MSFF-CNN-based post insulator fault identification method and system

Abstract

The invention discloses an MSFF-CNN-based post insulator fault identification method and system. The method comprises the steps of obtaining vibration data of a to-be-detected post insulator; performing multi-scale feature extraction on the vibration data by using a pre-trained MSFF-CNN model, performing fault identification based on the extracted multi-scale features, and determining the fault type of the post insulator to be detected; wherein the MSFF-CNN model comprises an input layer, a convolution module, a multi-scale convolution module, a cavity convolution module, a global average pooling layer and an output layer. According to the invention, through the convolution module, the multi-scale convolution module and the cavity convolution module in the MSFF-CNN model, the feature extraction capability of the vibration data of the post insulator can be improved, so that the feature extraction precision of the vibration data is effectively improved; fault recognition is carried out based on the extracted multi-scale features, so that the classification accuracy and robustness of fault types can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of insulator detection, and in particular to a post insulator fault identification method and system based on MSFF-CNN. Background Art

[0002] Porcelain post insulators are important high-voltage electrical equipment in substations. They are mainly used for supporting, fixing and insulating substation disconnectors and busbars. Their operating reliability directly affects the safety of power grid operation. In practical applications, the common type of post insulator failure is the fracture of the connection between the insulator shed and the flange. According to the location of the failure, the post insulator failure type is usually divided into upper flange failure and lower flange failure.

[0003] In the prior art, deep learning algorithms are usually used to extract features from different fault vibration signals of post insulators to achieve fault identification of post insulators. However, under some harsh conditions, the vibration signals collected by drones contain strong noise interference, which makes it difficult for the currently commonly used convolutional neural networks to effectively extract features, making it difficult to effectively identify faults of post insulators. Summary of the invention

[0004] The present invention provides a post insulator fault identification method and system based on MSFF-CNN. Through the convolution module, multi-scale convolution module and void convolution module in the MSFF-CNN model, the feature extraction capability of the vibration data of the post insulator can be improved, thereby effectively improving the feature extraction accuracy of the vibration data. Fault identification based on the extracted multi-scale features can effectively improve the classification accuracy and robustness of the fault type.

[0005] In order to solve the above technical problems, a first aspect of an embodiment of the present invention provides a post insulator fault identification method based on MSFF-CNN, comprising:

[0006] Obtain vibration data of the post insulator to be tested;

[0007] The vibration data is subjected to multi-scale feature extraction using a pre-trained MSFF-CNN model, and fault identification is performed based on the extracted multi-scale features to determine the fault type of the post insulator to be tested; wherein the MSFF-CNN model includes an input layer, a convolution module, a multi-scale convolution module, a hole convolution module, a global average pooling layer and an output layer.

[0008] As a preferred solution, the method of extracting multi-scale features from the vibration data using a pre-trained MSFF-CNN model, and performing fault identification based on the extracted multi-scale features to determine the fault type of the post insulator to be tested specifically includes:

[0009] Performing feature extraction on the vibration data received by the input layer through the convolution module to obtain initial fault features and output them to the multi-scale convolution module;

[0010] Performing multi-scale feature extraction on the initial fault feature through the multi-scale convolution module to obtain the initial multi-scale feature and output it to the hole convolution module;

[0011] Perform feature extraction of different receptive fields on the initial multi-scale features through the dilated convolution module to obtain the multi-scale features and output them to the global average pooling layer;

[0012] The activation values ​​of the multi-scale features are averaged through the global average pooling layer to generate a global feature vector and output it to the output layer;

[0013] The probability of each fault type corresponding to the global feature vector is obtained by using the Softmax function through the output layer, and the fault type with the highest probability is used as the fault type of the post insulator to be tested.

[0014] As a preferred solution, the convolution module is constructed by a first convolution layer, a first batch normalization layer and a first maximum pooling layer connected in sequence.

[0015] As a preferred solution, the convolution kernel size of the first convolution layer is 8×8.

[0016] As a preferred solution, the multi-scale convolution module is constructed by sequentially connected multi-scale convolution layers, first residual connection layers, second batch normalization layers and second maximum pooling layers; the multi-scale convolution layers include several second convolution layers with different convolution kernel sizes.

[0017] As a preferred solution, the multi-scale convolution layer includes three second convolution layers, and the convolution kernel sizes of the three second convolution layers are 3×3, 6×6 and 9×9 respectively.

[0018] As a preferred solution, the dilated convolution module is constructed by sequentially connected dilated convolution units, a second residual connection layer, a third batch normalization layer and a third maximum pooling layer; the dilated convolution unit includes several dilated convolution layers with different expansion rates.

[0019] As a preferred solution, the dilated convolution unit includes three dilated convolution layers, and the expansion rates of the three dilated convolution layers are 1, 3 and 5 respectively.

[0020] As a preferred solution, the method specifically trains the MSFF-CNN model through the following steps:

[0021] Collecting several post insulator fault vibration data and dividing them according to the preset single sample data volume to obtain the fault vibration data set;

[0022] Adding labels to the fault vibration data set to obtain a training data set; wherein the labels include normal data labels, upper flange fault data labels, and lower flange fault data labels;

[0023] The MSFF-CNN model is trained using the training data set to optimize the parameters of the MSFF-CNN model.

[0024] A second aspect of an embodiment of the present invention provides a post insulator fault identification system based on MSFF-CNN, comprising:

[0025] A data acquisition module, used to acquire vibration data of the post insulator to be tested;

[0026] A fault identification module is used to extract multi-scale features from the vibration data using a pre-trained MSFF-CNN model, and to perform fault identification based on the extracted multi-scale features to determine the fault type of the post insulator to be tested; wherein the MSFF-CNN model includes an input layer, a convolution module, a multi-scale convolution module, a hole convolution module, a global average pooling layer and an output layer.

[0027] Compared with the prior art, the beneficial effect of the embodiments of the present invention is that, through the convolution module, multi-scale convolution module and void convolution module in the MSFF-CNN model, the feature extraction capability of the vibration data of the post insulator can be improved, thereby effectively improving the feature extraction accuracy of the vibration data, and fault identification based on the extracted multi-scale features can effectively improve the classification accuracy and robustness of the fault type. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of a post insulator fault identification method based on MSFF-CNN in an embodiment of the present invention;

[0029] Figure 2 Schematic diagram of the structure of the MSFF-CNN model in an embodiment of the present invention;

[0030] Figure 3 is a schematic diagram of multi-scale convolution in an embodiment of the present invention;

[0031] Figure 4 is a schematic diagram of a dilated convolution in an embodiment of the present invention;

[0032] Figure 5 is a schematic diagram of residual connection in an embodiment of the present invention;

[0033] Figure 6 is an overall time domain diagram of data sets of different fault types in an embodiment of the present invention;

[0034] Figure 7 is a time domain diagram of a single sample of a data set of different fault types in an embodiment of the present invention;

[0035] Figure 8 Schematic diagram of the training accuracy curve and the test accuracy curve of the MSFF-CNN model in an embodiment of the present invention;

[0036] Fig. 9 Schematic diagram of the training loss curve and the test loss curve of the MSFF-CNN model in an embodiment of the present invention;

[0037] Fig.10 is a confusion matrix diagram of a test set in an embodiment of the present invention;

[0038] Fig.11 : is a KPCA dimensionality reduction visualization diagram of the distribution of raw data in an embodiment of the present invention;

[0039] Fig.12 : is a KPCA dimensionality reduction visualization diagram of the distribution of the data after training classification in an embodiment of the present invention;

[0040] Fig.13 Schematic diagram of test accuracy curves of various models in the embodiments of the present invention;

[0041] Fig.14 Schematic diagram of test loss curves of various models in the embodiments of the present invention;

[0042] Fig.15 is a confusion matrix diagram of a new test set in an embodiment of the present invention;

[0043] Fig.16 is a signal time domain diagram of adding Gaussian white noise with different signal-to-noise ratios to a normal vibration signal in an embodiment of the present invention;

[0044] Fig.17 is a signal time domain diagram in which Gaussian white noise with different signal-to-noise ratios is added to the upper flange fault signal in an embodiment of the present invention;

[0045] Fig.18 is a signal time domain diagram in which Gaussian white noise with different signal-to-noise ratios is added to the lower flange fault signal in an embodiment of the present invention;

[0046] Fig.19 It is a schematic diagram of the test accuracy of various models with different components in an embodiment of the present invention;

[0047] Fig. 20is a bar chart of the test accuracy of each model in the embodiment of the present invention under data sets divided according to different proportions;

[0048] Fig.21 It is a structural schematic diagram of a post insulator fault identification system based on MSFF-CNN in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] See also Figure 1 According to a first aspect of an embodiment of the present invention, a post insulator fault identification method based on MSFF-CNN is provided, comprising the following steps S1 and S2:

[0051] Step S1, obtaining vibration data of the post insulator to be tested;

[0052] Step S2, using a pre-trained MSFF-CNN model to extract multi-scale features from the vibration data, and performing fault identification based on the extracted multi-scale features to determine the fault type of the post insulator to be tested; wherein the MSFF-CNN model includes an input layer, a convolution module, a multi-scale convolution module, a hole convolution module, a global average pooling layer and an output layer.

[0053] It is worth noting that when a pillar insulator fails, the fault types are diverse, the transmission paths are complex, and it is easy to couple with other electrical components, causing the vibration signal to present multi-scale fault characteristics. The traditional convolutional neural network can only extract features of a single scale, so when the vibration signal contains strong noise interference, the feature extraction effect cannot be ensured, affecting the accuracy of fault identification. Compared with the traditional convolutional neural network, the MSFF-CNN model in this embodiment integrates the multi-scale convolution module and the hole convolution module on the basis of the conventional convolution module, so that it can extract features of different scales and different receptive fields, significantly improving the feature extraction capability for vibration data, so that the fault identification of the pillar insulator to be tested can be accurately performed based on the extracted multi-scale features. In addition, by replacing the fully connected layer of the traditional convolutional neural network with the global average pooling layer, the global feature information is integrated while reducing the dimension, and the classification probability of each fault type can be clarified in combination with the Softmax function of the output layer, thereby improving the interpretability of the model.

[0054] As a preferred solution, the method of extracting multi-scale features from the vibration data using a pre-trained MSFF-CNN model, and performing fault identification based on the extracted multi-scale features to determine the fault type of the post insulator to be tested specifically includes:

[0055] Performing feature extraction on the vibration data received by the input layer through the convolution module to obtain initial fault features and output them to the multi-scale convolution module;

[0056] Performing multi-scale feature extraction on the initial fault feature through the multi-scale convolution module to obtain the initial multi-scale feature and output it to the hole convolution module;

[0057] Perform feature extraction of different receptive fields on the initial multi-scale features through the dilated convolution module to obtain the multi-scale features and output them to the global average pooling layer;

[0058] The activation values ​​of the multi-scale features are averaged through the global average pooling layer to generate a global feature vector and output it to the output layer;

[0059] The probability of each fault type corresponding to the global feature vector is obtained by using the Softmax function through the output layer, and the fault type with the highest probability is used as the fault type of the post insulator to be tested.

[0060] It is worth noting that traditional convolutional neural networks usually use multiple fully connected layers to perform feature dimensionality reduction. However, since the fully connected layer adopts a fully connected method, a large number of parameters will be generated during the training process, resulting in high model calculation complexity and prone to overfitting problems. The global average pooling layer associates the fault type with the feature map, so that each channel has actual category meaning. Each feature map can be regarded as a confidence map of the fault category, and there are no parameters that need to be optimized inside it. While reducing the dimension, it greatly reduces the parameters of the network model and integrates the global feature information, which can effectively avoid overfitting. This embodiment uses the Softmax function to obtain the corresponding probability of each fault type based on the global feature vector related to the fault type output by the global average pooling layer, which can simplify the mapping process of features to fault types and improve the interpretability of the model.

[0061] As a preferred solution, the convolution module is constructed by a first convolution layer, a first batch normalization layer and a first maximum pooling layer connected in sequence.

[0062] As a preferred solution, the convolution kernel size of the first convolution layer is 8×8.

[0063] It is worth noting that if Figure 2As shown, it is a schematic diagram of the structure of the MSFF-CNN model in this embodiment. The convolution kernel size of the first convolution layer in this embodiment is preferably 8×8, so as to extract more comprehensive important fault features from the vibration data, and its expression is as follows: Among them, x i is the current input feature matrix, x i+1 is the calculated feature matrix, ω i is the convolution kernel weight parameter, b i is the bias parameter, Represents a convolution operation.

[0064] Furthermore, processing in combination with the first batch normalization layer can improve the training stability and convergence speed of the model. Subsequently, adding the first maximum pooling layer with a convolution kernel size of 2×2 and a step size of 1 helps to reduce the spatial dimension of the feature map while retaining key feature information. It is worth noting that the pooling layer is generally used to reduce the dimension of the features extracted by the convolution layer, increase the model's anti-distortion ability, and can effectively reduce the feature space and network parameters after convolution. There are two common pooling operations in convolutional neural networks, namely maximum pooling and average pooling. The present invention uses a maximum pooling layer to perform a maximum pooling operation on the feature information, reduce the data dimension while reducing the parameters, and select the largest statistic to obtain the local features of the area.

[0065] As a preferred solution, the multi-scale convolution module is constructed by sequentially connected multi-scale convolution layers, first residual connection layers, second batch normalization layers and second maximum pooling layers; the multi-scale convolution layers include several second convolution layers with different convolution kernel sizes.

[0066] As a preferred solution, the multi-scale convolution layer includes three second convolution layers, and the convolution kernel sizes of the three second convolution layers are 3×3, 6×6 and 9×9 respectively.

[0067] Specifically, this embodiment uses several parallel second convolution layers with different convolution kernel sizes to form a multi-scale convolution layer, so that features of different scales can be extracted, such as Figure 3 As shown in the figure, the features of different scales output by the multi-scale convolutional layer are fused through the first residual connection layer, and then processed by the second batch normalization layer to improve the expressiveness of the features. Finally, the second maximum pooling layer with a convolution kernel size of 3×3 and a stride of 2 is used to further reduce the spatial size of the feature map and reduce the computational complexity.

[0068] It is worth mentioning that residual connection solves the problem of gradient vanishing or gradient exploding that often occurs during training as the network depth increases. Through the skip connection of the shortcut path, the network can directly pass the input data to the subsequent layers of the network. Figure 5As shown, this connection forms an identity mapping that allows the network to learn the residual between the input and output. Figure 5 The residual block in mainly includes the main path and the shortcut path. The main path includes several convolutional layers, activation functions, etc. for extracting features; the shortcut path directly passes the input data to the output of the residual block. The output of the residual block is the element-by-element addition of the output of the main path and the output of the shortcut path. The operation process is as follows: H(x) = F(x) + x; where x is the input data, F(x) is the feature after processing by the main path, and H(x) is the final output of the residual block.

[0069] Batch normalization is a method used for data normalization. The process of batch normalization for any output data in training is as follows:

[0070] Calculate the average of a batch of data:

[0071]

[0072] Calculate the variance of batch data:

[0073]

[0074] Normalized data:

[0075]

[0076] Scaling and offsetting:

[0077]

[0078] Where m represents the batch size; ε is a constant term that ensures numerical stability; γ and β are scale factors and offset factors that can be learned by the network, respectively; is the output of the batch normalization layer. Through batch normalization, the output data of each layer can maintain a normal distribution, thereby accelerating convergence and reducing the offset of internal covariates during training, improving the generalization ability and training speed of the model.

[0079] As a preferred solution, the dilated convolution module is constructed by sequentially connected dilated convolution units, a second residual connection layer, a third batch normalization layer and a third maximum pooling layer; the dilated convolution unit includes several dilated convolution layers with different expansion rates.

[0080] As a preferred solution, the dilated convolution unit includes three dilated convolution layers, and the expansion rates of the three dilated convolution layers are 1, 3 and 5 respectively.

[0081] Specifically, dilated convolution is also called dilated convolution. It is a convolutional neural network method that expands the receptive field of the output unit without increasing the number of parameters. It is implemented by inserting holes between adjacent elements of the convolution kernel, which is equivalent to adding zero elements between these adjacent elements. The number of inserted holes is called the dilation rate, denoted as DR. Taking a 3×3 convolution kernel as an example, dilated convolutions with different dilation rates are as follows: Figure 4 As shown. It can be seen that the dilated convolution can expand the receptive field of the output unit, thereby enhancing the network's ability to extract features of different scales. This embodiment preferably uses a two-dimensional dilated convolution layer with expansion rates of 1, 3, and 5 to construct a dilated convolution unit, so as to extract features of different receptive fields in parallel, and then fuse the features through the second residual connection layer, followed by processing through the third batch normalization layer to improve the expressiveness of the features, and finally through the third maximum pooling layer with a convolution kernel size of 3×3 and a step size of 2, to further reduce the spatial size of the feature map and reduce the computational complexity.

[0082] As one of the preferred embodiments, the parameter settings of the MSFF-CNN model are shown in Table 1 below:

[0083] Table 1 Parameter settings of MSFF-CNN model

[0084]

[0085]

[0086] As a preferred solution, the method specifically trains the MSFF-CNN model through the following steps:

[0087] Collecting several post insulator fault vibration data and dividing them according to the preset single sample data volume to obtain the fault vibration data set;

[0088] Adding labels to the fault vibration data set to obtain a training data set; wherein the labels include normal data labels, upper flange fault data labels, and lower flange fault data labels;

[0089] The MSFF-CNN model is trained using the training data set to optimize the parameters of the MSFF-CNN model.

[0090] Specifically, this embodiment uses an SCT vibration acoustic detection device to collect vibration data of various faulty post insulators. The sampling frequency of the detection device is 50kHz, and the single sampling time is about 0.5s. Experimental tests have shown that the two probes of the detection device are against the lower flange of the post insulator, which can better obtain the vibration signal generated by the exciter and the flange surface. The data collected above under the normal state, upper flange fault and lower flange fault conditions of the post insulator are selected to construct a data set. Further, the fault vibration data set is obtained by dividing according to the preset single sample data volume. Exemplarily, the single sample data volume is 1024 sampling point data.

[0091] Furthermore, labels are added to the acquired fault vibration dataset and converted into One-hot encoding. Then, the Batch Size is set to 256, the Epoch is set to 300, the Adam optimizer is selected as the optimizer of the model, the learning rate is set to 0.001, and the cross entropy loss function is selected as the loss function of the model. The model training goal is to minimize the loss function, and the MSFF-CNN model is trained to tune the parameters of the MSFF-CNN model.

[0092] In order to verify the performance of the MSFF-CNN model in the embodiment of the present invention, a series of verification comparisons are used to illustrate it.

[0093] Firstly, several 110KV in-service porcelain post insulators in different regions were selected as research samples, and an experimental platform was built. The experimental platform consists of signal analysis software, insulator test bench, function arbitrary drum generator, experimental prototype and signal acquisition instrument. The signal acquisition instrument uses SCT vibration acoustic detection device to collect the vibration signal of porcelain post insulator. The sampling frequency of the instrument is 50kHz, and the single sampling time is about 0.5s. 256,000 sampling points are taken for each type of data, and each type of data is divided into 250 groups, each group is used as 1 sample, and each sample has 1024 sampling points, such as Figure 6 As shown in the figure, it is the overall time domain diagram of the constructed data set. Figure 7 As shown in the figure, it is a time domain diagram of a single sample of the constructed data set, and its label settings are shown in Table 2 below:

[0094] Table 2 Fault types and label numbers

[0095]

[0096] Divide it into training group and test group in the ratio of 8:2, of which 200 groups are put into training and the remaining 50 groups are tested. Batch Size is set to 256, Epoch is set to 300, Adam optimizer is selected as the optimizer of the model, learning rate is set to 0.001, cross entropy loss function is selected as the loss function of the model, and the model training goal is to minimize the validation set loss function.

[0097] The training and test accuracy curves are as follows Figure 8 As shown, the loss curve is Fig. 9 As shown in the figure. The experimental results show that the accuracy curve of the training set has completely converged after 10 iterations, and the accuracy reaches 100%. The loss curve decreases rapidly with the iteration and completely converges after 10 iterations, and the loss is infinitely close to 0. The accuracy curve of the test set has completely converged after 138 iterations, reaching 99.98%. The loss curve decreases rapidly with the iteration and completely converges after 150 iterations, and is infinitely close to 0.

[0098] The confusion matrix of the test set is as follows Fig.10 As shown in the figure, the horizontal axis is the predicted label and the vertical axis is the actual label. The main diagonal reaches 50, which is equal to the number of test sets for each data, and the recognition accuracy reaches up to 100%.

[0099] Fig.11 and Fig.12 The following are KPCA dimension reduction visualization diagrams of the distribution of the original data and the data after training and classification. By comparison, it can be seen that the distribution of the original data is relatively scattered, and various data are mixed together. As the training process progresses, different types of data points gradually disperse, while the same type of data points gradually gather and finally separate completely, achieving excellent classification results. It is proved that the MSFF-CNN model of the embodiment of the present invention has excellent classification performance for different fault data of porcelain post insulators under noise interference conditions.

[0100] To verify the superiority of the MSFF-CNN model of the embodiment of the present invention, it is compared with AlexNet, GhostNet and MobileNetV2. During the training process, the batch size is uniformly set to 256 and the number of iterations is 300. The final test accuracy and loss curve are visualized, as shown in Figure 2. Fig.13 and Fig.14As shown. It can be seen from the results that the MSFF-CNN model of the embodiment of the present invention has the best effect. The accuracy curve and the loss curve converge the fastest among various models. The accuracy reaches 99.98%, the highest among various networks, and the loss is the lowest among various models, infinitely close to 0. In contrast, the effect of directly using AlexNet, GhostNet and MobileNetV2 for diagnosis is poor. The test accuracy of the three models can only be maintained at around 95% at most, and the test loss is also much higher than the MSFF-CNN model of the embodiment of the present invention. And when the batch size and the number of iterations are the same, the test curves of AlexNet, GhostNet and MobileNetV2 fluctuate greatly, it is difficult to converge effectively, and the effect is poor.

[0101] In order to verify the robustness of the MSFF-CNN model of the embodiment of the present invention, new data of the same data type are selected to construct a new test data set. 102400 sampling points are taken for each type of data, and each type of data is divided into 100 samples, each sample includes 1024 sampling points, and its label setting is the same as that in Table 2 above. The models trained to convergence of MSFF-CNN, AlexNet, GhostNet and MobileNetV2 are saved in turn, and 5 experiments are carried out respectively using the new test data obtained.

[0102] The confusion matrix of the new test set is as follows Fig.15 As shown, the main diagonal basically reaches 100, which is equal to the number of test samples for each data, and the recognition accuracy rate reaches up to 100%.

[0103] The overall average accuracy of the test results of different models each time is shown in Table 3. As can be seen from Table 3, the test accuracy of the MSFF-CNN model is stable at more than 99%, the test accuracy of the other three models is unstable, the diagnostic effect is poor, and the highest test accuracy is 96.62%, which is much lower than the MSFF-CNN model of the embodiment of the present invention.

[0104] Table 3 Test accuracy of each model

[0105]

[0106] In order to verify the fault recognition effect of the MSFF-CNN model under noise conditions, Gaussian white noise with signal-to-noise ratios of -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, and 6dB was added to the new test data set. Fig.16 , Fig.17 and Fig.18As shown in Table 4. As the noise signal-to-noise ratio decreases, the insulator vibration signal is gradually covered by noise and is almost distorted below -4dB. The MSFF-CNN, AlexNet, GhostNet and MobileNetV2 models trained on data without adding noise were used for testing. Five test experiments were conducted using test data with different noises. The average test accuracy under each noise is the average of the accuracy of the five experiments, as shown in Table 4. As can be seen from Table 4, the average accuracy of the MSFF-CNN model is higher than that of other models in both the absence of noise interference and the presence of noise interference, which fully verifies that the model has excellent robustness.

[0107] Table 4 Test accuracy under different noise data

[0108]

[0109] In order to verify the rationality of each component in the MSFF-CNN model of the embodiment of the present invention and explore their role in the entire model, the following 9 models are set for comparative experiments. The specific settings are as follows:

[0110] Model 1: MSFF-CNN model of an embodiment of the present invention.

[0111] Model 2: The average pooling layer is used to replace the maximum pooling layer of the MSFF-CNN model.

[0112] Model 3: Remove the convolutional module of the MSFF-CNN model.

[0113] Model 4: Use conventional convolutional layers to replace the multi-scale convolution module of the MSFF-CNN model.

[0114] Model 5: Use conventional convolutional layers to replace the dilated convolution module of the MSFF-CNN model.

[0115] Model 6: Use the conventional concatenated feature fusion layer to replace the residual connection layer of the MSFF-CNN model.

[0116] Model 7: The convolution kernel sizes of the multi-scale convolutional layer of the MSFF-CNN model are set to (3×3), (7×7), and (12×12) respectively.

[0117] Model 8: The dilation rates of the atrous convolutional layers of the MSFF-CNN model are set to DR=1, DR=2, and DR=3 respectively.

[0118] Model 9: Use the fully connected layer to replace the global average pooling layer of the MSFF-CNN model.

[0119] The accuracy results of each model after 5 tests are as follows: Fig.19 As shown. Fig.19 It can be seen that the accuracy of the MSFF-CNN model is higher than that of the other eight models, which fully verifies the rationality and superiority of each component of the MSFF-CNN model in the embodiment of the present invention.

[0120] In order to verify the generalization of the MSFF-CNN model in the embodiment of the present invention, the above data sets are divided into training sets and test sets according to the ratio of 7:3 and 6:4 respectively. The Batch Size is also set to 256 and the Epoch is set to 300. Finally, the fault recognition accuracy of each model is Fig. 20 As shown. Fig. 20 It can be seen that with the change of the data set division ratio, the fault recognition accuracy of the MSFF-CNN model in the embodiment of the present invention remains stable at 99%, while the fault recognition accuracy of the AlexNet, GhostNet and MobileNetV2 models fluctuates greatly compared with when the data set is divided according to 8:2. This once again verifies that the MSFF-CNN model in the embodiment of the present invention has excellent effect on fault identification of porcelain post insulators under noise interference conditions.

[0121] Therefore, the post insulator fault identification method based on MSFF-CNN provided in an embodiment of the present invention can improve the feature extraction capability of the vibration data of the post insulator through the convolution module, multi-scale convolution module and void convolution module in the MSFF-CNN model, thereby effectively improving the feature extraction accuracy of the vibration data. Fault identification based on the extracted multi-scale features can effectively improve the classification accuracy and robustness of the fault type.

[0122] See also Fig.21 A second aspect of an embodiment of the present invention provides a post insulator fault identification system based on MSFF-CNN, comprising:

[0123] The data acquisition module 101 is used to acquire vibration data of the post insulator to be tested;

[0124] The fault identification module 102 is used to use a pre-trained MSFF-CNN model to extract multi-scale features from the vibration data, and to perform fault identification based on the extracted multi-scale features to determine the fault type of the post insulator to be tested; wherein the MSFF-CNN model includes an input layer, a convolution module, a multi-scale convolution module, a hole convolution module, a global average pooling layer and an output layer.

[0125] As a preferred solution, the fault identification module 102 is used to extract multi-scale features from the vibration data using a pre-trained MSFF-CNN model, and perform fault identification based on the extracted multi-scale features to determine the fault type of the post insulator to be tested, specifically including:

[0126] Performing feature extraction on the vibration data received by the input layer through the convolution module to obtain initial fault features and output them to the multi-scale convolution module;

[0127] Performing multi-scale feature extraction on the initial fault feature through the multi-scale convolution module to obtain the initial multi-scale feature and output it to the hole convolution module;

[0128] Perform feature extraction of different receptive fields on the initial multi-scale features through the dilated convolution module to obtain the multi-scale features and output them to the global average pooling layer;

[0129] The activation values ​​of the multi-scale features are averaged through the global average pooling layer to generate a global feature vector and output it to the output layer;

[0130] The probability of each fault type corresponding to the global feature vector is obtained by using the Softmax function through the output layer, and the fault type with the highest probability is used as the fault type of the post insulator to be tested.

[0131] As a preferred solution, the convolution module is constructed by a first convolution layer, a first batch normalization layer and a first maximum pooling layer connected in sequence.

[0132] As a preferred solution, the convolution kernel size of the first convolution layer is 8×8.

[0133] As a preferred solution, the multi-scale convolution module is constructed by sequentially connected multi-scale convolution layers, first residual connection layers, second batch normalization layers and second maximum pooling layers; the multi-scale convolution layers include several second convolution layers with different convolution kernel sizes.

[0134] As a preferred solution, the multi-scale convolution layer includes three second convolution layers, and the convolution kernel sizes of the three second convolution layers are 3×3, 6×6 and 9×9 respectively.

[0135] As a preferred solution, the dilated convolution module is constructed by sequentially connected dilated convolution units, a second residual connection layer, a third batch normalization layer and a third maximum pooling layer; the dilated convolution unit includes several dilated convolution layers with different expansion rates.

[0136] As a preferred solution, the dilated convolution unit includes three dilated convolution layers, and the expansion rates of the three dilated convolution layers are 1, 3 and 5 respectively.

[0137] As a preferred solution, the system further includes a model training module for:

[0138] Collecting several post insulator fault vibration data and dividing them according to the preset single sample data volume to obtain the fault vibration data set;

[0139] Adding labels to the fault vibration data set to obtain a training data set; wherein the labels include normal data labels, upper flange fault data labels, and lower flange fault data labels;

[0140] The MSFF-CNN model is trained using the training data set to optimize the parameters of the MSFF-CNN model.

[0141] The post insulator fault identification system based on MSFF-CNN provided by the embodiment of the present invention can improve the feature extraction capability of the vibration data of the post insulator through the convolution module, multi-scale convolution module and void convolution module in the MSFF-CNN model, thereby effectively improving the feature extraction accuracy of the vibration data. Fault identification based on the extracted multi-scale features can effectively improve the classification accuracy and robustness of the fault type.

[0142] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A post insulator fault identification method based on MSFF-CNN, characterized in that: include: Obtain vibration data of the post insulator to be tested; The vibration data is subjected to multi-scale feature extraction using a pre-trained MSFF-CNN model, and fault identification is performed based on the extracted multi-scale features to determine the fault type of the post insulator to be tested; wherein the MSFF-CNN model includes an input layer, a convolution module, a multi-scale convolution module, a hole convolution module, a global average pooling layer and an output layer.

2. The post insulator fault identification method based on MSFF-CNN according to claim 1, characterized in that: The method of extracting multi-scale features from the vibration data using a pre-trained MSFF-CNN model, and performing fault identification based on the extracted multi-scale features to determine the fault type of the post insulator to be tested specifically includes: Performing feature extraction on the vibration data received by the input layer through the convolution module to obtain initial fault features and output them to the multi-scale convolution module; Performing multi-scale feature extraction on the initial fault feature through the multi-scale convolution module to obtain the initial multi-scale feature and output it to the hole convolution module; Perform feature extraction of different receptive fields on the initial multi-scale features through the dilated convolution module to obtain the multi-scale features and output them to the global average pooling layer; The activation values ​​of the multi-scale features are averaged through the global average pooling layer to generate a global feature vector and output it to the output layer; The probability of each fault type corresponding to the global feature vector is obtained by using the Softmax function through the output layer, and the fault type with the highest probability is used as the fault type of the post insulator to be tested.

3. The post insulator fault identification method based on MSFF-CNN according to claim 1, characterized in that: The convolution module is constructed by sequentially connecting a first convolution layer, a first batch normalization layer, and a first maximum pooling layer.

4. The post insulator fault identification method based on MSFF-CNN according to claim 3 is characterized in that: The convolution kernel size of the first convolution layer is 8×8.

5. The post insulator fault identification method based on MSFF-CNN according to claim 1, characterized in that: The multi-scale convolution module is constructed by sequentially connected multi-scale convolution layers, a first residual connection layer, a second batch normalization layer and a second maximum pooling layer; the multi-scale convolution layer includes a plurality of second convolution layers with different convolution kernel sizes.

6. The post insulator fault identification method based on MSFF-CNN according to claim 5, characterized in that: The multi-scale convolution layer includes three second convolution layers, and the convolution kernel sizes of the three second convolution layers are 3×3, 6×6 and 9×9 respectively.

7. The post insulator fault identification method based on MSFF-CNN according to claim 1, characterized in that: The dilated convolution module is constructed by sequentially connected dilated convolution units, a second residual connection layer, a third batch normalization layer and a third maximum pooling layer; the dilated convolution unit includes several dilated convolution layers with different expansion rates.

8. The post insulator fault identification method based on MSFF-CNN according to claim 7, characterized in that: The atrous convolution unit includes three atrous convolution layers, and the expansion rates of the three atrous convolution layers are 1, 3, and 5, respectively.

9. The post insulator fault identification method based on MSFF-CNN according to claim 1, characterized in that: The method specifically trains the MSFF-CNN model through the following steps: Collecting several post insulator fault vibration data and dividing them according to the preset single sample data volume to obtain the fault vibration data set; Adding labels to the fault vibration data set to obtain a training data set; wherein the labels include normal data labels, upper flange fault data labels, and lower flange fault data labels; The MSFF-CNN model is trained using the training data set to optimize the parameters of the MSFF-CNN model.

10. A post insulator fault identification system based on MSFF-CNN, characterized in that: include: A data acquisition module, used to acquire vibration data of the post insulator to be tested; A fault identification module is used to extract multi-scale features from the vibration data using a pre-trained MSFF-CNN model, and to perform fault identification based on the extracted multi-scale features to determine the fault type of the post insulator to be tested; wherein the MSFF-CNN model includes an input layer, a convolution module, a multi-scale convolution module, a hole convolution module, a global average pooling layer and an output layer.

Images