Isolation switch opening and closing state detection method

Abstract

The invention discloses an isolating switch opening and closing state detection method, which comprises the following steps of: 1, acquiring a plurality of sample pictures of a transformer substation to obtain a data set, and dividing the data set to obtain a training set; in addition, an improved InceptionV3 model is constructed; 2, training the improved model by adopting training set data, and replacing a full connection layer, a Softmax layer and an output layer of the trained improved model by adopting a full connection layer, a Softmax layer and an output layer in an untrained improved model; then, performing fine tuning training on the replaced improved model to obtain an image classification network; and step 3, inputting a to-be-detected transformer substation picture into the image classification network to obtain an isolation switch opening and closing state identification result. The method has the advantages of being high in speed, better in recognition effect and higher in accuracy when detecting the opening and closing states of the disconnecting switch.

Description

Technical Field

[0001] The invention relates to the field of disconnection switch opening and closing detection methods, in particular to a disconnection switch opening and closing state detection method. Background Art

[0002] Once various types of high-voltage equipment in the power system fail, it may cause serious consequences and bring huge economic losses. Isolating switches are the most commonly used power equipment in substations. Their main function is to ensure safe isolation of the circuit when repairing other power transmission and transformation equipment. They play an important role in the power system. One of the most common abnormal conditions is that the isolating switch is not in place. The inadequate opening and closing of the isolating switch can easily lead to contact heating, arc discharge when operating adjacent isolating switches, and other faults or accidents. Therefore, correctly detecting the opening and closing status of the isolating switch plays an important role in the stable operation of the power system.

[0003] The reasons for the disconnector not opening and closing properly can be divided into two types: the disconnector's own fault and the external environment. Among them, the disconnector's own fault is mainly caused by product production quality, residual defects in the installation process, or aging of the transmission mechanism after long-term operation of the equipment, aging and sticking of the lubricant, etc.; the external environment is mainly caused by foreign objects such as plastic bags, kite lines, and bird nests hanging on the high-voltage end of the disconnector. During the operation of the disconnector, the high-voltage arm cannot move according to the correct trajectory, which eventually leads to the disconnector not opening and closing properly.

[0004] For a long time, regular inspections have been the main way to troubleshoot power system faults. Traditional manual inspections have certain problems in terms of inspection cycles, efficiency, and danger. Substation operators mainly conduct on-site inspections and use their naked eyes to observe the opening and closing of disconnectors after operation. Obviously, this method is highly subjective and is easily affected by the experience, working status, and even air visibility of operators, resulting in missed judgments and misjudgments, which poses a great safety hazard to the safe operation of the power system. With the introduction of the concept of digital power grids and the development of new artificial intelligence technologies, it is a general trend to realize intelligent inspections. This move is of great significance for monitoring the operation of disconnectors, improving the reliability of their transmission, ensuring the operation of the primary line of the substation, and the reliability and safety of maintenance. Summary of the invention

[0005] In order to solve the deficiencies in the prior art, the present invention proposes a method for detecting the opening and closing states of an isolating switch based on an improved InceptionV3 model.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for detecting the opening and closing state of an isolating switch comprises the following steps:

[0008] Step 1: obtain multiple sample images of the substation at different times, different light intensities, and different backgrounds, mark the disconnection and closing states of the disconnector in each sample image, thereby obtaining a data set, and divide the data set to obtain a training set;

[0009] In addition, an improved InceptionV3 model is constructed, wherein the improved InceptionV3 model includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first pooling layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, an Inception module a, an Inception module b, an Inception module c, a second pooling layer, a fully connected layer, a Softmax layer, and an output layer; wherein the first convolutional layer is composed of a 3×3 convolutional kernel with a step size of 2, the second convolutional layer is composed of a 3×3 convolutional kernel with a step size of 1, the third convolutional layer is composed of a 3×3 convolutional kernel with a step size of 1, and the pooling layer is composed of a 3×3 convolutional kernel with a step size of 2. The fourth convolution layer is composed of a 3×3 convolution kernel with a stride of 1, the fifth convolution layer is composed of a 3×3 convolution kernel with a stride of 2, and the sixth convolution layer is composed of a 3×3 convolution kernel with a stride of 1. The Inception module a is an Inception structure of two consecutive 3×3 convolution kernels, the Inception module b is an Inception structure that replaces the n×n convolution kernel with consecutive 1×n and n×1 convolution kernels, and the Inception module c is an Inception structure that replaces the 3×3 convolution kernel with parallel 1×3 and 3×1 convolution kernels. The second pooling layer is composed of 8×8 convolution kernels;

[0010] Step 2: Use the training set data to train the improved InceptionV3 model to obtain a trained improved InceptionV3 model;

[0011] Next, the fully connected layer, Softmax layer, and output layer of the untrained improved InceptionV3 model are used to replace the fully connected layer, Softmax layer, and output layer of the trained improved InceptionV3 model, and the number of categories of the replaced fully connected layer is adjusted;

[0012] Then, the Fine Tuning method is used to fine-tune the replaced improved InceptionV3 model to obtain an image classification network;

[0013] Step 3: Input the substation image to be detected into the image classification network obtained in step 2, and the image classification network outputs the recognition result of the disconnector opening and closing status.

[0014] Furthermore, in step 1, data enhancement processing is performed on the sample images to expand the number of sample images, and then a data set is constructed using the sample images with the expanded number.

[0015] Furthermore, in step 1, the optimizer used in the constructed improved InceptionV3 model is the stochastic steepest descent method.

[0016] Furthermore, in step 1, the LeakyReLU activation function is used in the Softmax layer of the improved InceptionV3 model constructed.

[0017] Furthermore, in step 2, the number of categories of the replaced fully connected layer is adjusted from 1000 to 2 categories.

[0018] The present invention proposes a disconnector opening and closing state detection method based on an improved InceptionV3 model. For the original InceptionV3 model, the random steepest descent method is used to replace the original steepest descent method as the optimizer of the InceptionV3 model; the Leaky ReLU function is used to replace the original ReLU function after the InceptionV3 full connection layer as a new activation function, thereby establishing an improved InceptionV3 model. Then, a new training network is constructed by applying transfer learning to the number of network layers on the improved InceptionV3 model, all layers except the last three layers are taken out from the pre-trained network, and then the last three layers are replaced by a fully connected layer, a Softmax layer and a classification layer, and the number of classifications of the fully connected layer is adjusted from 1000 to 2 categories, and then the network is fine-tuned and trained using the Fine Tuning method, and finally an image classification network is obtained. The improved model speeds up the speed of image detection. Therefore, the method of the present invention has the advantages of fast speed, better recognition effect and higher accuracy when detecting the disconnector opening and closing state. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of this embodiment.

[0020] Figure 2 : is the Inception module diagram in this embodiment, where (a) is Inception module a, (b) is Inception module b, and (c) is Inception module c.

[0021] Figure 3 This is a flow chart of data preprocessing in this embodiment.

[0022] Figure 4 This is an example diagram of the data enhancement effect in this embodiment.

[0023] Figure 5 This is a structural diagram of the improved InceptionV3 model in this embodiment.

[0024] Figure 6 This is a curve chart of the loss value change during the training process.

[0025] Figure 7 This is a curve chart of the accuracy value change during the training process.

[0026] Figure 8 ROC curves of four models.

[0027] Fig. 9 The following are the test result diagrams of four models, among which (a) is the result diagram of VGG16, (b) is the result diagram of ResNet50, (c) is the result diagram of InceptionV3, and (d) is the result diagram of InceptionV3 improved in this embodiment. DETAILED DESCRIPTION

[0028] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0029] like Figure 1 As shown, this embodiment discloses a method for detecting the opening and closing state of an isolating switch, comprising the following steps:

[0030] Step 1: Obtain multiple sample images of the substation at different times, different light intensities, and different backgrounds to construct a data set, and build an improved InceptionV3 model.

[0031] In this embodiment, sample images of different time, different light intensity and different background in the substation are obtained. The sample images are recorded by a 1080P high-definition camera and obtained by video screenshots. After screening, a total of 1500 original images are retained.

[0032] For each sample image, two types of targets, the disconnector open and closed, are labeled. During the labeling process, the disconnector open state is labeled as "open" and the closed state is labeled as "closed". After the labeling is completed, the label files corresponding to the images are obtained. These label files will be used for subsequent model training and target detection.

[0033] Since training the neural network model requires a sufficient amount of data, and the imbalance of the open and closed samples in the data set will also affect the recognition effect of the final model. Therefore, this embodiment uses data enhancement to expand the number of sample images. Specifically, data enhancement methods such as horizontal flipping, random shearing, rotation angle, adding noise, and changing the image brightness are used to expand the sample images and balance the positive and negative samples. The sample images after data enhancement can simulate complex situations such as changes in shooting angle, changes in light intensity, and image blur. It can not only increase the sample size and effectively avoid overfitting, but also enhance the generalization and robustness of the model. After data enhancement, there are a total of 3662 pictures, including 1857 closed pictures and 1805 open pictures. The experimental data preprocessing process is as follows. Figure 3 As shown, examples of data enhancement effects are shown in Figure 4 shown.

[0034] Furthermore, in order to meet the input requirements of the improved InceptionV3 model, this embodiment normalizes the images of different sizes in each sample image after data enhancement and adjusts them to 299×299 pixels, thereby constructing a data set.

[0035] Then, this embodiment uses a random selection method to divide the 3662 sample images in the data set into a training set, a validation set, and a test set. 80% of the data set is divided into one group (for training and validation, of which 90% is used for training and 10% is used for validation), and 20% is divided into another group (for testing). Ensure that the number of the two categories in each group is balanced, and the specific division is shown in Table 1.

[0036] Table 1 Dataset division

[0037]

[0038] like Figure 5 As shown, in this embodiment, the constructed improved InceptionV3 model includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first pooling layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, an Inception module a, an Inception module b, an Inception module c, a second pooling layer, a fully connected layer, a Softmax layer, and an output layer; wherein the first convolutional layer is composed of a 3×3 convolutional kernel with a step size of 2, the second convolutional layer is composed of a 3×3 convolutional kernel with a step size of 1, the third convolutional layer is composed of a 3×3 convolutional kernel with a step size of 1, and the pooling layer is composed of a 3×3 convolutional kernel with a step size of 2. The fourth convolution layer is composed of a 3×3 convolution kernel with a stride of 1, the fifth convolution layer is composed of a 3×3 convolution kernel with a stride of 2, and the sixth convolution layer is composed of a 3×3 convolution kernel with a stride of 1. The Inception module a is an Inception structure of two consecutive 3×3 convolution kernels, the Inception module b is an Inception structure that replaces the n×n convolution kernel with consecutive 1×n and n×1 convolution kernels, and the Inception module c is an Inception structure that replaces the 3×3 convolution kernel with parallel 1×3 and 3×1 convolution kernels. The second pooling layer is composed of 8×8 convolution kernels.

[0039] The biggest feature of the InceptionV3 model is the use of the Inception module, which aims to design a network with excellent local topological structure, that is, using convolution kernels of different sizes to extract different abstract features in the same area, and then fuse the extracted features.

[0040] The improved InceptionV3 model in this embodiment is based on the existing Inception model. In order to reduce the number of calculation parameters of the model and improve the calculation speed, the 5×5 convolution kernel is decomposed into two 3×3 convolution kernels, and the n×n convolution kernel is decomposed into two convolution kernels of 1×n and n×1. On this basis, the BatchNormalization auxiliary classifier is used to speed up the classification speed and smooth the labels to prevent overfitting.

[0041] Specifically, this embodiment improves the Inception module a in the InceptionV3 model as follows Figure 2 As shown in (a), it is the Inception structure of two consecutive 3×3 convolution kernels. Inception module b is as follows Figure 2 As shown in (b), the Inception structure uses continuous 1×n and n×1 convolution kernels instead of n×n convolution kernels. Figure 2 As shown in (c), the Inception structure uses parallel 1×3 and 3×1 convolution kernels instead of 3×3 convolution kernels. Obviously, in the improved InceptionV3 model of this embodiment, the Inception module not only improves the architecture to include three different structures, but also draws on the essence of VGGNet, using small-sized convolution kernels instead of large-sized convolution kernels to reduce the amount of calculation; at the same time, the increase in nonlinear mapping also expands the model's expression ability. The specific network structure of the improved InceptionV3 model of this embodiment is shown in Table 2. In the table, Inception module a, module b, and module c correspond to Figure 1 The three subgraphs in Table 2 are as follows:

[0042] Table 2 InceptionV3 network structure

[0043]

[0044]

[0045] The optimizer used in traditional convolutional neural networks is the steepest descent method. This embodiment improves the InceptionV3 model by using the random steepest descent method to optimize the original method. Assuming that the optimized function is f(x), find the optimal solution for the objective function on the training sample. The method of the steepest descent method to find the minimum value is to take a step along the derivative direction of the current position each time. At time t+1, the update iteration formula is shown in formula (1):

[0046] x t+1 =x t -η t ▽f(x t )(1)

[0047] Where x t is the position of step t, ▽f(x t ) is the derivative of the objective function at that position, η t is the step length.

[0048] The steepest descent method has two obvious disadvantages: first, when faced with very large data sets, the convergence time is too long and the operation speed is slow; second, it may fall into a saddle point or a relatively poor local optimum, at which point the algorithm cannot run because the derivative of the function is 0 at these points.

[0049] In large-scale optimization problems, the steepest descent method requires high computational complexity in each iteration, while the stochastic steepest descent method randomly takes a data sample to calculate the gradient direction at each iteration. The computational complexity of the stochastic steepest descent method is independent of the number of samples, which reduces the computational complexity. With the stochastic gradient descent of the stochastic steepest descent method, each weight update only uses one sample in the data set, that is, one epoch process has only one iteration and one data update. At time t+1, the update iterative calculation formula is shown in formula (2):

[0050] x t+1 =x t -η t g t (2)

[0051] Where gt is the random gradient value, satisfying E[g t ]=▽f(x t ), that is, the stochastic gradient gt is also the gradient ▽f(x t ) is an unbiased estimate of . In summary, the use of the random steepest descent method can solve the two defects of the steepest descent method. It is not only fast, but also can converge better. Therefore, the optimizer of the improved InceptionV3 model in this embodiment adopts the random steepest descent method.

[0052] Furthermore, in order to avoid the death of neurons when the input is negative, this embodiment uses the Leaky ReLU activation function as the activation function in the Softmax layer after the fully connected layer. As an extension of the ReLU function, the LeakyReLU activation function allows a small slope (generally 0.01) when the input is less than 0, so that the neurons can maintain a certain degree of activation when the input is negative, thereby avoiding the death of neurons and extracting more effective information. Its calculation formula is shown in formula (3):

[0053]

[0054] Where: f(x) is the optimized function, x is the input value. When x>0, the output is x; when x≤0, a<1, the gradient of the function is not 0, which can prevent the gradient value from tending to 0.

[0055] Step 2: Use the training set data to train the improved InceptionV3 model to obtain a trained improved InceptionV3 model.

[0056] In this embodiment, experiments have shown that the general feature migration effect obtained after training the first few layers of the neural network is better, while some specific features learned after the network becomes deeper are not suitable for migration; in addition, adding FineTuning to the deep migration network can better overcome the differences between data, and even perform better than the original network in the target task, and it is found that the migration of network layers can accelerate network learning and optimization.

[0057] Therefore, this embodiment applies transfer learning to the number of network layers on the improved InceptionV3 model and builds a new training network to classify the disconnector opening and closing images. By adjusting the structure of the improved InceptionV3 model predefined in MATLAB, the last three layers (i.e., fully connected layer, Softmax layer, output layer) of the pre-trained InceptionV3 model were originally used to identify objects of different categories. For the new image classification problem, these three layers must be adjusted.

[0058] Specifically, this embodiment takes out the last three layers, namely, the fully connected layer, the Softmax layer, and the output layer, from the trained improved InceptionV3 model, and uses the fully connected layer, the Softmax layer, and the output layer in the untrained improved InceptionV3 model to replace the fully connected layer, the Softmax layer, and the output layer of the trained improved InceptionV3 model, so as to migrate the originally trained network layer to the new classification task. In addition, this embodiment sets the parameters of the new fully connected layer according to the training data, and adjusts the number of classifications of the fully connected layer from 1000 to 2. Then, the Fine Tuning method is used to fine-tune the replaced improved InceptionV3 model, thereby obtaining an image classification network.

[0059] Step 3: Input the substation image to be detected into the image classification network obtained in step 2, and the image classification network outputs the recognition result of the disconnector opening and closing status.

[0060] The following is an experimental verification of the performance improvement of the InceptionV3 model in this embodiment

[0061] In order to evaluate the performance of the model as a disconnector opening and closing image classifier, this embodiment selects the AUC value, confusion matrix and prediction accuracy as evaluation indicators. AUC (Area Under Curve) is the area under the ROC curve (Receiver Operating Characteristic Curve), and its value range is between 0.5 and 1. Accuracy indicates the proportion of the number of samples predicted to be correctly classified to the total number of samples, and the calculation formula is shown in formula (4):

[0062]

[0063] Where: TP is the number of positive samples predicted as positive samples; FN is the number of positive samples predicted as negative samples; FP is the number of negative samples predicted as positive samples; TN is the number of negative samples predicted as negative samples.

[0064] Figure 6 , Figure 7 The changes in the loss value and accuracy of the improved InceptionV3 model proposed in this embodiment during the training process of 164 rounds are respectively shown. As the number of training rounds increases, the loss value gradually decreases and tends to be stable, and the accuracy gradually increases and tends to be stable. In order to verify the performance of the improved model in this article, the improved InceptionV3 model proposed in this embodiment is compared with the VGG16, ResNet50 and InceptionV3 models in the same data set and the same environment. The AUC value, confusion matrix and prediction accuracy of each model are calculated as evaluation indicators to test its performance. The ROC curve is obtained by adjusting the threshold of the model prediction, and then the value of each point is calculated after the threshold is automatically set by the perfcurve function, and finally these points are connected into a curve. Figure 8 The ROC curves of the four models and their corresponding AUC values ​​are shown in the figure. It can be seen from the figure that the AUC value of the improved model proposed in this paper is the largest.

[0065] Fig. 9 The confusion matrix results of the VGG16, ResNet50, InceptionV3 and the improved InceptionV3 model test set proposed in this embodiment are shown, where the horizontal labels represent the predicted results output by the model, the vertical labels represent the actual results, and the values ​​represent the prediction accuracy and the number of samples. The prediction accuracy of each model and the time consumption for training and recognition are shown in Table 3:

[0066] Table 3 Time consumption and prediction accuracy of different models

[0067]

[0068] From the perspective of numerical values, it can be seen from Table 3 that the training time consumption of the improved InceptionV3 model proposed in this embodiment is reduced by 53.8%, 50.0% and 33.3% respectively compared with the other three models; the recognition time consumption is reduced by 36.4%, 33.6% and 14.0% respectively compared with the other three models; the prediction accuracy is increased by 9.4%, 8.9% and 2.7% respectively compared with the other three models, indicating that the improved InceptionV3 model proposed in this embodiment can effectively identify the opening and closing states of the disconnector with high accuracy.

[0069] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0070] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.

Claims

1. A method for detecting the opening and closing state of an isolating switch, characterized in that: The following steps are involved: Step 1: obtain multiple sample images of the substation at different times, different light intensities and different backgrounds, mark the disconnection and closing states of the disconnector in each sample image, thereby obtaining a data set, and divide the data set to obtain a training set; In addition, an improved InceptionV3 model is constructed, wherein the improved InceptionV3 model includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first pooling layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, an Inception module a, an Inception module b, an Inception module c, a second pooling layer, a fully connected layer, a Softmax layer, and an output layer; wherein the first convolutional layer is composed of a 3×3 convolutional kernel with a step size of 2, the second convolutional layer is composed of a 3×3 convolutional kernel with a step size of 1, the third convolutional layer is composed of a 3×3 convolutional kernel with a step size of 1, and the pooling layer is composed of a 3×3 convolutional kernel with a step size of 2. The fourth convolution layer is composed of a 3×3 convolution kernel with a stride of 1, the fifth convolution layer is composed of a 3×3 convolution kernel with a stride of 2, and the sixth convolution layer is composed of a 3×3 convolution kernel with a stride of 1. The Inception module a is an Inception structure of two consecutive 3×3 convolution kernels, the Inception module b is an Inception structure that replaces the n×n convolution kernel with consecutive 1×n and n×1 convolution kernels, and the Inception module c is an Inception structure that replaces the 3×3 convolution kernel with parallel 1×3 and 3×1 convolution kernels. The second pooling layer is composed of 8×8 convolution kernels; Step 2: Use the training set data to train the improved InceptionV3 model to obtain a trained improved InceptionV3 model; Next, the fully connected layer, Softmax layer, and output layer of the untrained improved InceptionV3 model are used to replace the fully connected layer, Softmax layer, and output layer of the trained improved InceptionV3 model, and the number of categories of the replaced fully connected layer is adjusted; Then, the Fine Tuning method is used to fine-tune the replaced improved InceptionV3 model to obtain an image classification network; Step 3: Input the substation image to be detected into the image classification network obtained in step 2, and the image classification network outputs the recognition result of the disconnector opening and closing status.

2. A method for detecting the opening and closing state of an isolating switch according to claim 1, characterized in that: In step 1, data augmentation is performed on the sample images to expand the number of sample images, and then a data set is constructed using the sample images with the expanded number.

3. A method for detecting the opening and closing state of an isolating switch according to claim 1, characterized in that: In step 1, the optimizer used in the constructed improved InceptionV3 model is the stochastic steepest descent method.

4. A method for detecting the opening and closing state of an isolating switch according to claim 1, characterized in that: In step 1, the Leaky ReLU activation function is used in the Softmax layer of the constructed improved InceptionV3 model.

5. A method for detecting the opening and closing state of an isolating switch according to claim 1, characterized in that: In step 2, the number of categories of the replaced fully connected layer is adjusted from 1000 to 2 categories.

Images