CN113850320A - Transformer fault detection method based on improved support vector machine regression algorithm - Google Patents
Transformer fault detection method based on improved support vector machine regression algorithm Download PDFInfo
- Publication number
- CN113850320A CN113850320A CN202111118900.2A CN202111118900A CN113850320A CN 113850320 A CN113850320 A CN 113850320A CN 202111118900 A CN202111118900 A CN 202111118900A CN 113850320 A CN113850320 A CN 113850320A
- Authority
- CN
- China
- Prior art keywords
- transformer
- support vector
- vector machine
- model
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Neurology (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a transformer fault detection method based on an improved support vector machine regression algorithm, which optimizes punishment parameters and kernel width factors in a least square support vector machine regression model by adopting a chemical reaction optimization algorithm, adaptively adjusts parameters in the model, predicts transformer data by using the least square support vector machine regression model, generates a point prediction result of sample data, estimates a prediction interval under confidence coefficient, constructs a deep belief network model based on a TrAdaBoost algorithm as a fault diagnosis model to extract the depth characteristics of the transformer parameters, and realizes classification prediction of transformer faults.
Description
Technical Field
The invention belongs to the technical field of transformer fault detection, and particularly relates to a transformer fault detection method based on an improved support vector machine regression algorithm.
Background
The transformer is an important component and a core component of the whole transformer substation system and even a power grid and is responsible for completing transformation and transmission of electric energy. With the construction and development of power grids in China, the operation environment of transformers becomes more complicated, the state monitoring of the transformers becomes more important, and measures must be taken to ensure the good operation state of the transformers and ensure the safe operation of the transformer substations.
The method for detecting the abnormal condition of the transformer at present is to monitor and analyze the temperature of the top layer of the transformer oil. Because the transformer fault diagnosis procedure is complex and time-consuming, a large amount of data needs to be analyzed, and in order to solve the problem, domestic and foreign scholars adopt various artificial intelligence methods to carry out fault diagnosis on the transformer. The existing transformer fault detection technologies have certain defects, and the methods cannot well balance the convenience, sensitivity and cost of monitoring. The support vector machine has the specific advantages in solving the problems of small samples, nonlinearity and high-dimensional pattern recognition, and has good popularization capability under the condition of limited samples. The selection of the parameters of the training model greatly influences the fault diagnosis accuracy rate of the training model. Therefore, it is significant to consider the weighted least squares support vector machine regression algorithm for transformer fault detection.
Disclosure of Invention
Based on the problems, the invention provides a transformer fault detection method based on an improved support vector machine regression algorithm, which adopts a regression algorithm based on a chemical reaction optimization fusion weighted least square support vector machine to ensure the prediction precision, adopts a Chemical Reaction Optimization (CRO) algorithm to optimize a punishment parameter gamma and a kernel width factor sigma of the least square support vector machine regression (WLSSVM), obtains an optimal reaction parameter through iterative optimization of five reaction operators, generates an optimal prediction model to predict transformer data, and inputs the optimal prediction model into a fault diagnosis model of a deep belief network based on a TrAdaBoost algorithm to diagnose the transformer fault.
A transformer fault detection method based on an improved support vector machine regression algorithm comprises the following steps:
step 1: collecting transformer parameters as sample data;
step 2: constructing a weighted least square support vector machine regression model to generate a point prediction result of the sample data and estimating a prediction interval under the confidence coefficient;
and step 3: and constructing a deep belief network model based on the TrAdaBoost algorithm, and classifying the transformer faults according to the point prediction result.
The sample data in the step 1 comprises collecting electric energy parameters and environment parameters of the transformer at each moment in a period of time, wherein the electric energy parameters comprise three-phase current, three-phase voltage and zero-sequence current, and the environment parameters comprise oil temperature, oil level, environment temperature and winding temperature;
the step 2 comprises the following steps:
step 2.1: carrying out normalization processing on the sample data and converting the sample data into a range of [0,1 ];
step 2.2: dividing the normalized sample data into a training set and a test set, and constructing the training set into M pseudo sample sets by adopting a bootstrap method;
step 2.3: constructing a regression model of a weighted least square support vector machine;
step 2.4: optimizing a penalty parameter gamma and a kernel width factor sigma by fusing a chemical reaction optimization algorithm and a weighted least square support vector machine regression algorithm to finally obtain an optimized weighted least square support vector machine regression model;
step 2.5: training a regression model of a weighted least square support vector machine by using M pseudo sample sets, and selecting the model with the minimum error variance as an optimal prediction model;
step 2.6: carrying out interval estimation on a predicted value output by the optimal prediction model by adopting a bootstrap method, and estimating a prediction interval under the confidence coefficient;
step 2.7: and inputting the test data into a prediction model, wherein the obtained prediction value is a point prediction result.
The step 3 comprises the following steps:
step 3.1: constructing a deep belief network model based on a TrAdaBoost algorithm as a fault diagnosis model;
step 3.2: and (4) taking the prediction result obtained in the step (2.7) as the input of the fault diagnosis model to generate a fault classification result of the transformer.
The step 2.3 comprises:
step 2.3.1: given a training data setWherein xk∈RnFor N-dimensional input data, yk∈RnFor outputting data, in the main space RnThe optimization problem is established:
wherein J is a loss function, w ∈ RnIs the weight vector of the original weight space,representing a non-linear mapping function, b ∈ RnIs offset, gamma > 0 is a penalty parameter, ek∈RnIs the amount of error per sample;
step 2.3.2: solving the optimization problem by using a Lagrangian method, and introducing a Lagrangian function:
in the formula, akPartial differentiation of L (w, b, e, a) on w, b, e, a is respectively carried out for the kth Lagrangian, and w and e are respectively carried outkDenoted by a and b, yields:
wherein y is [ y ═ y1,y2,…,yN]T,I=[1,1,…,1]T,a=[a1,a2,…,aN]T,Ω∈RN×N,i,j=1,2,...,N,K(xi,xj) Is a kernel function, selecting an RBF kernel function:
in the formula, sigma represents the distance from the center of the RBF kernel function and is called a kernel width factor;
step 2.3.3: introducing a weighting factor vk(k 1.., N) correcting a residual vector e obtained by regression of a least squares support vector machine, and vkThe determination method of (2) is as follows:
wherein, residual ekObtained by a least squares support vector machine,for measuring the difference between the distribution of residual vector e and Gaussian distribution, where IQR is ekA quarter-bit pitch of c1、c2Is a constant.
The step 2.4 comprises:
step 2.4.1: initializing parameters of a chemical reaction optimization algorithm, and setting initial value ranges of penalty parameters and kernel width factors;
step 2.4.2: inputting a molecular structure consisting of a penalty parameter and a kernel width factor into a chemical reaction optimization algorithm, wherein the molecular structure is iteratively optimized through five chemical reaction operators, and each iteration process needs to be subjected to three judgment processes of a reaction type, a single-molecule reaction type and an intermolecular reaction type;
step 2.4.3: and stopping iteration when the maximum iteration times are reached, obtaining a molecular structure with the minimum molecular potential energy, namely a global optimal solution, and simultaneously distributing corresponding punishment parameters and kernel width factors to the weighted least square support vector machine regression model.
The step 3.1 comprises the following steps:
step 3.1.1: acquiring training sample data, acquiring sample data of a transformer in the transformer area of 5, taking one transformer area as a target area, taking the other four transformer areas as source areas, averagely dividing the data of the target area into five parts, selecting four-fifths of the data as a training set, and taking one-fifth of the data as a test set;
step 3.1.2: input source domain sample set DsTarget Domain tagged sample set DTA target field label-free sample set T and iteration times N';
step 3.1.3: initializing weight vector W ═ W'1,w′2,…,w′m+n) Weight change factor Wherein n is the number of source domain samples, m is the number of target domain samples, w'iFor the weight of each sample data, i ═ 1, 2, …, m + n;
step 3.1.4: by using Ds、DTCompleting the greedy training process of the depth belief network layer by layer and updating and calculating new sample weightCalculating a weight cost function hlxi is the classification result of the input value xi, and yxi is the label value;
step 3.1.5: performing reverse fine adjustment, and performing deep belief networkThe complex pair input value xiIs classified with its label value y (x)i) Equal, h (x)i)==yiIs 1, and updates the sample weight wiThe updating process is as follows; otherwise h (x)i)==yiIs 0;
step 3.1.6: when in useWhen x is greater than xiDeleted from Ds, where θ ═ { W, a ', b' } is a parameter of the restricted boltzmann model, a 'denotes the bias vector for the visible layer, b' denotes the bias vector for the hidden layer;
step 3.1.7: stopping training when iteration reaches the maximum number, and outputting a final classifier h (x)i) And as a fault diagnosis model, testing the fault diagnosis model by using a test set.
The invention has the beneficial effects that:
the invention provides a transformer fault detection method based on an improved support vector machine regression algorithm, aiming at the problems of small sample, high dimension and nonlinearity of fault parameter prediction, the method can quickly and accurately predict the parameter value in the future period; aiming at the uncertainty caused by model errors and data noise errors, a bootstrap method is adopted to construct a prediction interval, so that the uncertainty degree of a prediction result can be effectively reflected; meanwhile, a deep belief network based on transfer learning is established as a fault diagnosis model, so that the classification efficiency of the classifier is greatly improved, and the acquisition cost is also reduced.
Drawings
FIG. 1 is a flow chart of a transformer fault detection method based on an improved support vector machine regression algorithm in the present invention;
FIG. 2 is a graph of historical O TI data collected in the present invention;
FIG. 3 is a graph comparing MSE indicators for different algorithms according to the present invention;
FIG. 4 is a graph of MAE index contrast for different algorithms in the present invention;
FIG. 5 is a diagram of the predicted results of OTI time series in the present invention;
FIG. 6 is a graph of the three-layer TL-DBN fault identification accuracy based on transfer learning under different numbers of auxiliary samples in the present invention.
Detailed Description
The invention is further described with reference to the following figures and specific examples. The method adopts a bootstrap method to construct a data set and a prediction interval, obtains the penalty parameters and the kernel width factor of the optimal weighted least square support vector machine through a chemical reaction optimization algorithm, and further obtains the optimal prediction model and the prediction interval. And inputting the test data into the optimal prediction model to obtain the predicted values of the 11 transformer parameters, and inputting the prediction results into the fault diagnosis model to carry out fault diagnosis.
A transformer fault detection method based on an improved support vector machine regression algorithm comprises the steps of firstly, collecting sample data of a transformer; then generating a point prediction result of the sample data, and classifying the transformer faults according to the point prediction result; as shown in fig. 1, includes:
step 1: collecting transformer parameters as sample data; the sample data comprises electric energy parameters and environment parameters of the transformer at each moment within a period of time, wherein the electric energy parameters comprise three-phase current, three-phase voltage and zero-sequence current, and the environment parameters comprise oil temperature, oil level, environment temperature and winding temperature;
training data and test samples used by the model were taken from transformer data provided by a company at the equipment communication address of 9C58C1040000 in the left 3# platform area of PMS two-weft nail line 15. The time series of data samples was 2019-12-01 to 2020-07-01, with a monitoring period of 4 times per hour. The monitoring data from 12/1/2019 to 4/2020/30 are selected as training samples, and the monitoring data from 6/1/2020 to 7/1/2020 are selected as test samples. The original data of the oil temperature indicator OTI is shown in fig. 2.
Step 2: constructing a weighted least square support vector machine regression model to generate a point prediction result of the sample data and estimating a prediction interval under the confidence coefficient; the method comprises the following steps:
step 2.1: carrying out normalization processing on the sample data and converting the sample data into a range of [0,1 ];
step 2.2: dividing the normalized sample data into a training set and a test set, and constructing the training set into M pseudo sample sets by adopting a bootstrap method;
step 2.3: constructing a regression model of a weighted least square support vector machine; the method comprises the following steps:
step 2.3.1: given a training data setWherein xk∈RnFor N-dimensional input data, yk∈RnFor outputting data, in the main space RnThe optimization problem is established:
wherein J is a loss function, w ∈ RnIs the weight vector of the original weight space,representing a non-linear mapping function, b ∈ RnIs offset, gamma > 0 is a penalty parameter, ek∈RnIs the amount of error per sample;
step 2.3.2: solving the optimization problem by using a Lagrangian method, and introducing a Lagrangian function:
in the formula, akPartial differentiation of L (w, b, e, a) on w, b, e, a is respectively carried out for the kth Lagrangian, and w and e are respectively carried outkDenoted by a and b, yields:
wherein y is [ y ═ y1,y2,…,yN]T,I=[1,1,…,1]T,a=[a1,a2,…,aN]T,Ω∈RN×N,i,j=1,2,...,N,K(xi,xj) Is a kernel function, selecting an RBF kernel function:
in the formula, sigma represents the distance from the center of the RBF kernel function and is called a kernel width factor;
step 2.3.3: introducing a weighting factor vk(k 1.., N) correcting a residual vector e obtained by regression of a least squares support vector machine, and vkThe determination method of (2) is as follows:
wherein, residual ekObtained by a least squares support vector machine,for measuring the difference between the distribution of residual vector e and Gaussian distribution, where IQR is ekA quarter-bit pitch of c1、c2Is a constant;
step 2.4: fusing a chemical reaction optimization algorithm with a weighted least square support vector machine regression algorithm (the fused model is called as a CRO-WLSSVM model for short), optimizing a penalty parameter gamma and a kernel width factor sigma, and finally obtaining an optimized weighted least square support vector machine regression model; the method comprises the following steps:
step 2.4.1: initializing parameters of a chemical reaction optimization algorithm, setting initial value ranges of penalty parameters and kernel width factors, wherein the specific parameters are detailed in a table 1, and the initial values of the parameters are set in a table 2;
TABLE 1 parameter Table in chemical reaction optimization Algorithm
TABLE 2 initial CrO-WLSSVM Key parameters
Step 2.4.2: inputting a molecular structure consisting of penalty parameters and a kernel width factor into a chemical reaction optimization algorithm, wherein the molecular structure is iteratively optimized through five chemical reaction operators, basic chemical reactions are divided into monomolecular reactions and bimolecular reactions, and the monomolecular reactions comprise oxidation reduction and decomposition; bimolecular reactions including synthesis, redox, displacement; each iteration process needs three judgment processes of a reaction type, a single-molecule reaction type and an intermolecular reaction type;
step 2.4.3: and stopping iteration when the maximum iteration times are reached, obtaining a molecular structure with the minimum molecular potential energy, namely a global optimal solution, and simultaneously distributing corresponding punishment parameters and kernel width factors to the weighted least square support vector machine regression model. The original data set of the transformer is adopted to train the CRO-WLSSVM model, and the obtained optimal parameter gamma is 50.848, and sigma is 0.0945.
Step 2.5: training a regression model of a weighted least square support vector machine by using M pseudo sample sets, and selecting the model with the minimum error variance as an optimal prediction model;
the parameter prediction is more accurate by adopting the CRO-WLSSVM model, and in order to verify the optimization of the improved algorithm CRO-WLSSVM, the parameter setting when the CRO is applied to the optimization of the WLSSVM is analyzed by taking an XBTUSD data set from BitMex as an example. The data set was divided into training samples and test samples in a 3: 1 ratio. The setting parameter sets are InitialKE 1000, KELossRate 0.01, moleclol 0.6, α 50, and β 700. The number of termination iterations and the number of initial molecules of the chemical reaction optimization algorithm are defined as 50 and 10. The results obtained are shown in table 3:
TABLE 3 CRO-WLSSVM Experimental data results
For the data sets, the CRO-WLSSVM method provided by the invention is respectively used for carrying out comparison experiments with a GA-SVM and a PSO-SVM, a genetic algorithm is used for optimizing a support vector machine, the GA-SVM is abbreviated as the GA-SVM, and a particle swarm algorithm is used for optimizing the support vector machine, the PSO-SVM is abbreviated as the PSO-SVM. Learning factor c 1-c 2-1.5 and inertia factor w-0.75 in the PSO-SVM; the number of GA-SVN populations is set to be 5; the comparison experiment was performed using the Mean Absolute Error (MAE) and Mean Square Error (MSE) of the evaluation index, respectively. Through simulation test, the comparison results of the time series prediction experiments are shown in fig. 3 and fig. 4;
step 2.6: carrying out interval estimation on a predicted value output by the optimal model by adopting a bootstrap method, estimating a prediction interval under 95% confidence coefficient, wherein the prediction interval can effectively reflect the uncertainty degree of a prediction result, and thus, the potential risk is found;
the bootstrap method belongs to one of the non-parameter estimation methods in statistics, and the method adopts a method that the original sample set is repeatedly sampled, the sampling probability is equal, and the collected data is put back into the sample set again. The method has good solving effect on statistical problems of interval estimation, hypothesis test and the like of the parameters. The bootstrap method can preprocess the original time sequence data by taking a block as a unit to generate a plurality of new sub-samples, thereby avoiding the failure of the common method to the maximum extent and obtaining the prediction effect which is closer to the real distribution.
The principle of construction of the prediction interval is summarized as follows:
let x beiIs an input variable, tiIs a real sample value, a prediction modelWith the actual value tiThe error between can be described as:
wherein y (x)i) Is the true regression value of the prediction model,is the total error between the actual and predicted values, ε (x)i) Is data noise.
Prediction model y (x)i) The actual output value is the average of all the prediction model results and can be described by the following formula:
in the formulaThe predicted value is obtained by the ith data set and the corresponding CRO-WLSSVM model, and M is the sampling number.Is the error variance of the prediction model, described as:
after the model is trained, the significance level is alpha, and the prediction interval of the state parameters is as follows:
wherein z is1-α/2Is the 1-alpha/2 percentile of the standard normal distribution.
Step 2.7: the test data is input into the prediction model, and the prediction model can effectively predict the change process and the development trend of the oil top layer temperature. The uncertainty of the result can be analyzed according to the interval. When the interval width is increased, the predicted value of the transformer oil temperature indicator shows risks, and the reasons of changes, such as external interference, potential faults of the transformer, problems of the model or other conditions, should be analyzed in time. The point prediction and interval prediction results are obtained as shown in FIG. 5;
the Deep Belief Network (DBN) is composed of a plurality of layers of superposed Restricted Boltzmann Machine (RBM) networks and a last layer of logistic regression layer. The training process of the DBN network includes two processes: pre-training and fine-tuning. Firstly, the RBM is subjected to unsupervised pre-training layer by layer, and characteristic information implied by data is deeply mined. Only one layer of RBM is trained in each training, after the training is finished, a next layer of RBM is trained and stacked, and then supervised tuning is performed by using a back propagation algorithm by combining labels of samples, so that the classification effect is achieved.
The connection between RBM neurons is characterized in that: there is no connection between adjacent neurons in the layer, and the neurons in the layers are all connected. The RBM is an energy-based probability model, so an RBM energy function needs to be defined first, and then a corresponding distribution function is introduced through the defined energy function. W is a weight matrix connecting between visible layer neurons and hidden layer neurons. n isv,nhRepresenting the number of neurons of the visible layer and the hidden layer, respectively. v ═ v1,v2,...,vn]TRepresenting the state vector of the visible layer, wherein the state of the ith neuron is denoted by viAnd (4) showing. h ═ h1,h2,...,hn0]TState vector representing hidden layer, where the state of jth neuron is represented by hjAnd (4) showing. a ═ a1,a2,...,an]TA bias vector representing the visible layer, wherein the state of the ith neuron is represented by aiAnd (4) showing. b ═ b1,b2,...,bn]TA bias vector representing the hidden layer, wherein the state of the jth neuron is represented by bjAnd (4) showing. The RBM energy function E (v, h) is calculated as
The form of multiplication by a matrix can also be expressed as
Eθ(v,h)=-aTv-bTh-hTWv (11)
The joint probability distribution of the nodes in the visible layer and the hidden layer can be expressed as follows according to the energy function definition of equation (11):
in the formulae (13) and (14), θ ═ wij,ai,bj:1≤i≤nv,1≤j≤nh}; z (-) is a partition function, which acts as a normalization factor in the energy function.
The purpose of adjusting parameters of the RBM model is to enable the generated data to be consistent with the actual training data sample distribution, and to realize the maximum fitting of sample data. The goal of training the RBM is therefore a maximum likelihood function equation (15):
in the formula (15), N is the number of samples. And maximizing the likelihood function by adopting a gradient ascending algorithm, and continuously updating the parameters in an iterative mode. Iterate the formula of
In the formula (16), α > 0 and α is a learning rate. The partial derivatives of the log-likelihood function with respect to each parameter are
In the formulas (17) to (19), a summation operation with exponential complexity is required, the calculation difficulty is high, and the training efficiency is low. The log-likelihood gradient of the RBM is quickly calculated using a Contrast Divergence (CD) algorithm.
And step 3: constructing a deep belief network model based on a TrAdaBoost algorithm, and classifying transformer faults according to point prediction results; the method comprises the following steps:
step 3.1: constructing a deep belief network model (TL-DBN model for short) based on a TrAdaBoost algorithm as a fault diagnosis model; the method comprises the following steps:
step 3.1.1: acquiring training sample data, acquiring sample data of a transformer in the transformer area of 5, taking one transformer area as a target area, taking the other four transformer areas as source areas, averagely dividing the data of the target area into five parts, selecting four-fifths of the data as a training set, and taking one-fifth of the data as a test set;
the data of 5 transformer areas are obtained in total, the left 3# of the PMS _ second weft line 15 is used as a target field, the right 4 and left 1# of the PMS _ second weft line 25, the right 5# of the PMS _ second weft line 25, the left 2# of the PMS _ second weft line 9 and the right 4# of the PMS _ second weft line 19 are used as a source field, and the data of the target field are divided into a training set and a testing set according to the ratio of 4: 1.
Step 3.1.2: input source domain sample set DsTarget Domain tagged sample set DTA target field label-free sample set T, and training times are N;
step 3.1.3: initializing weight vector W ═ W'1,w′2,…,w′m+n) Weight change factorWherein n is the number of source domain samples, m is the number of target domain samples, w'iFor the weight of each sample data, i ═ 1, 2, …, m + n;
step 3.1.4: by using Ds、DTCompleting the greedy training process of the depth belief network layer by layer and updating and calculating new sample weightCalculating a weight cost function hlxi is the classification result of the input value xi, and yxi is the label value;
step 3.1.5: performing reverse fine adjustment when the deep belief network inputs the value xiIs classified with its label value y (x)i) Equal, h (x)i)==yiIs 1, and updates the sample weight wiThe updating process is as follows; otherwise h (x)i)==yiIs 0;
step 3.1.6: when in useWhen x is greater than xiFrom DsWhere θ ═ 0.1 is a parameter of the restricted boltzmann model, a 'denotes the bias vector for the visible layer, and b' denotes the bias vector for the hidden layer;
step 3.1.7: stopping training when iteration reaches the maximum number, and outputting a final classifier h (x)i) As a fault diagnosis model, testing the fault diagnosis model by using a test set;
step 3.2: and (3) taking the prediction result obtained in the step (2.7) as the input of a fault diagnosis model, generating a fault classification result of the transformer, and finally obtaining oil leakage fault, oil level fault, temperature difference fault, iron core insulation fault, winding deformation fault and normal six transformer state quantities.
In order to verify the effectiveness of classification of the deep belief network model (TL-DBN model) based on the TrAdaBoost algorithm, the deep belief network model and the TL-DBN model are respectively subjected to a comparison experiment, and the average percentage error result of fault diagnosis and evaluation obtained by the experiment is shown in Table 4; parameter beta in TL-DBN model tests=0.9,βT1.3, the threshold value theta is 0.1, the label layer selects softmax, and the BP training time is 100; the experimental results of the model are shown in fig. 6.
TABLE 4 Fault diagnosis of two sets of test transformers
Claims (7)
1. A transformer fault detection method based on an improved support vector machine regression algorithm is characterized by comprising the following steps:
step 1: collecting transformer parameters as sample data;
step 2: constructing a weighted least square support vector machine regression model to generate a point prediction result of the sample data and estimating a prediction interval under the confidence coefficient;
and step 3: and constructing a deep belief network model based on the TrAdaBoost algorithm, and classifying the transformer faults according to the point prediction result.
2. The method according to claim 1, wherein the sample data in step 1 includes collecting electric energy parameters and environmental parameters of the transformer at each moment in a period of time, the electric energy parameters include three-phase current, three-phase voltage and zero-sequence current, and the environmental parameters include oil temperature, oil level, environmental temperature and winding temperature.
3. The method for detecting the fault of the transformer based on the regression algorithm of the improved support vector machine according to claim 1, wherein the step 2 comprises the following steps:
step 2.1: carrying out normalization processing on the sample data and converting the sample data into a range of [0,1 ];
step 2.2: dividing the normalized sample data into a training set and a test set, and constructing the training set into M pseudo sample sets by adopting a bootstrap method;
step 2.3: constructing a regression model of a weighted least square support vector machine;
step 2.4: optimizing a penalty parameter gamma and a kernel width factor sigma by fusing a chemical reaction optimization algorithm and a weighted least square support vector machine regression algorithm to finally obtain an optimized weighted least square support vector machine regression model;
step 2.5: training a regression model of a weighted least square support vector machine by using M pseudo sample sets, and selecting the model with the minimum error variance as an optimal prediction model;
step 2.6: carrying out interval estimation on a predicted value output by the optimal prediction model by adopting a bootstrap method, and estimating a prediction interval under the confidence coefficient;
step 2.7: and inputting the test data into a prediction model, wherein the obtained prediction value is a point prediction result.
4. The method for detecting the fault of the transformer based on the regression algorithm of the improved support vector machine as claimed in claim 3, wherein said step 3 comprises:
step 3.1: constructing a deep belief network model based on a TrAdaBoost algorithm as a fault diagnosis model;
step 3.2: and (4) taking the prediction result obtained in the step (2.7) as the input of the fault diagnosis model to generate a fault classification result of the transformer.
5. The method for detecting the fault of the transformer based on the regression algorithm of the improved support vector machine as claimed in claim 3, wherein said step 2.3 comprises:
step 2.3.1: given a training data setWherein xk∈RnFor N-dimensional input data, yk∈RnFor outputting data, in the main space RnThe optimization problem is established:
wherein J is a loss function, w ∈ RnIs the weight vector of the original weight space,representing a non-linear mapping function, b ∈ RnIs offset, gamma > 0 is a penalty parameter, ek∈RnIs the amount of error per sample;
step 2.3.2: solving the optimization problem by using a Lagrangian method, and introducing a Lagrangian function:
in the formula, akPartial differentiation of L (w, b, e, a) on w, b, e, a is respectively carried out for the kth Lagrangian, and w and e are respectively carried outkDenoted by a and b, yields:
wherein y is [ y ═ y1,y2,...,yN]T,I=[1,1,...,1]T,a=[a1,a2,...,aN]T,Ω∈RN×N,K(xi,xj) Is a kernel function, selecting an RBF kernel function:
in the formula, sigma represents the distance from the center of the RBF kernel function and is called a kernel width factor;
step 2.3.3: introducing a weighting factor vk(k 1.., N) correcting a residual vector e obtained by regression of a least squares support vector machine, and vkThe determination method of (2) is as follows:
6. The method for detecting the fault of the transformer based on the regression algorithm of the improved support vector machine according to claim 3, wherein the step 2.4 comprises the following steps:
step 2.4.1: initializing parameters of a chemical reaction optimization algorithm, and setting initial value ranges of penalty parameters and kernel width factors;
step 2.4.2: inputting a molecular structure consisting of a penalty parameter and a kernel width factor into a chemical reaction optimization algorithm, wherein the molecular structure is iteratively optimized through five chemical reaction operators, and each iteration process needs to be subjected to three judgment processes of a reaction type, a single-molecule reaction type and an intermolecular reaction type;
step 2.4.3: and stopping iteration when the maximum iteration times are reached, obtaining a molecular structure with the minimum molecular potential energy, namely a global optimal solution, and simultaneously distributing corresponding punishment parameters and kernel width factors to the weighted least square support vector machine regression model.
7. The method for detecting the fault of the transformer based on the regression algorithm of the improved support vector machine as claimed in claim 4, wherein said step 3.1 comprises:
step 3.1.1: acquiring training sample data, acquiring sample data of a transformer in the transformer area of 5, taking one transformer area as a target area, taking the other four transformer areas as source areas, averagely dividing the data of the target area into five parts, selecting four-fifths of the data as a training set, and taking one-fifth of the data as a test set;
step 3.1.2: input source domain sample set DsTarget Domain tagged sample set DTA target field label-free sample set T and iteration times N';
step 3.1.3: initializing weight vector W ═ W'1,w′2,…,w′m+n) Weight change factor Wherein n is the number of source domain samples, m is the number of target domain samples, w'iFor the weight of each sample data, i ═ 1, 2, …, m + n;
step 3.1.4: by using Ds、DTCompleting the greedy training process of the depth belief network layer by layer and updating and calculating new sample weightCalculating a weight cost function hl(xi) Is an input value xiThe classification result of (2), y (x)i) Is a tag value;
step 3.1.5: performing reverse fine adjustment when the deep belief network inputs the value xiIs classified with its label value y (x)i) Equal, h (x)i)==yiIs 1, and updates the sample weight wiThe updating process is as follows; otherwise h (x)i)==yiIs 0;
step 3.1.6: when in useWhen x is greater than xiFrom DsWhere θ ═ { W, a ', b' } is a parameter of the restricted boltzmann model, a 'denotes a bias vector for the visible layer, and b' denotes a bias vector for the visible layerA bias vector of the hidden layer;
step 3.1.7: stopping training when iteration reaches the maximum number, and outputting a final classifier h (x)i) And as a fault diagnosis model, testing the fault diagnosis model by using a test set.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111118900.2A CN113850320A (en) | 2021-09-24 | 2021-09-24 | Transformer fault detection method based on improved support vector machine regression algorithm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111118900.2A CN113850320A (en) | 2021-09-24 | 2021-09-24 | Transformer fault detection method based on improved support vector machine regression algorithm |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113850320A true CN113850320A (en) | 2021-12-28 |
Family
ID=78979608
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111118900.2A Pending CN113850320A (en) | 2021-09-24 | 2021-09-24 | Transformer fault detection method based on improved support vector machine regression algorithm |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113850320A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114326406A (en) * | 2021-12-31 | 2022-04-12 | 山西世纪中试电力科学技术有限公司 | Coordination control method based on vector machine online identification prediction |
CN114936669A (en) * | 2022-04-06 | 2022-08-23 | 武汉大学 | Hybrid ship rolling prediction method based on data fusion |
CN117574141A (en) * | 2024-01-17 | 2024-02-20 | 朗峰新材料启东有限公司 | Wireless charging equipment fault detection method and system based on artificial intelligence |
-
2021
- 2021-09-24 CN CN202111118900.2A patent/CN113850320A/en active Pending
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114326406A (en) * | 2021-12-31 | 2022-04-12 | 山西世纪中试电力科学技术有限公司 | Coordination control method based on vector machine online identification prediction |
CN114936669A (en) * | 2022-04-06 | 2022-08-23 | 武汉大学 | Hybrid ship rolling prediction method based on data fusion |
CN114936669B (en) * | 2022-04-06 | 2024-05-17 | 武汉大学 | Mixed ship rolling prediction method based on data fusion |
CN117574141A (en) * | 2024-01-17 | 2024-02-20 | 朗峰新材料启东有限公司 | Wireless charging equipment fault detection method and system based on artificial intelligence |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wang et al. | A critical review of improved deep learning methods for the remaining useful life prediction of lithium-ion batteries | |
CN110263866B (en) | Power consumer load interval prediction method based on deep learning | |
CN109635928B (en) | Voltage sag reason identification method based on deep learning model fusion | |
CN113850320A (en) | Transformer fault detection method based on improved support vector machine regression algorithm | |
CN112051481B (en) | Alternating current-direct current hybrid power grid fault area diagnosis method and system based on LSTM | |
CN113702895B (en) | Online quantitative evaluation method for error state of voltage transformer | |
CN111339712A (en) | Method for predicting residual life of proton exchange membrane fuel cell | |
CN111965476B (en) | Low-voltage diagnosis method based on graph convolution neural network | |
CN111768000A (en) | Industrial process data modeling method for online adaptive fine-tuning deep learning | |
CN112180204A (en) | Power grid line fault diagnosis method based on electric quantity information | |
CN105825271A (en) | Satellite fault diagnosis and prediction method based on evidential reasoning (ER) | |
CN110866366A (en) | XGboost algorithm-based island detection method for photovoltaic microgrid containing PHEV | |
CN112507479B (en) | Oil drilling machine health state assessment method based on manifold learning and softmax | |
CN115015760B (en) | Lithium battery health state assessment method based on neural network and migration integrated learning | |
CN115358347B (en) | Method for predicting remaining life of intelligent electric meter under different subsystems | |
CN113537469B (en) | Urban water demand prediction method based on LSTM network and Attention mechanism | |
CN113917334A (en) | Battery health state estimation method based on evolution LSTM self-encoder | |
CN116679211A (en) | Lithium battery health state prediction method | |
CN108520310A (en) | Wind speed forecasting method based on G-L mixed noise characteristic v- support vector regressions | |
Xue et al. | Estimating state of health of lithium-ion batteries based on generalized regression neural network and quantum genetic algorithm | |
CN115561005A (en) | Chemical process fault diagnosis method based on EEMD decomposition and lightweight neural network | |
CN112149896A (en) | Attention mechanism-based mechanical equipment multi-working-condition fault prediction method | |
CN116109039A (en) | Data-driven anomaly detection and early warning system | |
Wang et al. | An efficient state-of-health estimation method for lithium-ion batteries based on feature-importance ranking strategy and PSO-GRNN algorithm | |
CN112014757A (en) | Battery SOH estimation method integrating capacity increment analysis and genetic wavelet neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |