Transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost

Abstract

The invention discloses a transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost, and the method comprises the following steps: calculating an out-of-bag error rate through a random forest, screening out a gas combination which is most beneficial to the improvement of diagnosis precision, carrying out the optimization of parameters of a double-hidden-layer back propagation network through an improved particle swarm algorithm, and carrying out the optimization of the parameters of the double-hidden-layer back propagation network. And an Adaboost algorithm is used to improve the robustness of the IPSO-BP diagnosis model. According to the method, four strategies of SPM chaotic mapping based on reverse learning, adaptive inertia weight, spiral flight search and crisscross variation are innovatively introduced to improve the particle swarm optimization. According to the method, the problems that a traditional particle swarm algorithm is prone to falling into local optimum, low in convergence speed and insufficient in population diversity are solved; the defect that a traditional BPNN network is sensitive to the weight and bias of neurons is overcome; the accuracy of the diagnosis result is greatly improved, and the stability of the model diagnosis result is ensured.

Description

Technical Field

[0001] The invention relates to the field of oil-immersed power transformer fault monitoring, and in particular to a transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost. Background Art

[0002] The dissolved gas analysis method used in traditional transformer fault diagnosis is mainly based on the fact that when the insulation material fails or ages, it will produce H 2 , CH 4 , C 2 H 6 , C 2 H 4 and C 2 H 2 The relative content of these gases reflects the severity of the transformer failure. The characteristic gases dissolved in the insulating oil of the power transformer can be analyzed. However, the disadvantage of the traditional method is that it relies on expert experience and observation of equipment parameters. With the introduction of artificial intelligence technology, diagnostic models such as support vector machines, multi-layer perceptrons, and extreme learning machines have begun to be applied to transformer fault diagnosis. However, there is no recognized feature set for diagnosing power transformer faults. In previous studies, there are many types of gas feature quantities and insufficient analysis of fault-related features. This not only greatly increases the operating pressure of the diagnostic model, but also causes the problem of low diagnostic accuracy. Secondly, the parameters of the diagnostic model itself are also highly random, which makes it difficult to further improve the diagnostic accuracy.

[0003] Therefore, a new method is needed to effectively screen the feature set and select a suitable optimization algorithm to optimize the parameters of the traditional diagnostic model to improve the accuracy of the diagnostic model in transformer fault detection. Summary of the invention

[0004] Purpose of the invention: In view of the above shortcomings, this paper proposes a transformer fault diagnosis model based on feature selection and IPSO-BP-Adaboost. An optimal feature set is established according to the different detection accuracies caused by different input features, and the random forest (RF) feature selection method is used to reduce the bias in feature selection and improve the effectiveness of diagnosis. A multi-strategy improved particle swarm combined with back propagation network (IPSO-BPNN) is used to identify the best feature subset and optimize the key parameters of the classifier at the same time. Finally, Adaboost is used to enhance the robustness and generalization of the model, and real fault data is used for detection, which proves that the scheme has high practical value.

[0005] Technical solution: To solve the above problems, the present invention adopts a transformer fault diagnosis based on feature selection and improved IPSO-BP-Adaboost, including the following steps:

[0006] Step S1, select five gases as original characteristic quantities, and combine gas concentration, three ratio method and uncoded ratio to obtain a total of 26 groups of gas characteristics;

[0007] Step S2, using a random forest algorithm to select the gas combination that is most conducive to improving the diagnostic accuracy;

[0008] Step S3, using four strategies, namely, SPM chaos mapping based on reverse learning, adaptive inertia weight, spiral flight search, and vertical and horizontal cross mutation, to improve the particle swarm algorithm;

[0009] Step S4, using the improved particle swarm algorithm to optimize the weights and biases in the BPNN network;

[0010] Step S5, using Adaboost to improve the robustness of the IPSO-BP network;

[0011] Further, step S1 includes the following steps:

[0012] Step S11, select H 2 , CH 4 , C 2 H 6 , C 2 H 4 and C 2 H 2 As the original feature quantity;

[0013] In step S12, a total of 26 groups of gas characteristics are obtained by combining gas concentration, three-ratio method and uncoded ratio.

[0014] Further, step S2 includes the following steps:

[0015] Step S21, according to the number of samples and features in the data set, the minimum number of leaves of the random forest is set to 8, and the number of decision trees is set to 400;

[0016] Step S22, calculating the importance of each gas combination using the error rate of out-of-bag data of random forest.

[0017] Further, step S3 includes the following steps:

[0018] Step S31, setting and initializing the parameters in the genetic particle swarm, the parameters include: maximum number of iterations t max , population size N, dimension D, upper and lower limits of speed and position, initial value w of inertia weight, learning factor C 1,C 2 , the selection probability p of the search method m and search step length k;

[0019] Step S32: Use the SPM function to map and generate N solutions, and use reverse learning to generate N reverse solutions, calculate the fitness of a total of 2N solutions, and select the top N solutions in terms of fitness;

[0020] Step S33: Calculate the inertia weight of the current iteration cycle and the fitness values ​​of all particles to obtain the individual optimal pbest and global optimal gbest of the particles;

[0021] Step S34: According to the selection probability p m Choose to perform a standard PSO search or a spiral search;

[0022] Step S35: Calculate the speed and position of the particle according to the search method;

[0023] Step S36: performing a lateral crossover operation on the particle with an odd index and the particle after it, and performing a lateral crossover operation on the particle with an even index and the particle before it;

[0024] Step S37: According to the dimensions of each particle, two different dimensions are randomly selected to perform a vertical cross operation;

[0025] Step S38: Calculate the fitness value according to the latest position of each particle and update pbest and gbest;

[0026] Step S39: Determine t, if t=t max , then stop searching and output the obtained gbest and its fitness value, that is, the minimum value of the objective function; otherwise, return to step S33.

[0027] On the other hand, the SMP chaos mapping function used in the particle swarm initialization population is constructed as:

[0028]

[0029] Among them, x(i+1) represents the (i+1)th particle, and r is a random number between 0 and 1. η∈(0,1), and when μ∈(0,1), the function is in a chaotic state, and according to the upper and lower bounds U of the optimized weights and biases b and L b , determine the generated initial solution.

[0030] Further, in step S35, the speed update formula when spiral search is adopted is:

[0031] x(t+1)=D′·e cl ·cos(2πl)+x* (t+1)

[0032] where x * (t+1) is the position of the particle in the last iteration, D' is the distance between the current particle and the global optimal particle, l is a random number between (-1,1), and C is a parameter of a spiral search shape, which is calculated by the formula:

[0033] In another aspect, the two horizontal cross formulas and one vertical cross formula used are:

[0034]

[0035] Among them, r 1 , r 2 is a random number in (0,1), C 1 and C 2 is a random number in (-1,1), and p is a random number in (0,1).

[0036] The beneficial effects of the present invention are as follows: the optimal feature subset obtained by screening improves the accuracy of the diagnosis result while reducing the model operation pressure; a new inertia weight calculation formula is proposed, which is beneficial to the global and local search requirements of particles in the entire iteration cycle; the spiral search strategy is adopted to enhance the traditional PSO search performance; the vertical and horizontal cross mutation strategy is used to update the particles so that the algorithm jumps out of the local optimum; compared with the traditional diagnosis model, the addition of the Adaboost algorithm can enable the IPSO-BP to obtain more stable diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0038] Figure 1 It is a flow chart of the feature selection and improved IPSO-BP-Adaboost diagnostic method of the present invention;

[0039] Figure 2 is a characteristic importance evaluation and diagnostic accuracy curve diagram of the gas combination of the present invention;

[0040] Figure 3 is a structural diagram of a double hidden layer back propagation network used in the present invention;

[0041] Figure 4 It is a structural diagram of the AdaBoost algorithm model used in the present invention;

[0042] Figure 5It is a comparison chart of the mapping effects of the SPM mapping function used in the particle swarm initialization stage of the present invention and the traditional mapping function, Logistic function, and Chebyshev function;

[0043] Figure 6 It is a comparison diagram of the inertia weight change curve of the present invention and the traditional change curve;

[0044] Figure 7 It is a comparison chart of iterative curves when the present invention adds four strategies to PSO one by one;

[0045] Figure 8 It is an iterative curve of BPNN parameter optimization using traditional particle swarm algorithm, sparrow search algorithm (SSA), dung beetle optimization algorithm (DBO), and hippopotamus optimization algorithm (HO) for transformer fault data set in the implementation case of the present invention;

[0046] Fig. 9 It is a comprehensive comparison between the average values ​​of statistical data obtained after inputting different models into the traditional three-ratio method, the Dornenburg method and the gas characteristic combination method adopted by the present invention in the implementation case of the present invention. DETAILED DESCRIPTION

[0047] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0048] like Figure 1 As shown in the figure, the specific implementation steps of the transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost are as follows:

[0049] Step S1, select H 2 , CH 4 , C 2 H 6 , C 2 H 4 and C 2 H 2 As the original characteristic quantity, combined with gas concentration, three-ratio method and uncoded ratio, a total of 26 groups of gas characteristics were obtained;

[0050] Step S2, using a random forest algorithm to select the gas combination that is most conducive to improving the diagnostic accuracy;

[0051] Step S3, using four strategies, namely, SPM chaos mapping based on reverse learning, adaptive inertia weight, spiral flight search, and vertical and horizontal cross mutation, to improve the particle swarm algorithm;

[0052] Step S4, using the improved particle swarm algorithm to optimize the weights and biases in the BPNN network;

[0053] Step S5, using Adaboost to improve the robustness of the IPSO-BP network;

[0054] Specifically, step S1 includes the following steps:

[0055] Step S11, select H 2 , CH 4 , C 2 H 6 , C 2 H 4 and C 2 H 2 As the original feature quantity;

[0056] In step S12, a total of 26 groups of gas characteristics are obtained by combining gas concentration, three-ratio method and uncoded ratio.

[0057] Specifically, the 26 groups of gas feature combinations are as follows, where TH is the sum of all hydrocarbon gas contents; TG is the sum of the contents of 5 original feature gases.

[0058]

[0059] Step S2 is to filter the optimal feature subset and combine Figure 2 As shown in the figure, the minimum number of leaves of the random forest is set to 8, and the number of decision trees n is set to 400. Sort all features in descending order. The higher the ranking, the smaller the contribution of the feature in fault diagnosis. The optimal gas combinations finally screened are: TH, C 2 H 4 / TH,(H 2 +C 2 H 4 ) / TG,C 2 H 2 / C 2 H 4 ,TG,(CH 4 +C 2 H 4 ) / TH,C 2 H 2 / TH,H 2 , C 2 H 6 / TH

[0060] The BPNN network structure used, combined with Figure 3As shown in the figure, the input nodes are set to 9, the number of neurons in the two hidden layers is 20, and the number of output nodes is set to 6. The number of training times is set to 1000, the learning rate is 0.05, and the training target error is 0.000001.

[0061] The Adaboost operation process used, combined with Figure 4 As shown, the number of weak regressors is set to 20, and the exponential function is selected to adjust the weight according to the error.

[0062] Figure 5 It is the particle distribution diagram and its frequency diagram of SPM chaotic function and other chaotic functions, where the SPM function is constructed based on the following formula:

[0063]

[0064] Among them, x(i+1) represents the (i+1)th particle, and r is a random number between 0 and 1. η∈(0,1), and when μ∈(0,1), the function is in a chaotic state, and according to the upper and lower bounds U of the optimized weights and biases b and L b , determine the generated initial solution.

[0065] Combination Figure 6 As shown, the calculation of inertia weight adopts a new nonlinear function, w = (1-t / t max ) (1 / 3) Among them, t max Represents the maximum number of iterations. Compared with the traditional linear curve, the new nonlinear function better balances the local and global search capabilities of particles in the entire iteration cycle.

[0066] More specifically, the ablation test results before and after PSO improvement are as follows: Figure 7 As shown, PSO means no improvement; IPSO1 means the introduction of nonlinear weights and cross-cutting strategies; IPSO2 means the introduction of SPM mapping, nonlinear weights and cross-cutting strategies; IPSO3 means the introduction of chaotic mapping based on reverse learning, nonlinear weights and cross-cutting strategies; IPSO4 means the introduction of chaotic mapping based on reverse learning, nonlinear weights, cross-cutting and spiral search strategies. The above algorithms are used to optimize BPNN at the same time, the objective function is set to the minimum classification accuracy, and the real transformer fault data is used, the maximum number of iterations t max is set to 30, the population number N is set to 50, and the learning factors are all set to 1 and 2. Obviously, IPSO4, after being improved by the four strategies, has better optimization ability and can obtain the minimum fitness, that is, it can minimize the error of the diagnostic results output by BPNN and achieve the highest accuracy.

[0067] Figure 8In the implementation case, based on real fault data, the optimization based on different optimization algorithms was compared, and the traditional particle swarm algorithm, sparrow search algorithm (SSA), dung beetle optimization algorithm (DBO), and hippopotamus optimization algorithm (HO) were used to optimize the key parameters of BPNN. The basic parameters of the model in the experiment are: the number of iterations of each algorithm t max The number of BPNN input nodes is set to 9, the number of neurons in the two hidden layers is set to 20, and the number of output nodes is set to 6. Obviously, the improved IPSO algorithm can obtain the minimum fitness value while converging the fastest. Some other algorithms lack local search capabilities in the later stages of iterations, or require longer iterations to complete the optimization. This shows that the improved IPSO algorithm has excellent global and local search capabilities.

[0068] Fig. 9 In the implementation case, the traditional three-ratio method, Dornenburg method and the optimal gas feature combination after screening were input into different diagnostic models to compare their diagnostic accuracy. Each method was tested independently 30 times and the average value was taken to eliminate the randomness of the single experimental results as much as possible. Obviously, the average diagnostic accuracy on the optimal feature subset used in this paper is between 79% and 91.7%, among which the diagnostic model proposed in this paper has the highest average accuracy of 90.9%.

[0069] Through the above description, within the knowledge of ordinary technicians in this field, relevant personnel can make various changes and modifications without departing from the technical concept of this invention. The technical scope of this invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost, characterized in that: The following steps are involved: Step S1, selecting five original characteristic gases as original characteristic quantities, and combining gas concentration, three ratio method and uncoded ratio to obtain a total of 26 groups of gas characteristics; Step S2, using a random forest algorithm to select the gas combination that is most conducive to improving the diagnostic accuracy; Step S3, using four optimization strategies to improve the particle swarm algorithm; Step S4, using the improved particle swarm algorithm to optimize the weights and biases in the BPNN network; Step S5, using Adaboost to improve the robustness of the IPSO-BP network; The step S3 comprises the following steps: Step S31, setting and initializing the parameters in the genetic particle swarm, the parameters include: maximum number of iterations t max , population size N, dimension D, speed v ij and position x ij The upper and lower limits, the initial value of the inertia weight w, the learning factors c1, c2, and the selection probability p of the search method m and search step length k; Step S32: Use the SPM function to map and generate N solutions, and use reverse learning to generate N reverse solutions, calculate the fitness of a total of 2N solutions, and select the top N solutions in terms of fitness; Step S33: Calculate the inertia weight of the current iteration cycle and the fitness value f(x ij ), obtain the individual optimal pbest and global optimal gbest of the particle; Step S34: According to the selection probability p m Choose to perform a standard PSO search or a spiral search; Step S35: Generate the speed and position of the particle according to the search method; Step S36: performing a lateral crossover operation on the particle with an odd index and the particle after it, and performing a lateral crossover operation on the particle with an even index and the particle before it; Step S37: According to the dimensions of each particle, two different dimensions are randomly selected to perform a vertical cross operation; Step S38: Calculate the fitness value according to the latest position of each particle and update pbest and gbest; Step S39: Determine t, if t=t max , then stop searching and output the obtained gbest and its fitness value, that is, the minimum value of the objective function; otherwise, return to step S33.

2. The transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21, the minimum number of leaves is set to 8, and the number of decision trees is set to 400; Step S22, calculating the importance of each gas combination using the error rate of out-of-bag data of random forest.

3. The transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost according to claim 1 is characterized in that: The chaotic function used to produce the initial population is constructed as: Among them, x(i+1) represents the (i+1)th particle, and r is a random number between 0 and 1. η∈(0,1), and when μ∈(0,1), the function is in a chaotic state, and according to the upper and lower bounds U of the optimized weights and biases b and L b , determine the generated initial solution.

4. The transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost according to claim 1 is characterized in that: In step S33, the calculation of inertia weight adopts a new nonlinear function, w = (1-t / t max ) (1 / 3) in, t max Represents the maximum number of iterations.

5. The transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost according to claim 1 is characterized in that: In step S35, the speed update formula when spiral search is used is: x(t+1)=D'·e cl ·cos(2πl)+x * (t+1), where x * (t+1) is the position of the particle in the last iteration, D' is the distance between the current particle and the global optimal particle, l is a random number between (-1,1), and C is a parameter of a spiral search shape, which is calculated by the formula:

6. The transformer fault diagnosis method based on feature selection and improved IPSO-BP-Adaboost according to claim 1 is characterized in that: Steps 36 and 37 have horizontal and vertical crossover formulas respectively: Among them, r1, r2 is a random number in (0,1), C1 and C2 is a random number in (-1,1), and p is a random number in (0,1).

Images