Information identification network training method and high-voltage switch maintenance method
Through convolutional neural network learning the fault status and maintenance solutions of high-voltage switching equipment, a one-to-one network model is established, which solves the problem that high-voltage switching equipment fault maintenance in the existing technology depends on manual judgment, and improves the reliability and efficiency of maintenance.
Patent Information
- Application Number
- CN202411007743.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-12-06
AI Technical Summary
The fault maintenance of medium and medium-voltage switching equipment in the prior art relies on manual judgment, resulting in ineffective solution reliability and efficiency, and depends on the personal ability and experience accumulation of engineers.
The information recognition network training method is adopted to learn the fault status and maintenance scheme of high-voltage switching equipment through convolutional neural networks, establish one-to-one corresponding network models, and realize automated identification and maintenance scheme generation.
It improves the reliability and efficiency of fault maintenance of high-voltage switching equipment, reduces the dependence on manual judgment, and shortens the time for generation of maintenance plans.
Smart Images

Figure CN119089946A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field, and in particular to an information recognition network training method and a high-voltage switch maintenance method. Background Art
[0002] Usually, when a high-voltage switchgear fails in operation or is discovered during maintenance, the user will report the fault phenomenon to the equipment manufacturer through text description, taking photos, recording videos, etc.
[0003] Equipment manufacturers rely on their knowledge of high-voltage switches and historical experience to determine the status of high-voltage switches, such as whether the contacts are damaged, the degree of wear, whether the pull rods are cracked, whether the sealing rings are damaged, etc., and then issue a detailed maintenance plan. However, this method greatly reduces the reliability and efficiency of the plan because the maintenance plan issued manually depends on the personal ability of the engineer and requires a lot of time to accumulate experience. Summary of the invention
[0004] Therefore, the present invention aims to overcome the technical problems that need to be solved in the prior art, thereby providing an information recognition network training method and a high-voltage switch maintenance method.
[0005] The present invention provides an information recognition network training method, comprising:
[0006] Step 11, obtaining information samples, which are divided into picture samples, video samples, text samples, and voice samples based on their categories;
[0007] Step 12, processing the information samples respectively to obtain key features;
[0008] Step 13, establishing a first convolutional neural network based on the key features and the fault state of the high-voltage switch;
[0009] Step 14: If the test error exceeds the acceptable range, the image sample and the text sample are corrected respectively, a second convolutional neural network is established, and the intersection of the first convolutional neural network and the second convolutional neural network is taken to obtain a third convolutional neural network.
[0010] Furthermore, in step 12, the image samples are processed by dividing and tracking the images using a mixed Gaussian model to obtain a first image training sample with key points;
[0011] To process the speech samples, an acoustic model is used to convert the speech samples into text samples;
[0012] The text samples are processed by using the TextRank algorithm to obtain a first text training sample having a number of keywords;
[0013] The image sample and the text sample are associated, and the above method is combined to obtain a first associated training sample;
[0014] For a video sample with voice, it is decomposed into associated picture samples and voice samples, and combined with the above method, a second associated training sample with both keywords and key points is obtained.
[0015] Furthermore, in step 13, the first picture training sample, the first voice training sample, the first associated training sample, and the first associated training sample, and the corresponding high-voltage switch fault state are respectively divided into a training set and a verification set, and after network training, a first convolutional neural network model corresponding to the training sample and the high-voltage switch fault state is obtained.
[0016] Furthermore, in step 14, the image sample is corrected by binarizing the image sample and obtaining a second image training sample with binary features based on the grayscale of the pixels.
[0017] To correct the text samples, the text samples are randomly interpolated by random interpolation method, and the union is taken, and then the second text training sample is obtained through the speech recognition model.
[0018] The samples associated with the picture sample and the text sample are corrected, and the third associated training sample is obtained by combining the above method.
[0019] For the correction of the video sample with speech, it is decomposed into associated picture samples and text samples, and a fourth associated training sample with both binary features and random interpolation is obtained.
[0020] Furthermore, the second picture training sample, the second voice training sample, the third associated training sample and the fourth associated training sample, and the corresponding high-voltage switch fault state are respectively divided into a training set and a verification set. After network training, a second convolutional neural network model corresponding to the training samples and the high-voltage switch fault state is obtained.
[0021] Furthermore, if the overlap between the first convolutional neural network model and the second convolutional neural network model is higher than a preset value, the correction is successful, and the first convolutional neural network model and the second convolutional neural network model are intersected to obtain a third convolutional neural network.
[0022] An information identification method, comprising:
[0023] Step 21, obtaining high-voltage switch information, wherein the high-voltage switch information includes a high-voltage switch to be identified;
[0024] Step 22, processing the high voltage switch information using the method according to claim 2 to obtain target information;
[0025] Step 23, input the target information into the third convolutional neural network model of the reward to obtain the high-voltage switch fault state of the high-voltage switch to be identified.
[0026] A maintenance plan network training method, comprising:
[0027] Step 31, obtaining a high-voltage switch fault state sample and a corresponding maintenance plan sample;
[0028] Step 32: divide the high-voltage switch fault state samples and maintenance plan samples into a training set and a validation set, and establish a fourth convolutional neural network model based on the high-voltage switch fault state and maintenance plan.
[0029] A high voltage switch maintenance method, comprising:
[0030] Step 41, obtaining a high-voltage switch fault state of the high-voltage switch to be identified, wherein the high-voltage switch fault state is output by the third convolutional neural network model;
[0031] Step 42: input the fault state of the high-voltage switch into the fourth convolutional neural network model to obtain a corresponding maintenance plan.
[0032] The technical solution of the present invention has the following advantages:
[0033] The present invention combines the original historical maintenance information with the network model and establishes a high-voltage switch network model by learning the knowledge of high-voltage switch equipment. In the high-voltage switch network model, the one-to-one correspondence between maintenance information (text description, photo shooting, video recording, etc.) and high-voltage switch fault status is learned and trained through a neural network, and the network model in which the maintenance information corresponds to the high-voltage switch fault status is determined and corrected. Then, the one-to-one correspondence between the high-voltage switch fault status and the maintenance plan is learned and trained through a neural network, and the network model in which the high-voltage switch fault status corresponds to the maintenance plan is determined. The output of the former network model is the input of the latter network model. After the establishment of the network model is completed, when the maintenance information is obtained again, the corresponding maintenance plan can be directly obtained. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0035] Figure 1 This is a schematic diagram of the principle of the information recognition network training method of the present invention;
[0036] Figure 2It is a schematic diagram of the principle of the information identification method of the present invention;
[0037] Figure 3 This is a schematic diagram of the principle of the maintenance scheme network training method of the present invention;
[0038] Figure 4 It is a schematic diagram of the principle of the high-voltage switch maintenance method of the present invention. DETAILED DESCRIPTION
[0039] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. 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.
[0040] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.
[0041] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0042] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0043] Example
[0044] Usually, when a high-voltage switchgear fails in operation or is discovered during maintenance, the user will report the fault phenomenon to the equipment manufacturer through text description, taking photos, recording videos, etc.
[0045] Equipment manufacturers rely on their knowledge of high-voltage switches and historical experience to determine the status of high-voltage switches, such as whether the contacts are damaged, the degree of wear, whether the pull rods are cracked, whether the sealing rings are damaged, etc., and then issue a detailed maintenance plan. However, this method greatly reduces the reliability and efficiency of the plan because the maintenance plan issued manually depends on the personal ability of the engineer and requires a lot of time to accumulate experience.
[0046] like Figures 1 to 4 A specific implementation method of an information recognition network training method is shown, and the information recognition network training method includes:
[0047] Step 11, obtaining information samples, which are divided into picture samples, video samples, text samples, and voice samples based on their categories;
[0048] Step 12, processing the information samples respectively to obtain key features;
[0049] Step 13, establishing a first convolutional neural network based on the key features and the fault state of the high-voltage switch;
[0050] Step 14: If the test error exceeds the acceptable range, the image sample and the text sample are corrected respectively, a second convolutional neural network is established, and the intersection of the first convolutional neural network and the second convolutional neural network is taken to obtain a third convolutional neural network.
[0051] The training sample set (x i ,y i ), then the neural network algorithm can provide a complex and nonlinear hypothesis model h w,b (x), which has parameters W and b, which can be used to fit the data.
[0052] When a neuron is an arithmetic unit with (x1, x2, x3) and an intercept term +1 as input values, its output is
[0053] The function f is called the activation function. Sigmoid function is used as the activation function Then the input-output mapping relationship of this neuron is actually a logistic regression.
[0054] Furthermore, in step 12, the image samples are processed by dividing and tracking the images using a mixed Gaussian model to obtain a first image training sample with key points;
[0055] In image processing, since images often have too many pixels, if they are put into neural network training, the amount of calculation will be too huge. Therefore, it is necessary to first reduce the parameters input to the neural network, and convolutional neural networks can effectively do this through convolution and pooling operations.
[0056] Convolution: Natural images have inherent properties, that is, the statistical properties of one part of the image are the same as those of other parts, which means that the features learned in one part of the image can also be used in another part, so the same learned features can be used for all positions in the image.
[0057] When a small piece, say 8x8, is randomly selected from a large-size image as a sample, and some features are learned from this small sample, the features learned from this 8x8 sample can be used as a detector and applied to any part of the image. In particular, the features learned from the 8x8 sample can be convolved with the original large-size image to obtain a different feature activation value for any position on the large-size image.
[0058] Assume that a large image of size r×c is given, which is defined as x large. First, a sparse autoencoder is trained by extracting a×b small-size image samples x small from the large-size image, and f = σ(W(1)x small+b(1)) is calculated, where σ is a sigmoid function, and k features are obtained, where W(1) and b(1) are the weights and bias values between the visible layer unit and the hidden unit. For each small image xs of size a×b, the corresponding value fs = σ(W(1)x s+b(1)) is calculated. By convolving these values, a matrix of k×(r-a+1)×(c-b+1) convolved features can be obtained.
[0059] Pooling: After obtaining features through convolution, overfitting is prone to occur, and there are still too many features. Pooling is used to further reduce the feature values. The assumption of pooling is that features that are useful in one image area are very likely to be applicable in another area. Therefore, in order to describe large images, a natural idea is to aggregate statistics of features at different locations. For example, one can calculate the average (or maximum) value of a specific feature in an area of the image. These summary statistical features not only have a much lower dimension (compared to using all extracted features), but also improve the results (not prone to overfitting). This aggregation operation is called pooling, sometimes also called average pooling or maximum pooling (depending on the method of calculating the pooling).
[0060] Generally, multi-layer convolution + pooling is used, and then a fully connected layer (hidden layer) is used for training. The purpose of multi-layer convolution is that the features learned by a layer of convolution are often local. The higher the number of layers, the more global the features learned.
[0061] To process the speech samples, an acoustic model is used to convert the speech samples into text samples;
[0062] The processing of speech samples mainly involves converting the known content of speech audio files into audio features through feature extraction methods, and then training a decoder through the audio features and the corresponding audio content through the acoustic model. Then, feature extraction is performed on unknown audio samples, and the extracted audio features are decoded through the decoder to obtain the corresponding audio content. The audio content obtained through analysis is then corrected by the language model after audio recognition, and finally the speech content of the unknown audio file is output.
[0063] After obtaining the speech sample, the acoustic model is established through the GMM-HMM (Gaussian Mixture-Hidden Markov Model) algorithm to recognize the speech. HMM (Hidden Markov Model) is a statistical model used to describe a Markov process with hidden unknown parameters. It is used to solve the problem that the state cannot be observed directly but can be determined through the observation sequence. Each observation sequence is a variety of states expressed through certain probability density distributions.
[0064] For HMM, there are three important assumptions:
[0065] Markov hypothesis: p(Xi|Xi-1,…,X1)=P(Xi|Xi-1)
[0066] Immobility assumption: P(Xi+1|Xi)=P(Xj+1|Xj)
[0067] Output independence assumption: P(01,…,0T|X1,…,XT)=ПP(O t|X t)
[0068] There is a probabilistic relationship between the hidden state and the observed sequence. When the hidden state H is considered to be a probability of a sequence 0 that can be observed, P(O|H). Assuming there are three observed sequences, P(01|H)+P(02|H)+P(03|H)=1. From this, the corresponding probability between the state and the observed sequence (called the confusion matrix) can be constructed:
[0069] Observation sequence 1 Observation sequence 2 State 1 P(01丨H1) P(02丨H1) State 2 P(01丨H2) P(02丨H2)
[0070] A HMM has 5 elements {N, M, π, A, B}, where N is the number of hidden states, which is known by the requirements, M is the number of observation sequences, which is obtained through the training set, π is the initial state probability, A is the hidden state transition probability matrix, and B is the confusion matrix. The HMM model can be written as: λ = (A, B, π). The HMM model can be established by obtaining the various parameters of λ through the training set. For the acoustic model using the RMM model, the actual content of the audio speech is the state, and the observation sequence is obtained by fitting the MFCC features of each frame of the audio through GMM. After constructing the five necessary elements of the HMM model, the various parameters are obtained through an iterative method to complete the establishment of the acoustic model.
[0071] When recognizing, the MFCC features of the speech are used and the observation sequence is obtained through GMM calculation. The state with the highest probability is calculated through this observation sequence and output as the result, which is the content of the audio.
[0072] When calculating the state with the highest probability among all states, if there are too many states, the amount of calculation will be too large. Usually, the Viterbi search method (Viterbi algorithm) is used to find the state output with the highest probability in the RMM model. The Viterbi algorithm is a dynamic programming algorithm used to find the Viterbi path that is most likely to generate an observation sequence from a state.
[0073] Assume that in a hidden Markov model (HMM), the state space is S, the initial probability of state i is πi, and the transition probability from state i to j is ai,j. The observation sequence is y1,…,yT. The most likely state sequence that produces the observation value is x1,x2,…,xT, which can be obtained by the iterative relationship:
[0074] P1, k = πk·P(y1|k)
[0075] Pt,k=max{Pt-1,x·ax,k·P(yT|k)}
[0076] Where Pt,k is the probability of the most likely state sequence corresponding to the first t observations with k as the final state, and P(yT|k) is the probability of generating the observation sequence yT corresponding to the hidden state k.
[0077] The HMM model is essentially a sequence classifier, which recognizes a sequence of one length as a sequence of another length. Its training process is to use each training sample and its corresponding sentence to use the Viterbi method to iteratively update the mean vector and covariance matrix of each state in the GMM and their state transition probabilities, and finally reach convergence.
[0078] The text samples are processed by using the TextRank algorithm to obtain a first text training sample having a number of keywords;
[0079] The image sample and the text sample are associated, and the above method is combined to obtain a first associated training sample;
[0080] For a video sample with voice, it is decomposed into associated picture samples and voice samples, and combined with the above method, a second associated training sample with both keywords and key points is obtained.
[0081] Furthermore, in step 13, the first picture training sample, the first voice training sample, the first associated training sample, and the first associated training sample, and the corresponding high-voltage switch fault state are respectively divided into a training set and a verification set. After network training, a first convolutional neural network model corresponding to the training sample and the high-voltage switch fault state is obtained. The implementation method of the first convolutional neural network model can be achieved by establishing a matrix relationship between the training sample and the high-voltage switch fault state.
[0082] Optionally, for the fault state of the high-voltage switch, electrical principle samples and mechanical principle samples in the original database can be obtained, wherein the electrical principle samples and the mechanical principle samples can both include picture samples, video samples, text samples, and voice samples. Through the above step 12, the electrical principle samples and the mechanical principle samples are processed to obtain the first picture principle sample, the first voice principle sample, the first associated principle sample, and the first associated principle sample, wherein the features in the principle samples obtained through step 12 include: electrical part: closing insurance connection state, voltage operating range, operating handle working position, contactor closing coil working state, closing loop resistance connection state; mechanical part: switch body and contactor working contour position, main shaft string movement or pin working contour position, closing bracket working contour position, bracket working angle, opening pin hook working contour position, closing coil working contour position.
[0083] To build a feature comparison network, you can use the SimCLR comparative learning network to compare the image training samples with the image principle samples and establish associations based on similarity.
[0084] Optional, including:
[0085] 1. Take an input image: perform two types of data augmentation on the same image to form a positive sample pair; different images are negative sample pairs.
[0086] 2. Prepare 2 random image enhancements: rotation, color, saturation, brightness change, scaling, cropping, etc. (Construct positive samples: image SimCLR-data enhancement, text SimCSE-Dropout, image-text CLIP-image-text pair)
[0087] 3. Feature extraction: Run a deep neural network (can be a convolutional neural network such as ViT, Bert, ResNet50) to obtain image feature representation of the enhanced image.
[0088] 4. Feature Projection: Run a fully connected linear neural network to project the embeddings into another vector space.
[0089] 5. Calculate loss: Calculate contrast loss and back propagate through both networks. When projections from the same image are similar, the contrast loss decreases. The similarity between projections can be arbitrary, and cosine similarity can be used.
[0090] 6. Downstream tasks: Use contrastive learning to obtain the Encoder as a feature extractor, and fine-tune Finetuin according to the dataset of the downstream task.
[0091] Output the difference features and similarities between the image training samples and the image principle samples.
[0092] Optionally, the RoBERTa Base Chinese contrastive learning network can be used to compare the text training samples and the mosquito principle samples, and establish associations based on similarity. The difference features and similarities between the text training samples and the mosquito principle samples are output, and the association and comparison between the training samples and the principle samples are established.
[0093] Build a discrete selection model to select the training samples and principle samples with the highest similarity.
[0094] Establish an association matrix between difference features and switch fault states, where the switch fault states may include electrical circuit faults and mechanical faults. Electrical circuit faults include: closing fuse blown or poor contact, DC voltage too low, operating handle, switch auxiliary contact poor or broken, contactor closing coil short-circuited, tiger gate closed by motor, closing circuit resistor broken or not returning after tripping. Mechanical faults include: switch body and contactor stuck (such as SN1-10 guide tube detached, DW2-35 lifting pin too long, etc.), large shaft string movement or pin falling off, closing bracket stuck due to excessive oil mud, bracket slope large, not straight or small, three points too high, opening pin hook not firmly closed, mechanism stuck, not returned to the pre-closing position, closing buffer gap small, closing coil core overrun small.
[0095] By associating the training samples and principle samples, the training samples and principle samples with higher similarity are selected, and then the corresponding switch fault status is obtained based on the difference characteristics.
[0096] Furthermore, in step 14, the image sample is corrected by binarizing the image sample and obtaining a second image training sample with binary features based on the grayscale of the pixels.
[0097] The text samples are modified by performing random interpolation on the text samples through a random interpolation method, such as the Lagrange interpolation method, taking the union, and then obtaining a second text training sample through a speech recognition model.
[0098] The samples associated with the picture sample and the text sample are corrected, and the third associated training sample is obtained by combining the above method.
[0099] For the correction of the video sample with speech, it is decomposed into associated picture samples and text samples, and a fourth associated training sample with both binary features and random interpolation is obtained.
[0100] Furthermore, the second picture training sample, the second voice training sample, the third associated training sample and the fourth associated training sample, and the corresponding high-voltage switch fault state are respectively divided into a training set and a verification set. After network training, a second convolutional neural network model corresponding to the training samples and the high-voltage switch fault state is obtained.
[0101] Furthermore, if the overlap between the first convolutional neural network model and the second convolutional neural network model is higher than a preset value, the correction is successful, and the first convolutional neural network model and the second convolutional neural network model are intersected to obtain a third convolutional neural network.
[0102] An information identification method, comprising:
[0103] Step 21, obtaining high-voltage switch information, wherein the high-voltage switch information includes a high-voltage switch to be identified;
[0104] Step 22, processing the high voltage switch information using the method according to claim 2 to obtain target information;
[0105] Step 23, input the target information into the third convolutional neural network model of the reward to obtain the high-voltage switch fault state of the high-voltage switch to be identified.
[0106] A maintenance plan network training method, comprising:
[0107] Step 31, obtaining a high-voltage switch fault state sample and a corresponding maintenance plan sample;
[0108] Step 32: divide the high-voltage switch fault state samples and maintenance plan samples into a training set and a validation set, and establish a fourth convolutional neural network model based on the high-voltage switch fault state and maintenance plan.
[0109] A high voltage switch maintenance method, comprising:
[0110] Step 41, obtaining a high-voltage switch fault state of the high-voltage switch to be identified, wherein the high-voltage switch fault state is output by the third convolutional neural network model;
[0111] Step 42: input the fault state of the high-voltage switch into the fourth convolutional neural network model to obtain a corresponding maintenance plan.
[0112] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.
Claims
1. A method for training an information recognition network, It is characterized in that include: Step 11, obtaining information samples, which are divided into picture samples, video samples, text samples, and voice samples based on their categories; Step 12, processing the information samples respectively to obtain key features; Step 13, establishing a first convolutional neural network based on the key features and the fault state of the high-voltage switch; Step 14: If the test error exceeds the acceptable range, the image sample and the text sample are corrected respectively, a second convolutional neural network is established, and the intersection of the first convolutional neural network and the second convolutional neural network is taken to obtain a third convolutional neural network.
2. The information recognition network training method according to claim 1, It is characterized in that In step 12, the image samples are processed by dividing and tracking the images using a mixed Gaussian model to obtain a first image training sample with key points; To process the speech samples, an acoustic model is used to convert the speech samples into text samples; The text samples are processed by using the TextRank algorithm to obtain a first text training sample having a number of keywords; The image sample and the text sample are associated, and the above method is combined to obtain a first associated training sample; For a video sample with voice, it is decomposed into associated picture samples and voice samples, and combined with the above method, a second associated training sample with both keywords and key points is obtained.
3. The information recognition network training method according to claim 2, It is characterized in that In step 13, the first picture training sample, the first voice training sample, the first associated training sample, the first associated training sample, and the corresponding high-voltage switch fault state are divided into a training set and a verification set respectively. After network training, a first convolutional neural network model corresponding to the training sample and the high-voltage switch fault state is obtained.
4. The information recognition network training method according to claim 3, It is characterized in that In step 14, the image sample is corrected by binarizing the image sample and obtaining a second image training sample with binary features based on the grayscale of the pixels. To correct the text samples, the text samples are randomly interpolated by random interpolation method, and the union is taken, and then the second text training sample is obtained through the speech recognition model. The samples associated with the picture sample and the text sample are corrected, and the third associated training sample is obtained by combining the above method. For the correction of the video sample with speech, it is decomposed into associated picture samples and text samples, and a fourth associated training sample with both binary features and random interpolation is obtained.
5. The information recognition network training method according to claim 4, It is characterized in that The second picture training sample, the second voice training sample, the third associated training sample and the fourth associated training sample, and the corresponding high-voltage switch fault state are respectively divided into a training set and a verification set. After network training, a second convolutional neural network model corresponding to the training samples and the high-voltage switch fault state is obtained.
6. The information recognition network training method according to claim 5, It is characterized in that If the overlap between the first convolutional neural network model and the second convolutional neural network model is higher than a preset value, the correction is successful, and the first convolutional neural network model and the second convolutional neural network model are intersected to obtain a third convolutional neural network.
7. A method for identifying information, It is characterized in that include: Step 21, obtaining high-voltage switch information, wherein the high-voltage switch information includes a high-voltage switch to be identified; Step 22, processing the high voltage switch information using the method according to claim 2 to obtain target information; Step 23, input the target information into the third convolutional neural network model of the reward to obtain the high-voltage switch fault state of the high-voltage switch to be identified.
8. A maintenance plan network training method, It is characterized in that include: Step 31, obtaining a high-voltage switch fault state sample and a corresponding maintenance plan sample; Step 32: divide the high-voltage switch fault state samples and maintenance plan samples into a training set and a verification set, and establish a fourth convolutional neural network model based on the high-voltage switch fault state and maintenance plan.
9. A method for repairing a high voltage switch. It is characterized in that include: Step 41, obtaining a high-voltage switch fault state of the high-voltage switch to be identified, wherein the high-voltage switch fault state is output by the third convolutional neural network model; Step 42: input the fault state of the high-voltage switch into the fourth convolutional neural network model to obtain a corresponding maintenance plan.
