Intelligent high-voltage switch state evaluation method and device, terminal and storage medium

Abstract

The application provides a state evaluation method and device of an intelligent high-voltage switch, a terminal and a storage medium. The method comprises the following steps: obtaining a sensing data set of the intelligent high-voltage switch; training a stacked denoising autoencoder (SDAE) and a deep belief network (DBN) respectively according to the sensing data set to obtain the weight of the SDAE and the weight of the DBN; fusing the weight of the SDAE and the weight of the DBN by using a weighted fusion method to obtain a fusion weight, initializing the weight of a target evaluation model according to the fusion weight, and testing the feasibility of the target evaluation model; when the feasibility of the target evaluation model meets a preset condition, obtaining operation data of the intelligent high-voltage switch, inputting the operation data into the target evaluation model, and obtaining the operation state of the operation data. The application can accurately and efficiently evaluate the state of the high-voltage switch.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent high-voltage switch state evaluation, and in particular to a state evaluation method and device of an intelligent high-voltage switch, a terminal and a storage medium. BACKGROUND

[0002] The high-voltage switch is a core interlocking device in the power system, and abnormal operation state thereof can cause large-scale power outage and affect normal production and life of enterprises and residents. The high-voltage switch is the most complex part of the entire power transmission and transformation device, and many index data need to be evaluated during operation, resulting in uncertainty of related indexes. Therefore, the performance of the high-voltage switch needs to be evaluated from multiple angles.

[0003] At present, domestic and foreign power institutions mainly obtain data based on daily inspection, live detection, online monitoring and power-off test, and determine the operation state of the high-voltage switch according to relevant standards and expert experience. However, this evaluation method has three problems: 1. It is difficult to form universal parameters and is not suitable for automatic evaluation by traditional programming method; 2. The overall score deviation is prone to occur when the staff evaluates the state according to their own experience; and 3. There is a lack of comprehensive measurement of the internal correlation of the state. In addition, if regular maintenance is completely adopted, it may cause over-maintenance or under-maintenance, resulting in waste of manpower and material resources and reduction of power supply reliability. Therefore, how to accurately evaluate the state of the high-voltage switch and carry out maintenance work on demand to reduce operation and maintenance cost is one of the key works of the power industry. SUMMARY

[0004] The embodiments of the present application provide a state evaluation method, device, terminal and storage medium of an intelligent high-voltage switch to solve the problem that the state of the high-voltage switch cannot be accurately and efficiently evaluated in the prior art.

[0005] In a first aspect, the embodiments of the present application provide a state evaluation method of an intelligent high-voltage switch, comprising:

[0006] obtaining a sensing data set of the intelligent high-voltage switch;

[0007] training a stacked denoised autoencoder (SDAE) and a deep boltzmann machine (DBN) according to the sensing data set respectively to obtain weights of the SDAE and weights of the DBN;

[0008] fusing the weights of the SDAE and the weights of the DBN by using a weighted fusion method to obtain fusion weights, initializing weights of a target evaluation model according to the fusion weights, and testing the feasibility of the target evaluation model;

[0009] When the feasibility of the target evaluation model meets a preset condition, operation data of the intelligent high-voltage switch is acquired, the operation data is input into the target evaluation model, and an operation state of the operation data is obtained.

[0010] In a possible implementation, the training of the stack denoising autoencoder (SDAE) and the deep belief network (DBN) according to the sensor data set respectively to obtain the weight of the SDAE and the weight of the DBN includes:

[0011] The sensor data set is divided into an initial training sample set and an initial test sample set;

[0012] The initial training sample set and the initial test sample set are respectively subjected to dimension reduction processing to obtain a target training sample set and a target test sample set;

[0013] The stack denoising autoencoder (SDAE) is trained according to the target training sample set to obtain the weight of the SDAE;

[0014] The deep belief network (DBN) is trained according to the target training sample set to obtain the weight of the DBN.

[0015] In a possible implementation, the dimension reduction processing of the initial training sample set and the initial test sample set respectively includes:

[0016] The initial training sample set is subjected to a decentralization operation to obtain a first training sample set;

[0017] A covariance matrix of the first training sample set is calculated, eigenvalues of the covariance matrix and eigenvectors corresponding to the eigenvalues are calculated, the eigenvalues are sorted in descending order, the contribution rate and the cumulative contribution rate of the eigenvectors are calculated based on the eigenvalues, and eigenvectors with a cumulative contribution rate greater than a first preset value are selected to form a projection matrix;

[0018] The first training sample set is multiplied by the projection matrix to obtain a dimension-reduced first training sample set, and the dimension reduction processing is implemented.

[0019] In a possible implementation, the training of the deep belief network (DBN) according to the target training sample set to obtain the weight of the DBN includes:

[0020] The target training sample set is input into a visible layer neuron, and the probability that a hidden layer neuron is activated is calculated according to The visible layer neuron and the hidden layer neuron are in a symmetrical full connection relationship, where v represents a visible layer neuron, h j represents the jth element of a hidden layer neuron, and vi represents the i-th element of the visible layer neuron, i represents the current sample number, m represents the total number of samples in the sample set, W ij represents the element in the i-th row and the j-th column of the weight matrix connecting the visible layer neuron and the hidden layer neuron, c j represents the j-th element of the hidden layer bias vector;

[0021] extracting a random sample, reconstructing the visible layer neuron according to the random sample to obtain a reconstructed visible layer neuron, and calculating the probability of the visible layer neuron being activated according to i represents the i-th element of the reconstructed visible layer neuron, h represents the hidden layer neuron, b i represents the i-th element of the visible layer bias vector;

[0022] reconstructing the visible layer neuron, calculating the probability of the hidden layer neuron being activated according to j represents the j-th element of the hidden layer neuron on the basis of the reconstructed visible layer neuron;

[0023] updating the weight of the DBN according to W = W + η (h'v T T represents the transpose of the visible layer neuron.

[0024] In a possible implementation, the method for fusing the weight of the SDAE and the weight of the DBN to obtain the fused weight comprises the following steps.

[0025] ds d s ds s d

[0026] In a possible implementation, the method for testing the feasibility of the target evaluation model comprises the following steps.

[0027] inputting the target test sample set into the target evaluation model, and outputting the running state of each sample in the target test sample set;

[0028] ​​​​​​​​​​According to and determining accuracy of the target evaluation model, testing feasibility of the target evaluation model according to the accuracy;

[0029] wherein, MAPE represents mean absolute percentage error, MAE represents mean absolute error, n represents total number of samples in the target test sample set, t represents current sample number, y t represents output running state, y t represents real running state.

[0030] In a second aspect, an embodiment of the present application provides a state evaluation device of an intelligent high-voltage switch, comprising:

[0031] an acquisition module, configured to acquire a sensing data set of the intelligent high-voltage switch;

[0032] a training module, configured to train a stack denoising auto-encoder (SDAE) and a deep belief network (DBN) respectively according to the sensing data set, to obtain weights of the SDAE and weights of the DBN;

[0033] a fusion verification module, configured to fuse the weights of the SDAE and the weights of the DBN by using a weighted fusion method to obtain fusion weights, to initialize weights of a target evaluation model according to the fusion weights, and to test feasibility of the target evaluation model;

[0034] an evaluation module, configured to acquire running data of the intelligent high-voltage switch when the feasibility of the target evaluation model meets a preset condition, to input the running data into the target evaluation model, and to obtain a running state of the running data.

[0035] In a third aspect, an embodiment of the present application provides a terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements steps of the method according to the first aspect or any possible implementation manner of the first aspect.

[0036] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program is executable on a processor to implement steps of the method according to the first aspect or any possible implementation manner of the first aspect.

[0037] The embodiment of the present application provides a state evaluation method and device of an intelligent high-voltage switch, a terminal and a storage medium, the state evaluation method comprises the following steps: obtaining a sensing data set of the intelligent high-voltage switch; training a stacked denoising autoencoder (SDAE) and a deep belief network (DBN) according to the sensing data set respectively to obtain the weight of the SDAE and the weight of the DBN; fusing the weight of the SDAE and the weight of the DBN by using a weighted fusion method to obtain a fusion weight; initializing the weight of a target evaluation model according to the fusion weight and testing the feasibility of the target evaluation model; when the feasibility of the target evaluation model meets a preset condition, obtaining operation data of the intelligent high-voltage switch, inputting the operation data into the target evaluation model to obtain the operation state of the operation data. The SDAE and the DBN are trained, the weight of the SDAE and the weight of the DBN are fused, the weight of the target evaluation model is initialized, the target evaluation model has the properties of the SDAE and the DBN, the characteristics of the SDAE and the DBN are fully utilized, when the operation data is input, the time complexity of evaluation can be reduced, and the evaluation accuracy can be observed, and the operation state of the operation data can be obtained efficiently and accurately. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0039] Figure 1 is the implementation flowchart of the state evaluation method of the intelligent high-voltage switch provided by the embodiment of the present application;

[0040] Figure 2 is the network structure diagram of the SDAE provided by the embodiment of the present application;

[0041] Figure 3 is the training flowchart of the SDAE provided by the embodiment of the present application;

[0042] Figure 4 is the network structure diagram of the DBN provided by the embodiment of the present application;

[0043] Figure 5 is the structural schematic diagram of the state evaluation device of the intelligent high-voltage switch provided by the embodiment of the present application;

[0044] Figure 6 is the schematic diagram of the terminal provided by the embodiment of the present application. DETAILED DESCRIPTION

[0045] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known methods, devices, circuits, and

[0046] In order to make the objects, technical solutions and advantages of the present application clearer, the following will be described in conjunction with the accompanying drawings and specific embodiments.

[0047] Figure 1 The implementation flowchart of the state evaluation method of the intelligent high-voltage switch provided by the embodiments of the present application is described in detail as follows.

[0048] In step 101, a sensing data set of the intelligent high-voltage switch is acquired.

[0049] Optionally, the data in the sensing data set can include the closing and opening coil current measured by the current sensor, the heating fault temperature measured by the temperature sensor, the vibration signal of the rotating mechanical bearing measured by the vibration sensor, the humidity data measured by the humidity sensor, the insulation fault data measured by the partial discharge sensor, the sound signal during closing and opening measured by the sound sensor, the angle of the rotating mechanical shaft measured by the angle sensor, the rotating attitude angle of the disconnecting switch contact measured by the attitude sensor, the normalized pressure of SF6 in the high-voltage switch measured by the SF6 sensor, the microswitch in-place information measured by the position sensor, the spring pressure data measured by the energy storage sensor, and the energy storage motor current data measured by the current sensor.

[0050] In order to ensure the accuracy of the sensing data set, the sensing data set needs to be down-sampled and data-interpolated to obtain an intermediate sensing data set before grouping the sensing data set. The down-sampling can avoid the problem of wasting training time and resources caused by too dense and slow-changing data in the sensing data set. The data interpolation refers to adding or deleting data points in the data at random. The data interpolation can make the sensing data set have more random factors and meet the characteristics of the pseudo-periodic data.

[0051] After the above operations on the sensing data set, the intermediate sensing data set needs to be filtered to obtain normal data and fault data, and typical fault samples are selected, and then the intermediate sensing data set is normalized.

[0052] Optionally, according to the normalization of the intermediate sensing data set, the normalization of the target sensing data set is performed. The normalization of the intermediate sensing data set is performed, wherein X nor represents the target sensing data set, X z represents the data in the intermediate sensing data set, and X minX represents the minimum value of data in the intermediate sensing data set max X represents the maximum value of data in the intermediate sensing data set.

[0053] In step 102, the stack denoising auto-encoder SDAE and the deep belief network DBN are respectively trained according to the sensing data set, so as to obtain the weight of the SDAE and the weight of the DBN.

[0054] In a possible implementation, the sensing data set is divided into an initial training sample set and an initial test sample set;

[0055] Optionally, the proportion of the initial training sample set and the initial test sample set is 4:1.

[0056] The initial training sample set and the initial test sample set are respectively subjected to dimension reduction processing, so as to obtain a target training sample set and a target test sample set;

[0057] Optionally, the dimension reduction processing can be principal component analysis (PCA) dimension reduction. The PCA algorithm is a feature extraction technology based on linear mapping, which transforms the multi-dimensional high-voltage switch data to a new low-dimensional space, so that the maximum variance of the high-dimensional data is projected on the first low-dimensional space coordinate (i.e. the first principal component), the second largest variance is projected on the second low-dimensional space coordinate (the second principal component), and so on. The original multi-dimensional high-voltage switch data is maximally retained by using a few principal components, and the first principal component contains most of the information in the original multi-dimensional high-voltage switch data. The PCA algorithm mainly utilizes the property that the covariance matrix is a real diagonal matrix, i.e. variance maximization and covariance minimization, to perform dimension reduction.

[0058] In a possible implementation, the initial training sample set is subjected to a decentralization operation to obtain a first training sample set;

[0059] The covariance matrix of the first training sample set is calculated, the eigenvalues of the covariance matrix and the eigenvectors corresponding to the eigenvalues are calculated, the eigenvalues are sorted in descending order, the contribution rate and the cumulative contribution rate of the eigenvectors are calculated based on the eigenvalues, and the eigenvectors with the cumulative contribution rate greater than a first preset value are selected to form a projection matrix;

[0060] According to The covariance matrix of the first training sample set is calculated, wherein X' represents the first training sample set, C represents the covariance matrix, and X' T X represents the transpose matrix of the first training sample set, and x represents the number of samples in the first training sample set.

[0061] Optionally, according to The contribution rate and the cumulative contribution rate of the feature vector are calculated, wherein, σ y represents the contribution rate of the feature vector corresponding to any one eigenvalue, and δ represents the cumulative contribution rate of the feature vector, λ y represents any one eigenvalue, λ P represents the last eigenvalue, σ P represents the contribution rate of the feature vector corresponding to the last eigenvalue.

[0062] In a possible implementation, the eigenvalues are sorted in descending order, and the feature vectors are arranged in sequence following the order of the eigenvalues, and the contribution rates of the feature vectors calculated based on the eigenvalues are arranged in descending order.

[0063] As shown in the fault feature table in Table 1, the fault types and fault features of the fault samples are listed in Table 1, the fault features are numbered, and 31 fault features can be obtained.

[0064] Table 1 Fault feature table

[0065]

[0066]

[0067] The PCA principal component analysis is performed on the fault features of all the fault samples, the contribution rates of all the principal component vectors are sequentially calculated, and then all the fault features are arranged in descending order according to the contribution rates. As shown in the contribution rate table in Table 2, the contribution rates and the cumulative contribution rates corresponding to the fault features are listed in Table 2.

[0068] Table 2 Contribution rate table

[0069]

[0070]

[0071] Optionally, the present application does not limit the value of the first preset value, for example, the first preset value can be 90%, 85%, and 80%. When the first preset value is 90%, it can be obtained from Table 2 that the cumulative contribution rate of the PCA numbers 1-22 is 90.9184%, and therefore, the fault features corresponding to the first 22 PCA numbers can be selected as the best fault features. In addition, the PCA number corresponding to the cumulative contribution rate reaching the first preset value is taken as the dimension of the projection matrix.

[0072] Optionally, it is assumed that the PCA number corresponding to the cumulative contribution rate reaching the first preset value is k, and the first k feature vectors are selected to form the projection matrix.

[0073] The first training sample set is multiplied by the projection matrix to obtain a first training sample set after dimension reduction, so as to realize dimension reduction processing.

[0074] The dimension of the sample set can be reduced, the calculation complexity can be reduced, and the redundant feature vectors can be reduced by performing dimension reduction processing on the sample set data.

[0075] The stack denoising auto-encoder SDAE is trained according to the target training sample set to obtain the weight of the SDAE;

[0076] The stack denoising auto-encoder SDAE is composed of multiple denoising auto-encoders (DAE), Figure 2 is a network structure diagram of the SDAE. w is a weight matrix of the (w-1) th hidden layer and the w th hidden layer, and B w is a threshold matrix of the (w-1) th hidden layer and the w th hidden layer, and the training process includes two steps of greedy layer-by-layer unsupervised pre-training and supervised fine-tuning. The greedy layer-by-layer unsupervised pre-training obtains the initial weight and threshold of the network by layer-by-layer training. The input of the bottom DAE is the original data, and the hidden layer output data is used as the input data of the upper DAE. Each layer needs to erode the input data.

[0077] As shown in the network structure diagram of the SDAE, Figure 2 The SDAE is composed of an input layer, multiple hidden layers and a classifier, and the output of the previous layer is the input of the next layer.

[0078] In one possible implementation, as shown in the training flowchart of the SDAE, Figure 3 The training steps include:

[0079] S1, the target training sample set is divided into M groups of sub-sample sets as the current sample set, and the M groups of sub-sample sets are sequentially input to the current hidden layer of the SDAE. The weight and threshold of the SDAE corresponding to each group of sub-sample sets are calculated; and the feature vector of the current hidden layer is calculated according to the weight and threshold of the SDAE corresponding to each group of sub-sample sets.

[0080] The SDAE structure is composed of an encoder and a decoder. The input vector is mapped to the feature vector in the hidden layer through the encoder, and then the feature vector is reconstructed into the original input vector through the decoder.

[0081] Optionally, a sample set X is set, X={x i 1≤i≤m} is calculated according to H=S f (WX+B) to calculate the feature vector corresponding to the sample set, wherein H represents the feature vector corresponding to the sample set, S frepresents the neuron activation function of the encoder. The activation function usually adopts a sigmoid function, which has good feature recognition.

[0082] The decoder is an inverse operation of the encoder, taking the feature vector of the hidden layer as the input vector, and setting as the output vector set, as x i The corresponding output vector, then the expression of the decoder is: W' is the weight of the hidden layer and the output layer, B' is the threshold value of the hidden layer and the output layer, s g is the neuron activation function of the decoder.

[0083] In a possible implementation, during the process of inputting the sample set for training, a random inactivation technique is further adopted, and part of the neurons are randomly selected to temporarily stop working, which can reduce the common action between specific nodes, effectively improve the generalization ability, and avoid the overfitting problem.

[0084] Based on the operation process of the above-mentioned encoder and decoder, the reconstruction error between the output vector and the sample set is minimized. Optionally, according to the reconstruction error is determined, wherein, represents the reconstruction error between the output vector and the sample set.

[0085] In a possible implementation, a gradient descent algorithm is adopted to continuously adjust the weight and the threshold value, and to reduce the reconstruction error, and the formula of the gradient descent algorithm is wherein, represents the reconstruction error between the output vector and the sample set, represents the second-order partial derivative of the weight, represents the second-order partial derivative of the threshold value.

[0086] Optionally, during the process of continuously adjusting the weight and the threshold value by using the gradient descent algorithm, the weight and the threshold value of the current hidden layer are saved.

[0087] S2, taking the feature vector as the current sample set and the next layer of the hidden layer of the SDAE as the current hidden layer, jump to step S1 for execution until the maximum number of hidden layers is reached;

[0088] S3, according to the weight and the threshold value of the SDAE are fine-tuned;

[0089] Optionally, after the hierarchical pre-training is completed, the hidden layers are stacked, and the input data and the output data relationship is represented as:

[0090] Objective function of the construction parameter optimization, that is, error loss function of the actual value and the predicted value Fine-tune the whole network weight and threshold by top-down back propagation, reduce the error of the predicted value and the actual value.

[0091] Calculate the gradient according to the error loss function, including:

[0092] Calculate the square gradient, including:

[0093] Update the weight and threshold according to the gradient and the square gradient, including:

[0094] Wherein, is x i The corresponding predicted value, W represents the weight of the SDAE, B represents the threshold of the SDAE, r1 and r2 represent the cumulative variables of the weight and the threshold respectively, w represents the weight gradient, b represents the threshold gradient, x i represents the i-th sample, y i represents x i The corresponding actual value, m represents the total number of samples in the sample set, i represents the current sample number, represents the second derivative symbol, l represents the global learning rate, ρ represents the decay rate, f represents the relationship function between the weight and the threshold, and represents the element-wise product symbol.

[0095] In a possible implementation, after the weight of the SDAE is obtained by training the stack denoising auto-encoder SDAE according to the target training sample set, the method further includes:

[0096] According to and Determine the accuracy of the SDAE, and test the feasibility of the SDAE according to the accuracy.

[0097] Wherein, MAPE represents the mean absolute percentage error, MAE represents the mean absolute error, n represents the total number of samples in the target test sample set, t represents the current sample number, y t ' represents the output running state, y t represents the real running state.

[0098] Optionally, in the present application, the values of MAPE and MAE are not limited, and the values of MAPE and MAE can be set according to actual conditions, for example, when the MAPE is less than 15%, the SDAE is considered feasible.

[0099] In a possible implementation, the SDAE increases the depth relative to the original DAE, improves the learning ability, and is more conducive to training, each layer of the DAE obtains the weight and threshold value that minimizes the reconstruction error after being trained, then the output of the first layer of the DAE is taken as the input of the next layer of the DAE, and finally the new weight and threshold value are obtained after being trained. The network structure with the depth can be obtained through such multiple layer-by-layer training, and the application of the deep or layered architecture can overcome the difficulty in learning highly nonlinear and complex patterns.

[0100] Optionally, the SDAE model automatically extracts relevant features based on the reconstruction idea, and the deep auto-encoding neural network is trained by using the unsupervised layer-by-layer greedy pre-training and supervised fine-tuning method. In the application, the SDAE network structure is set to three hidden layers, and the nodes of the hidden layers are n1=600, n2=200 and n3=100 respectively.

[0101] The DBN is trained according to the target training sample set, and the weight of the DBN is obtained;

[0102] The DBN is a deep graph model, which extracts the features of the input data through the neuron transmission, and then reconstructs the data of the network, and the DBN model structure includes a data input layer, an output layer and a hidden layer. The input layer is a layer for intelligently inputting high-voltage switch data, the hidden layer is mainly used for feature extraction, that is, mapping low-dimensional features to high-dimensional features, and the output layer is the top layer.

[0103] As shown in the DBN network structure diagram shown in Figure 4 According to the application scene of the intelligent high-voltage switch, the softmax algorithm is adopted. The target training sample set is taken as the input data, the input layer and the first hidden layer constitute the first RBM (Restricted Boltzmann machine), the output data of the hidden layer is taken as the new input data and sent to the next hidden layer, the current hidden layer and the next hidden layer constitute the second RBM, and the pre-training is repeatedly performed, and finally the parameter adjustment is performed according to the error BP (BackPropagation) algorithm.

[0104] Optionally, the state of the hidden layer neuron in the network is:

[0105]

[0106] Wherein, v represents the visual layer neuron, h j represents the jth element of the hidden layer neuron, B represents the basis vector of the DBN, P(h j represents the probability that the hidden layer neuron is activated, H represents the vector, and H jdenotes the jth element of vector H, u denotes a probability and obeys a uniform distribution of 0-1, e denotes a mathematical constant, W up denotes the upward weight.

[0107] The DBN model is composed of multiple RBMs and an output layer. The initial weights and biases are obtained by training the RBMs layer by layer, and then the weights and biases of the model are fine-tuned. In the unsupervised training phase of the DBN model training, the weight relationship between neurons and the generation weight is When the upward and downward weights are fixed and satisfy the relationship , the DBN can be regarded as the superposition of multiple RBMs, that is, At this time, W denotes the weight matrix connecting the visible layer neurons and the hidden layer neurons. For each sample in the target training sample set, the input is input into the visible layer, and the RBM1 is trained. After the RBM1 is fully trained, the RBM1 weight is fixed, and then the RBM2 is trained. The training of all RBMs is sequentially carried out, and the pre-processing of the DBN weight is realized.

[0108] Optionally, for the data-driven network of intelligent high-voltage switch, using probability instead of state value to train the RBM network can reduce the overfitting phenomenon. The specific training process is as follows:

[0109] The target training sample set is input into the visible layer neurons, and the probability of the activation of the hidden layer neurons is calculated according to The visible layer neurons and the hidden layer neurons have a symmetrical full connection relationship, wherein v denotes the visible layer neuron, h j denotes the jth element of the hidden layer neuron, v i denotes the ith element of the visible layer neuron, i denotes the current sample number, m denotes the total number of samples in the sample set, W ij denotes the element in the ith row and jth column of the weight matrix connecting the visible layer neurons and the hidden layer neurons, c j denotes the jth element of the hidden layer bias vector;

[0110] A random sample is extracted, the visible layer neurons are reconstructed according to the random sample, and the reconstructed visible layer neurons are obtained. The probability of the activation of the visible layer neurons is calculated according to , wherein v' denotes the ith element of the reconstructed visible layer neuron, h denotes the hidden layer neuron, and b i denotes the ith element of the reconstructed visible layer neuron, h denotes the hidden layer neuron, and b i denotes the ith element of the visible layer bias vector;

[0111] On the basis of the reconstructed visible layer neurons, the probability of the activation of the hidden layer neurons is calculated according to , wherein v' denotes the reconstructed visible layer neuron, and h'j represents the jth element of the hidden layer neurons based on the reconstructed visual layer neurons;

[0112] According to W=W+η(hv T update the weights of the DBN, wherein W represents the weight matrix connecting the visual layer neurons and the hidden layer neurons, η represents the learning rate of the DBN, h' represents the hidden layer neurons based on the reconstructed visual layer neurons, v T represents the transpose of the visual layer neurons.

[0113] Optionally, the network structure of the DBN model is set by training the restricted Boltzmann machine layer by layer through a layer-by-layer greedy training algorithm, and finally forming the model through BP algorithm for network fine-tuning. In the present application, the DBN network structure is set to three hidden layers, and the nodes of the hidden layers are n1=300, n2=100 and n3=100 respectively.

[0114] Optionally, after the deep belief network DBN is trained according to the target training sample set, the weights of the DBN are obtained, and the method further comprises:

[0115] According to and determine the accuracy of the DBN, and test the feasibility of the DBN according to the accuracy;

[0116] Wherein, MAPE represents the mean absolute percentage error, MAE represents the mean absolute error, n represents the total number of samples in the target test sample set, t represents the current sample number, y t ' represents the output running state, y t represents the real running state.

[0117] Optionally, in the present application, the values of MAPE and MAE are not limited, and the values of MAPE and MAE can be set according to actual conditions, for example, when MAPE is less than 15%, it is considered that the DBN is feasible.

[0118] Step 103, the weights of SDAE and DBN are fused by using a weighted fusion method to obtain a fusion weight, the weights of the target evaluation model are initialized according to the fusion weight, and the feasibility of the target evaluation model is tested.

[0119] According to W ds =aW d +bW s to obtain the fusion weight, wherein W ds represents the fusion weight, W s represents the weight of SDAE, W d represents the weight of DBN, and a and b represent the weighting coefficients of DBN and SDAE respectively.

[0120] Optionally, W s and W d may be a number or a matrix, and the type of the two is not limited herein, for example, W s and W s may be a matrix.

[0121] Optionally, a, b ∈ [0, 1], a + b = 1. When selecting the weighting coefficient, the range of the optimal coefficient needs to be determined, and after the determination, the coefficient is refined.

[0122] The target test sample set is input into the target evaluation model, and the running state of each sample in the target test sample set is output;

[0123] According to and , the accuracy of the target evaluation model is determined, and the feasibility of the target evaluation model is tested according to the accuracy;

[0124] Wherein, MAPE represents the mean absolute percentage error, MAE represents the mean absolute error, n represents the total number of samples in the target test sample set, t represents the current sample number, y t ' represents the output running state, y t represents the real running state.

[0125] Optionally, the values of MAPE and MAE are not limited in the present application, and the values of MAPE and MAE can be set according to actual conditions, for example, when MAPE is less than 15%, it is considered that the target evaluation model is feasible.

[0126] Step 104, when the feasibility of the target evaluation model meets the preset condition, the running data of the intelligent high-voltage switch is obtained, the running data is input into the target evaluation model, and the running state of the running data is obtained.

[0127] Optionally, the running state can include a normal state, an attention state, an abnormal state and a serious state. Among them, the normal state corresponds to a normal operation, that is, no maintenance; the attention state corresponds to a planned maintenance; the abnormal state corresponds to a maintenance as soon as possible; and the serious state corresponds to an immediate maintenance.

[0128] The embodiment of the application provides a state evaluation method of an intelligent high-voltage switch, the method comprises the following steps: obtaining a sensing data set of the intelligent high-voltage switch; training a stacked denoising autoencoder (SDAE) and a deep belief network (DBN) according to the sensing data set, obtaining the weight of the SDAE and the weight of the DBN; fusing the weight of the SDAE and the weight of the DBN by using a weighted fusion method, obtaining a fusion weight; initializing the weight of a target evaluation model according to the fusion weight, and testing the feasibility of the target evaluation model; when the feasibility of the target evaluation model meets a preset condition, obtaining operation data of the intelligent high-voltage switch, inputting the operation data into the target evaluation model, and obtaining the operation state of the operation data. The SDAE and the DBN are trained, the weight of the SDAE and the weight of the DBN are obtained, the weight of the SDAE and the weight of the DBN are fused, the weight of the target evaluation model is initialized, the target evaluation model has the properties of the SDAE and the DBN, the characteristics of the SDAE and the DBN are fully utilized, when the operation data is input, the time complexity of the evaluation can be reduced, and the evaluation accuracy can be observed, and the operation state of the operation data can be obtained efficiently and accurately.

[0129] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the application.

[0130] The following is the device embodiment of the application, and for the details not described in detail, reference can be made to the corresponding method embodiments described above.

[0131] Figure 5 The structure schematic diagram of the state evaluation device of the intelligent high-voltage switch provided by the embodiment of the application is shown, only the parts related to the embodiment of the application are shown for the convenience of description, and the details are described as follows:

[0132] As shown in Figure 5 , the state evaluation device 5 of the intelligent high-voltage switch comprises:

[0133] The acquisition module 51 is configured to acquire a sensing data set of the intelligent high-voltage switch.

[0134] The training module 52 is configured to train a stacked denoising autoencoder (SDAE) and a deep belief network (DBN) according to the sensing data set, and obtain the weight of the SDAE and the weight of the DBN.

[0135] The fusion verification module 53 is configured to fuse the weight of the SDAE and the weight of the DBN by using a weighted fusion method, obtain a fusion weight, initialize the weight of a target evaluation model according to the fusion weight, and test the feasibility of the target evaluation model.

[0136] The evaluation module 54 is configured to, when the feasibility of the target evaluation model meets a preset condition, acquire operation data of the intelligent high-voltage switch, input the operation data into the target evaluation model, and obtain an operation state of the operation data.

[0137] In a possible implementation, the training module 52 respectively trains the stack denoising autoencoder SDAE and the deep belief network DBN according to the sensing data set, to obtain the weight of the SDAE and the weight of the DBN, for:

[0138] dividing the sensing data set into an initial training sample set and an initial test sample set;

[0139] respectively performing dimension reduction processing on the initial training sample set and the initial test sample set to obtain a target training sample set and a target test sample set;

[0140] training the stack denoising autoencoder SDAE according to the target training sample set to obtain the weight of the SDAE;

[0141] training the deep belief network DBN according to the target training sample set to obtain the weight of the DBN.

[0142] In a possible implementation, the training module 52 respectively performs dimension reduction processing on the initial training sample set and the initial test sample set, for:

[0143] performing a decentralization operation on the initial training sample set to obtain a first training sample set;

[0144] calculating a covariance matrix of the first training sample set, calculating eigenvalues of the covariance matrix and eigenvectors corresponding to the eigenvalues, sorting the eigenvalues in descending order, calculating a contribution rate and a cumulative contribution rate of the eigenvectors based on the eigenvalues, and selecting eigenvectors with a cumulative contribution rate greater than a first preset value to form a projection matrix;

[0145] multiplying the first training sample set by the projection matrix to obtain a dimension-reduced first training sample set, to realize dimension reduction processing.

[0146] In a possible implementation, the training module 52 trains the stack denoising autoencoder SDAE according to the target training sample set, for:

[0147] S1, dividing the target training sample set into M groups of sub-sample sets as a current sample set, sequentially inputting the M groups of sub-sample sets into a current hidden layer of the SDAE, calculating the weight and the threshold value of the SDAE corresponding to each group of sub-sample sets, and calculating the eigenvectors of the current hidden layer according to the weight and the threshold value of the SDAE corresponding to each group of sub-sample sets;

[0148] S2, taking the feature vector as the current sample set, taking the next layer of hidden layer of the SDAE as the current hidden layer, jumping to step S1 for execution until the maximum number of hidden layers is reached;

[0149] S3, according to fine-tuning the weights and thresholds of the SDAE;

[0150] wherein W represents the weights of the SDAE, B represents the thresholds of the SDAE, r1 and r2 respectively represent the cumulative variables of the weights and the thresholds, w represents the weight gradient, b represents the threshold gradient, x i represents the i-th sample, y i represents x i the corresponding actual value, m represents the total number of samples in the sample set, i represents the current sample number, represents the second derivative symbol, l represents the global learning rate, p represents the decay rate, f represents the relationship function between the weights and the thresholds, and represents the element-wise product symbol.

[0151] In a possible implementation, the training module 52 trains the deep belief network DBN according to the target training sample set to obtain the weights of the DBN, which are used for:

[0152] inputting the target training sample set into the visible layer neurons, and according to calculating the probability of activation of the hidden layer neurons, the visible layer neurons and the hidden layer neurons being in a symmetrical full connection relationship, wherein v represents the visible layer neuron, h j represents the j-th element of the hidden layer neuron, v i represents the i-th element of the visible layer neuron, i represents the current sample number, m represents the total number of samples in the sample set, W ij represents the element in the i-th row and the j-th column of the weight matrix connecting the visible layer neuron and the hidden layer neuron, c j represents the j-th element of the hidden layer bias vector;

[0153] extracting a random sample, reconstructing the visible layer neuron according to the random sample to obtain a reconstructed visible layer neuron, and according to calculating the probability of activation of the visible layer neuron, wherein v i represents the i-th element of the reconstructed visible layer neuron, h represents the hidden layer neuron, b i represents the i-th element of the visible layer bias vector;

[0154] based on the reconstructed visible layer neuron, according to calculating the probability of activation of the hidden layer neuron, wherein v' represents the reconstructed visible layer neuron, h' jrepresents the jth element of the hidden layer neurons based on the reconstructed visual layer neurons;

[0155] According to W = W + η (h'v T update the weights of the DBN, where W represents the weight matrix connecting the visual layer neurons and the hidden layer neurons, η represents the learning rate of the DBN, h' represents the hidden layer neurons based on the reconstructed visual layer neurons, v T represents the transpose of the visual layer neurons.

[0156] In a possible implementation, the fusion verification module 53 fuses the weights of the SDAE and the weights of the DBN by using a weighted fusion method to obtain the fused weights, which are used for:

[0157] According to W ds = aW d +bW s to obtain the fused weights, where W ds represents the fused weights, W s represents the weights of the SDAE, W d represents the weights of the DBN, and a and b respectively represent the weighting coefficients of the DBN and the SDAE.

[0158] In a possible implementation, the fusion verification module 53 tests the feasibility of the target evaluation model, for:

[0159] inputs the target test sample set into the target evaluation model, and outputs the running states of each sample in the target test sample set;

[0160] According to and determines the accuracy of the target evaluation model, and tests the feasibility of the target evaluation model according to the accuracy;

[0161] where MAPE represents the mean absolute percentage error, MAE represents the mean absolute error, n represents the total number of samples in the target test sample set, t represents the current sample number, y t ' represents the output running state, y t represents the real running state.

[0162] This invention provides a state assessment device for intelligent high-voltage switches. The device acquires a sensor dataset of the intelligent high-voltage switch; trains a stacked denoising autoencoder (SDAE) and a deep belief network (DBN) based on the sensor dataset to obtain the weights of the SDAE and DBN respectively; fuses the weights of the SDAE and DBN using a weighted fusion method to obtain fused weights; initializes the weights of the target assessment model based on the fused weights, and tests the feasibility of the target assessment model; when the feasibility of the target assessment model meets preset conditions, acquires the operating data of the intelligent high-voltage switch, inputs the operating data into the target assessment model, and obtains the operating state of the operating data. This invention trains the SDAE and DBN, obtains their weights, fuses the weights, and initializes the weights of the target assessment model, enabling the target assessment model to possess the properties of both the SDAE and DBN. By fully utilizing the features of the SDAE and DBN, and after inputting operating data, it can reduce the time complexity of the assessment and achieve considerable assessment accuracy, efficiently and accurately obtaining the operating state of the operating data.

[0163] Figure 6 This is a schematic diagram of a terminal provided in an embodiment of the present invention. Figure 6 As shown, the terminal 6 in this embodiment includes a processor 60, a memory 61, and a computer program 62 stored in the memory 61 and executable on the processor 60. When the processor 60 executes the computer program 62, it implements the steps in the above-described embodiments of the state evaluation methods for intelligent high-voltage switches, for example... Figure 1 Steps 101 to 104 are shown. Alternatively, when processor 60 executes computer program 62, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 5 The functions of modules / units 51 to 54 shown.

[0164] For example, computer program 62 can be divided into one or more modules / units, one or more of which are stored in memory 61 and executed by processor 60 to complete the present invention. One or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 62 in terminal 6. For example, computer program 62 can be divided into... Figure 5 Modules / units 51 to 54 are shown.

[0165] Terminal 6 may include, but is not limited to, a processor 60 and a memory 61. Those skilled in the art will understand that... Figure 6 This is merely an example of terminal 6 and does not constitute a limitation on terminal 6. It may include more or fewer components than shown, or combine certain components, or different components. For example, the terminal may also include input / output devices, network access devices, buses, etc.

[0166] The processor 60 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0167] The memory 61 can be an internal storage unit of the terminal 6, such as a hard disk or a memory of the terminal 6. The memory 61 can also be an external storage device of the terminal 6, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 61 can include both the internal storage unit and the external storage device of the terminal 6. The memory 61 is used to store computer programs and other programs and data required by the terminal. The memory 61 can also be used to temporarily store data that has been output or is to be output.

[0168] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to complete all or part of the above-described functions. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit, and the integrated unit can be realized in the form of hardware or in the form of software functional unit. In addition, the specific names of each functional unit and module are only for easy distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0169] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0170] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0171] In the embodiments provided by the present application, it should be understood that the disclosed apparatus / terminal and method can be implemented in other ways. For example, the apparatus / terminal embodiments described above are merely schematic. For example, the division of the modules or units is merely a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between the units can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or in other forms.

[0172] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0173] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically independently, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0174] The integrated module / unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various intelligent high-voltage switch state evaluation method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, an executable file, or some intermediate form. The computer readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0175] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features. Such modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for condition assessment of an intelligent high voltage switch, characterized in that, The method comprises the following steps: obtaining a sensing data set of an intelligent high-voltage switch; dividing the sensing data set into an initial training sample set and an initial test sample set; performing dimension reduction processing on the initial training sample set and the initial test sample set respectively to obtain a target training sample set and a target test sample set; training a stacked denoising autoencoder (SDAE) according to the target training sample set to obtain the weight of the SDAE; training a deep belief network (DBN) according to the target training sample set to obtain the weight of the DBN; fusing the weight of the SDAE and the weight of the DBN by using a weighted fusion method to obtain a fusion weight, initializing the weight of a target evaluation model according to the fusion weight, inputting the target test sample set into the target evaluation model, and outputting the running state of each sample in the target test sample set; determining the accuracy of the target evaluation model according to the MAPE and the MAE, and testing the feasibility of the target evaluation model according to the accuracy; when the feasibility of the target evaluation model meets a preset condition, obtaining running data of the intelligent high-voltage switch, inputting the running data into the target evaluation model, and obtaining the running state of the running data; the training of the SDAE according to the target training sample set comprises the following steps: S1, dividing the target training sample set into M groups of sub-sample sets as a current sample set, inputting the M groups of sub-sample sets into a current hidden layer of the SDAE in turn, calculating the weight and threshold value of the SDAE corresponding to each group of sub-sample sets, and calculating the feature vector of the current hidden layer according to the weight and threshold value of the SDAE corresponding to each group of sub-sample sets; S2, taking the feature vector as a current sample set, taking a next layer of hidden layer of the SDAE as a current hidden layer, and jumping to step S1 for execution until the maximum number of hidden layers is reached; S3, according to fine-tuning the weights and thresholds of the SDAE; wherein, W denotes the weights of the SDAE, B denotes the thresholds of the SDAE, , denote the accumulated variables of the weights and thresholds, respectively, denotes the weight gradient, denotes the threshold gradient, denotes the i-th sample, denotes the denotes the corresponding actual value, denotes the total number of samples in the sample set, denotes the current sample number, denotes the second derivative sign, denotes the global learning rate, denotes the weight decay rate, denotes the relationship function between the weights and the thresholds, denotes the element-wise product sign.

2. The method of condition assessment of an intelligent high voltage switch according to claim 1, characterized in that, the dimension reduction processing on the initial training sample set and the initial test sample set comprises the following steps: performing a decentralization operation on the initial training sample set to obtain a first training sample set; calculating a covariance matrix of the first training sample set, calculating the eigenvalue of the covariance matrix and the feature vector corresponding to the eigenvalue, sorting the eigenvalues in descending order, calculating the contribution rate and cumulative contribution rate of the feature vector based on the eigenvalues, and selecting the feature vectors with a cumulative contribution rate greater than a first preset value to form a projection matrix; multiplying the first training sample set by the projection matrix to obtain a first training sample set after dimension reduction, and realizing dimension reduction processing.

3. The method of condition assessment of an intelligent high voltage switch according to claim 1, wherein, the training of the DBN according to the target training sample set to obtain the weight of the DBN comprises the following steps: The target training sample set is input into the visible layer neurons, according to... Calculate the probability of activation of hidden layer neurons, where there is a symmetrical fully connected relationship between the visible layer neurons and the hidden layer neurons. Represents neurons in the visible layer. The first hidden layer neuron represents the first... j One element, The first neuron representing the visible layer i One element, Indicates the current number of samples. This represents the total number of samples in the sample set. The weight matrix representing the connection between visible layer neurons and hidden layer neurons is represented by the [number]th [unit]. i Line number j Column elements, The first hidden layer bias vector represents the... j One element; extracting a random sample, reconstructing the visual layer neurons according to the random sample to obtain reconstructed visual layer neurons, and calculating the probability of the visual layer neurons being activated according to the reconstructed visual layer neurons, wherein represents an element of the reconstructed visual layer neurons, i represents an element of the visual layer bias vector; and h represents a hidden layer neuron, represents an element of the reconstructed visual layer neurons, i represents an element of the visual layer bias vector; and On the basis of the reconstructed visual layer neurons, according to calculating the probability of the activation of the hidden layer neurons, wherein, denotes the reconstructed visual layer neurons, denotes the jth element of the hidden layer neurons on the basis of the reconstructed visual layer neurons; According to updating the weights of the DBN, wherein, denotes a weight matrix connecting the visible layer neurons with the hidden layer neurons, denotes a learning rate of the DBN, denotes the hidden layer neurons based on the reconstructed visible layer neurons, denotes the transpose of the visible layer neurons.

4. The method of condition assessment of an intelligent high voltage switch according to claim 1, wherein, the fusion of the weight of the SDAE and the weight of the DBN by using the weighted fusion method to obtain the fusion weight comprises the following steps: According to obtained fusion weights, wherein, denote the fusion weights, denote the weights of the SDAE, denote the weights of the DBN, a, b denote the weighting coefficients of the DBN and the SDAE, respectively.

5. The method of condition assessment of an intelligent high voltage switch according to claim 1, wherein, the testing of the feasibility of the target evaluation model comprises the following steps: the calculation formula of the MAPE is: the calculation formula of the MAE is: wherein, MAPE denotes the mean absolute percentage error, MAE denotes the mean absolute error, n denotes the total number of samples in the target test sample set, t denotes the current number of samples, denotes the output operational status, denotes the true operational status.

6. A state evaluation device of an intelligent high-voltage switch, characterized by, The method comprises the following steps: an acquisition module configured to obtain a sensing data set of an intelligent high-voltage switch; a training module configured to divide the sensing data set into an initial training sample set and an initial test sample set; The initial training sample set and the initial test sample set are respectively subjected to dimension reduction processing to obtain a target training sample set and a target test sample set; the stack denoising auto-encoder SDAE is trained according to the target training sample set to obtain the weight of the SDAE; The deep belief network DBN is trained according to the target training sample set to obtain the weight of the DBN; The fusion verification module is configured to fuse the weight of the SDAE and the weight of the DBN by using a weighted fusion method to obtain a fusion weight; the weight of the target evaluation model is initialized according to the fusion weight, the target test sample set is input into the target evaluation model, and the running state of each sample in the target test sample set is output; the accuracy of the target evaluation model is determined according to the MAPE and the MAE, and the feasibility of the target evaluation model is tested according to the accuracy; The evaluation module is configured to, when the feasibility of the target evaluation model meets a preset condition, acquire the running data of the intelligent high-voltage switch, input the running data into the target evaluation model, and obtain the running state of the running data. The training module is specifically configured to: S1, divide the target training sample set into M groups of sub-sample sets as a current sample set, input the M groups of sub-sample sets into a current hidden layer of the SDAE in turn, calculate the weight and threshold value of the SDAE corresponding to each group of sub-sample sets, and calculate the feature vector of the current hidden layer according to the weight and threshold value of the SDAE corresponding to each group of sub-sample sets; S2, take the feature vector as a current sample set, take a next layer of hidden layer of the SDAE as a current hidden layer, and jump to step S1 for execution until the maximum number of hidden layers is reached. S3, according to fine-tuning the weights and thresholds of the SDAE; wherein, W denotes the weights of the SDAE, B denotes the thresholds of the SDAE, , denote accumulated variables for the weights and thresholds, respectively, denotes the weight gradient, denotes the threshold gradient, denotes the i-th sample, denotes the denotes the corresponding actual value, denotes the total number of samples in the sample set, denotes the current sample number, denotes the second derivative sign, denotes the global learning rate, denotes the weight decay rate, denotes the relation function between weights and thresholds, denotes the element-wise product sign. 7.A terminal, comprising a memory for storing a computer program and a processor for invoking and running the computer program stored in the memory, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 5.

Images