation module and the output end of the second transformation module are respectively connected to the first input end and the second input end of the second matrix multiplier; the output end of the second matrix multiplier is connected to the input end of the fully connected layer, the output end of the compression module is connected to the input end of the fully connected layer, and the output end of the fully connected layer and the output end of the branch network are respectively connected to the two input ends of the second matrix dot multiplier;
[0070] The second concatenation module is used to concatenate the dot product results of the output of each branch network and the output of the fully connected layer.
[0071] Based on the technical solution of this embodiment, the three modules of the first feature extractor, the source domain classifier and the domain discriminator work together, and the source domain labeled fault data and the target domain unlabeled fault diagnosis data are used as unsupervised transfer training sets to perform unsupervised transfer training on the first feature extractor, so that the feature extractor can extract domain-invariant features.
[0072] The first feature extractor uses a feature extraction block with a parallel multi-branch structure to extract scale features with different receptive fields. As an implementation example, its working principle can be found in Figure 2 In this embodiment, four branch networks are set up. Each branch network first uses a one-dimensional convolution with different void rates to extract the input feature F in Features with different resolutions can be extracted from the network. For example, a one-dimensional convolution with a dilation rate of 2 is used in the second branch network, which means that there is a zero value between any convolution kernel weight parameters in the convolution kernel; a one-dimensional convolution with a dilation rate of 3 is used in the third branch network, which means that there are two zero values between any convolution kernel weight parameters in the convolution kernel. The increase in the number of zero values leads to an increase in the receptive field of the convolution kernel, so the third branch can extract more global features than the second branch. At the same time, zero padding is used in the design of the one-dimensional dilated convolution kernel, and the padding and dilation rate are set equal, with a step size of 1, to keep the sizes of the output feature maps F1, F2, F3, and F4 of each branch network consistent.
[0073] The size of the above output feature maps F1, F2, F3, and F4 is C×l×W, where C represents the number of channels in the feature map and l×W is the size of each channel feature map. The W elements in the feature map of each channel are obtained by convolving the input feature with a one-dimensional convolution kernel in different spaces, and different fault signals have different manifestations in different spaces. The attention mechanism in each network branch assigns different attention weights to the elements in each channel feature map of the feature map.
[0074] The attention module used in this embodiment first uses l×l convolution to reduce the dimension of the feature map F of each scale size C×l×W, removes the channel dimension, and obtains a one-dimensional attention weight vector Fl of size l×W, and then transposes Fl T Perform matrix multiplication with Fl to obtain the attention weight matrix A of size W×W. Combining the matrix multiplication operation process, it can be seen that the element A in the weight matrix A ij Represents the influence of the i-th pixel on the j-th pixel in the feature map. After obtaining the weight matrix A, it is compressed from the column dimension, and the mean of each column element is taken as the result of the column. Finally, a one-dimensional attention weight vector A' with a size of l×W is obtained, which contains the dependency relationship between elements at each spatial position. The expression of the compression module is:
[0075]
[0076] In order to continuously correct A' during the training process and avoid introducing more training parameters, a one-dimensional convolution layer with a convolution kernel width of 5 can be used to correct A', and the calculated result is normalized by the softmax function to obtain the final spatial attention vector A". Finally, the l×W spatial attention vector element A" is multiplied by the i-th spatial position pixel on all channels of the feature map F to obtain the feature map F with the spatial attention weight applied. s .
[0077] The final fusion network is equivalent to taking into account the different effects of feature maps of different scales on fault classification, calculating the weight coefficients of output feature maps of different scales based on the scale attention mechanism, and using the weight coefficients to weight the output feature maps of the corresponding scales, and finally splicing and fusion.
[0078] How Converged Networks Work Figure 3 , its working method is to first concatenate the output feature maps of each scale into a feature map F' of size C×4×W, where 4 represents the number of scale feature maps. Then, each scale feature map in the feature map F is reduced in dimension to obtain a feature map F2 of size 4×N, where N=C×W. Then, the same method as the spatial attention mechanism is used to establish a correlation matrix B between scales of size 4×4, and the same method as the formula is used to compress the scale weight vector B' of 4×1. Then, the initialized weight vector B' is input into a fully connected layer with both input and output of 4 to achieve the correction learning of the weight coefficients of each scale. Finally, the weight coefficients of each scale are normalized through the softmax operation to obtain the final scale weight vector α=[α1,α2,α3,α4]. Based on the weight vector α and the original scale feature maps F1 s , F2 s , F3s , F4 s , and then perform element-wise multiplication and concatenation in the channel dimension to obtain the fused feature map F con :
[0079] F con =[α1·F1 s ,α2·F2 s ,α3·F3 s ,α4·F4 s ].
[0080] Since the above processing concatenates and fuses the output feature maps of each scale, the number of channels of the output feature map of the feature extraction module increases compared to the number of channels of the input feature map. in And the output feature map F con The size of the feature map F is consistent, so the feature map F can be fused con A 1×1 convolution operation is performed to resize it to the same size as the input feature map.
[0081] The following is a further explanation of the structural parameter selection of each module in this embodiment:
[0082] The feature extractor network layer parameters of the first feature extractor module used in this embodiment can be found in Table 1:
[0083] Table 1. Feature extractor network layer parameters
[0084]
[0085] Its input is a one-dimensional fault vibration signal of length. In order to reduce the parameter scale of the model, a one-dimensional convolution block with a convolution kernel width of 15 and an output channel of 16 can be used to reduce the length of the fault vibration signal. Then a one-dimensional convolution block with a width of 3 and an output channel of 32 is used to increase the number of channels of the feature map and extract the basic features of the fault signal. Then a maximum pooling layer with a window length of 2 and a step size of 2 is used to halve the length of the feature map. After performing the above simple feature extraction of the vibration signal, two MSA blocks are used to extract deep and shallow fault features. Each MSAblock contains four parallel branches with void rates of 1, 2, 3, and 4, a window length of 3, an input channel of 32, and an output channel of 32. The network layers contained in the spatial attention mechanism and the scale attention mechanism are consistent with those in Section 4.5.2. A maximum pooling layer with a window length of 2 and a step size of 2 is added between the two MSA blocks to halve the length of the feature map. Similar to common convolutional neural networks, in addition to the convolutional layer, the above convolutional block also adds a BN layer and a ReLU activation function to improve the generalization and nonlinear expression capabilities of the model. The parameters of each layer of the feature extractor are shown in 1 below, where KC (15, 1, 16) indicates that the convolution kernel width is 15, the input channel is 1, and the output channel is 16; KMP (2, 2) indicates that the convolution kernel width is 2 and the step size is 2; MS1_block1 represents the one-dimensional convolution layer of the first branch MS1 in the first MSA Block.
[0086] The input of the source domain classifier module is the fault feature map extracted by the feature extractor from the source domain sample, and the fault category to which the fault feature map belongs is output. The source domain classifier can use the global average pooling GAP layer instead of the fully connected layer to reduce the dimension of the feature map, thereby improving the generalization ability of the entire network. Therefore, the source domain classifier first includes two convolutional pooling modules to further nonlinearly transform and reduce the dimension of the high-dimensional feature map extracted by the feature extractor, and then uses an auxiliary convolution block ConvBlock3 to adjust the number of channels of the feature map to be consistent with the number of label categories, and then performs a global average pooling operation on the feature map to replace the features of each channel with the mean, and finally inputs the mean feature into the Softmax layer to output the classification probability. The detailed parameters of each layer of the network are shown in Table 2, where k represents the number of categories of fault classification:
[0087] Table 2 Network layer parameters of source domain classifier module
[0088]
[0089]
[0090] The domain discriminator can also receive high-dimensional feature maps corresponding to samples from the source domain dataset and the target domain unlabeled dataset, and determine their sources based on the feature maps. The network structure of the domain discriminator is similar to that of the source domain classifier module, consisting of a convolutional pooling layer and a global average pooling layer. Since the domain discrimination task is a binary classification task, the Softmax classification function is replaced by a Sigmoid function. In addition, a gradient reversal layer can be added to the head of the domain discriminator network so that the gradient of the discriminator is automatically reversed during back propagation, and adversarial training of the feature extractor and the domain discriminator can be achieved in one training process. The parameters of each layer of the domain discriminator network are shown in Table 3 below:
[0091] Table 3 Domain discriminator network layer parameters
[0092]
[0093] Example 4
[0094] This embodiment further illustrates the prototype network based on the technical solution of Embodiment 1.
[0095] In this embodiment, the prototype network includes a second feature extractor and a feature encoder connected in sequence; the feature encoder maps the basic fault features extracted by the second feature extractor into a metric feature vector.
[0096] Furthermore, the feature encoder works as follows:
[0097] Using two convolutional layers and two maximum pooling layers to perform nonlinear transformation and dimension reduction on the fault features output by the second feature extractor;
[0098] The auxiliary convolution layer is used to adjust the number of feature map channels to d;
[0099] Use global average pooling to convert the feature map into a d-dimensional feature vector.
[0100] The prototype network adopts the meta-training method, and takes a large number of meta-training tasks sampled and divided from a large number of source domain labeled fault data as the training set. After multiple rounds of training, the model learns a large amount of small sample fault diagnosis knowledge, and then performs fault diagnosis on several meta-test tasks sampled and divided from the target domain labeled fault data. The fault features extracted by the second feature extractor from the one-dimensional fault vibration signal as basic features can be well used for common classification tasks. Since the second feature extractor has a good fault feature extraction ability under the unsupervised transfer pre-training of a large number of source domain and target domain samples, the parameters of the feature extractor can be fixed during the meta-training process. At the same time, a feature encoder with a simple structure and a small number of parameters is introduced to map the basic fault features extracted by the feature extractor into a metric feature vector suitable for distance calculation. In the feature encoder, two convolutional layers and two maximum pooling layers are first used to perform nonlinear transformation and dimensionality reduction on the fault features output by the feature extractor, and then an auxiliary convolutional layer is used to adjust the number of feature map channels to d. Finally, global average pooling is used to convert the feature map into a d-dimensional feature vector as the representation of the fault sample in the metric space. The parameters of each network layer of the feature encoder are shown in Table 4 below:
[0101] Table 4 Feature encoder network layer parameters
[0102]
[0103]
[0104] As a preferred solution of this embodiment, the prototype network adopts the distance penalty loss function L pd (θ):
[0105]
[0106] Among them, m1 and m2 are constants, and m1 <m2, To support the k-th type of fault samples in the set, c i 、c j 、c k are the prototype centers calculated from the features of the i-th, j-th, and k-th samples in the support set, respectively. d(.,.) represents the function of finding the distance between two parameters. for Medium Sample Eigenvectors in metric space.
[0107] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for diagnosing electrical equipment production faults for small sample learning, characterized in that: The following steps are involved: Install multiple sensors and collect fault signals through the multiple sensors respectively; Perform data preprocessing to obtain unsupervised transfer training sets, meta-training tasks, and meta-testing tasks; Identify production faults of electrical equipment through trained small sample fault diagnosis models; The small sample fault diagnosis model includes an unsupervised transfer training network and a prototype network, the unsupervised transfer training set is connected to the input end of the unsupervised transfer training network, the meta-training task and the meta-test task are connected to the input end of the prototype network, and the feature extractor of the unsupervised transfer training network is connected to the second feature extractor of the prototype network.
2. The method for diagnosing electrical equipment production faults for small sample learning according to claim 1, characterized in that: The data preprocessing is performed by: Obtaining labeled data in a source domain and unlabeled data in a target domain as the unsupervised transfer training set; Obtaining labeled data in the source domain as the meta-training task; Obtain labeled data of the target domain as the meta-testing task.
3. The method for diagnosing electrical equipment production faults for small sample learning according to claim 1, characterized in that: The unsupervised transfer training network includes a first feature extractor, a source domain classifier and a domain discriminator; the source domain classifier outputs the fault category, and the domain discriminator is used to identify the source of the feature; The unsupervised transfer training set is connected to an input end of the first feature extractor, and an output end of the first feature extractor is connected to an input end of the source domain classifier and an input end of the domain discriminator.
4. The method for diagnosing electrical equipment production faults for small sample learning according to claim 3, characterized in that: The first feature extractor adopts a feature extraction block based on a multi-scale attention mechanism, including multiple branch networks and a fusion network; the output ends of the multiple branch networks are respectively connected to the input end of the fusion network; Each of the branch networks includes a one-dimensional convolution module and an attention module connected in sequence; The one-dimensional convolution module uses different void rates to extract features with different resolutions.
5. The method for diagnosing electrical equipment production faults for small sample learning according to claim 4, characterized in that: The attention module includes a dimensionality reduction module, a matrix transposition module, a first matrix multiplier, a compression module, a correction module and a first matrix dot multiplier; The dimension reduction module is used to reduce the dimension of the output of the one-dimensional convolution module. The output end of the dimension reduction module is connected to the input end of the matrix transposition module and the first input end of the first matrix multiplier. The output end of the matrix transposition module is connected to the second input end of the first matrix multiplier. The output end of the first matrix multiplier is connected to the input end of the compression module. The output end of the compression module is connected to the input end of the correction module. The output end of the one-dimensional convolution module and the output end of the correction module are respectively connected to the first input end and the second input end of the first matrix dot multiplier.
6. The method for diagnosing electrical equipment production faults for small sample learning according to claim 5, characterized in that: The fusion network includes a first splicing module, a first transformation module, a second transformation module, a second matrix multiplier, a compression module, a fully connected layer, a second matrix point multiplier, and a second splicing module; The first concatenation module is used to concatenate the outputs of the branch networks together; the first transformation module is used for dimensionality reduction, and the second transformation module is used for dimensionality reduction and then matrix transposition; The output end of the first concatenation module is connected to the input end of the first transformation module and the input end of the second transformation module, the output end of the first transformation module and the output end of the second transformation module are respectively connected to the first input end and the second input end of the second matrix multiplier; the output end of the second matrix multiplier is connected to the input end of the fully connected layer, the output end of the compression module is connected to the input end of the fully connected layer, and the output end of the fully connected layer and the output end of the branch network are respectively connected to the two input ends of the second matrix dot multiplier; The second concatenation module is used to concatenate the dot product results of the output of each branch network and the output of the fully connected layer.
7. The method for diagnosing electrical equipment production faults for small sample learning according to claim 1, characterized in that: The prototype network includes a second feature extractor and a feature encoder connected in sequence; the feature encoder maps the basic fault features extracted by the second feature extractor into a metric feature vector.
8. The method for diagnosing electrical equipment production faults for small sample learning according to claim 7, characterized in that: The feature encoder works as follows: Using two convolutional layers and two maximum pooling layers to perform nonlinear transformation and dimension reduction on the fault features output by the second feature extractor; The auxiliary convolution layer is used to adjust the number of feature map channels to d; Use global average pooling to convert the feature map into a d-dimensional feature vector.
9. The method for diagnosing electrical equipment production faults for small sample learning according to claim 8, characterized in that: The prototype network adopts the distance penalty loss function L pd (θ): Among them, m1 and m2 are constants, and m1<m2, To support the k-th type of fault samples in the set, c i 、c j 、c k are the prototype centers calculated from the features of the i-th, j-th, and k-th samples in the support set, respectively. d(.,.) represents the function of finding the distance between two parameters. for Medium Sample Eigenvectors in metric space.
Citation Information
Patent Citations
Complex working condition gear fault diagnosis method based on matching element learning under few samples
CN116593157A
