Power transformer fault diagnosis method based on particle swarm optimization

Abstract

The invention discloses a power transformer fault diagnosis method based on a particle swarm algorithm. The method comprises the following steps: 1, designing a BP neural network model for transformer fault diagnosis; 2, selecting the number of particle network layers of the BP neural network; 3, determining the number of input layer nodes of the BP neural network; 4, determining the number of output layer nodes of the BP neural network; 5, selecting and processing samples of the BP neural network; 6, making a BP neural network strategy improved by a particle swarm optimization algorithm; the method has the advantages that the accuracy of transformer fault diagnosis is improved, the network error precision is reduced, simulation and test results show that the convergence speed and the diagnosis accuracy of a system are obviously improved, and the method is suitable for large-scale popularization and application. And the system is trained by collecting a large number of representative samples, so that the reliability and the accuracy of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis methods, and more specifically to the technical field of power transformer fault diagnosis methods. Background Art

[0002] With the development of national economy and industry, the demand for electric energy in various industries is increasing day by day, and the structural complexity and capacity of the power grid are also increasing to meet the demand for electric energy. In the development of power system, power transformer plays the role of energy transmission and conversion. It is the key equipment for power conversion in power system, which directly affects the safe operation of power system and the quality of electric energy provided by the system. Once the power transformer fails, it will not only cause damage to the equipment, but also have a great impact on the power system, industrial and agricultural production and national defense. Therefore, it is of great significance to the safe and stable operation of the power system to be able to detect the operation failure of the transformer in time and block the expansion of the failure.

[0003] At present, the most commonly used scheme for transformer fault diagnosis is to use the characteristic gas dissolved in the transformer insulating oil to complete the fault judgment, and has achieved certain results. Therefore, it can be understood that the fault diagnosis method of dissolved gas in insulating oil has a certain degree of applicability. However, this type of diagnostic method is based on rich experience and has some shortcomings. The accuracy and integrity of the detection need to be improved. Therefore, it is of great significance to propose a new diagnostic scheme to ensure the normal operation of power transformers. Artificial neural network technology shows strong performance in self-organization and self-learning, and has the mapping characteristics of completing highly nonlinear input and output relationships. Therefore, it is feasible to use it to diagnose transformer faults.

[0004] With the rise of artificial intelligence, BP neural network has been widely used and has important functions in various fields, especially in the field of fault diagnosis. However, there are also some problems. According to the analysis, the standard BP neural network algorithm uses the gradient descent method to correct the weights during the training process. The gradient descent blindly pursues the monotonic decrease of the error, which makes the network easy to fall into the local minimum and difficult to obtain the global optimal value; the convergence and learning of the BP algorithm are somewhat contradictory. In order to achieve the purpose of rapid convergence, the value of the learning rate must be limited, which limits the learning ability of the BP algorithm. In order to ensure the learning ability of the algorithm, the convergence speed of the BP neural network will be slow; the learning and memory ability of the BP neural network is greatly affected by the data samples of the network, and has a great relationship with the weights in the neural network, the number of layers and units of the network, the quality of the samples, etc. The stability of the network is greatly affected and it is difficult to ensure stability.

[0005] Commonly used optimization algorithms include particle swarm optimization, ant colony optimization, neural network algorithm, and genetic algorithm and other intelligent control algorithms. Among these algorithms, particle swarm optimization (PSO) is the most commonly used search algorithm for population optimization. It has good performance when optimizing control parameters, and the algorithm structure is simple and has high control reliability. Therefore, this method proposes to use particle swarm optimization to optimize the BP neural network control parameters for transformer fault diagnosis. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a method for diagnosing power transformer faults based on particle swarm algorithm. The accuracy of transformer fault diagnosis is improved by optimizing BP method with particle swarm algorithm, and the network error precision is reduced.

[0007] The technical solutions adopted by the present invention to solve the above technical problems are as follows:

[0008] The power transformer fault diagnosis method based on particle swarm algorithm has the following steps:

[0009] The first step is to design the BP neural network model for transformer fault diagnosis;

[0010] The second step is to select the number of particle network layers of the BP neural network;

[0011] The third step is to determine the number of nodes in the input layer of the BP neural network;

[0012] The fourth step is to determine the number of nodes in the output layer of the BP neural network;

[0013] Step 5: Sample selection and processing of BP neural network;

[0014] Step 6: Particle swarm optimization algorithm improves BP neural network strategy formulation;

[0015] The seventh step is to determine the optimal weights and threshold parameters of the BP neural network.

[0016] Preferably, in the first step, the BP neural network model design for transformer fault diagnosis is as follows:

[0017] The BP neural network is used to describe the existing mapping relationship, find the nature of transformer faults and the relationship between them and the dissolved gas in the oil, and use it as the basis for fault diagnosis. After collecting transformer operation data, fault characteristics and dissolved gas data samples, the BP neural network fault diagnosis model for transformer fault diagnosis is established by detecting the changes in the characteristic gas dissolved in the transformer oil to judge the transformer fault or predict the fault.

[0018] Preferably, in the second step, the number of particle network layers of the BP neural network is selected as follows:

[0019] According to the characteristics of BP neural network, the particle characteristics of transformer operation status are used as samples as input, and the transformer fault characteristics are used as output. The relationship between input and output is unknown, and is simply processed into three levels and parameter weights in BP neural network to determine the relationship between the input layer and output layer of BP neural network.

[0020] Preferably, in the third step, the number of nodes in the input layer of the BP neural network is determined as follows:

[0021] The sample data is input through the input layer, and the number of its nodes is equal to the dimension of the space vector. When the transformer is in fault-free operation, the gases dissolved in the oil are mainly oxygen and nitrogen. With the long-term operation of the transformer and the occurrence of faults, the gas components in the oil will gradually increase to hydrogen, low molecular hydrocarbon gases, carbon monoxide and carbon dioxide. The content of dissolved carbon monoxide and carbon dioxide is used as the basis for judging whether the transformer has solid insulation faults. The ratios of hydrogen / hydrogen hydrocarbons, methane / total hydrocarbons, ethane / total hydrocarbons, ethylene / total hydrocarbons, and acetylene / total hydrocarbons are used as input vectors, so the number of input layer nodes is 5;

[0022] Preferably, in the fourth step, the number of nodes in the output layer of the BP neural network is determined as follows:

[0023] The fault diagnosis of the transformer is based on the fault mode corresponding to the sample input by the network. According to the corresponding relationship between the fault and the phenomenon, the fault type is output, which is divided into: medium and low temperature overheating fault type, low energy discharge fault type and high energy discharge fault type. There are 5 types of fault types for the data processing transformer, and the number of output layer nodes is 5;

[0024] Preferably, in the fifth step, the sample selection and processing of the BP neural network are as follows:

[0025] The training sample data is linearly normalized. After processing, the training sample data is distributed in the interval [0, 1].

[0026] The calculation formula applied to the input and processed data is:

[0027]

[0028] In formula (1), X jmax and X jm ix represents the training sample data X j The maximum and minimum values ​​in X j Get the processed data X' j, the dissolved gases in the transformer oil are classified, and the content of each gas is calculated according to their respective data characteristics, including hydrogen / hydrogen hydrocarbons, methane / total hydrocarbons, ethane / total hydrocarbons, ethane / total hydrocarbons, ethylene / total hydrocarbons, and acetylene / total hydrocarbons, and the content is used to represent the composition of the data sample.

[0029] For the output quantity, the number of output layer nodes is 5, and the corresponding relationship between the expected values ​​of the transformer fault type network output layer output is shown in Table 1 below

[0030] Table 1 Fault types and their corresponding output values

[0031]

[0032] Preferably, in the sixth step, the particle swarm optimization algorithm improves the BP neural network strategy formulation, as follows:

[0033] For a 3-layer BP network with a topological structure of nkm, the network has n*k+k*m weights and k+m thresholds, so n=5, m=5. The weights and thresholds of the particle swarm BP network are initially optimized, that is, in the entire search space of the particle swarm, the optimal solution that can make the BP network reach a certain convergence accuracy is found from the random solution, so as to form the initial weights and thresholds of the BP network, and then the BP network searches in the negative gradient direction to achieve the training goal. The new improved particle swarm algorithm model is:

[0034]

[0035] In equations (2) and (3), x represents the particle position, v represents the particle velocity, p represents the extreme value of the individual particle, g represents the global optimal solution, c represents the learning acceleration factor, c1 adjusts the step size of the particle flying to the optimal individual position, and c2 adjusts the step size of the particle flying to the global optimal position; r is a random number in the interval [0,1]; k represents the number of iterations of the current particle, and the inertia weight ω is adaptively adjusted using the following formula to balance the global and local search capabilities of the particle swarm algorithm, namely:

[0036]

[0037] In formula (4): k and k max They represent the current number of iterations and the maximum number of iterations of the particle swarm algorithm, ω max represents the maximum inertia weight, ω min Represents the minimum value of the inertia weight.

[0038] The calculation formula of the mean square error index J is:

[0039]

[0040] In the above formula (5), N is the total number of training samples, C is the number of output neurons number, which is the ideal input of the neuron i,j The output value is the actual output value of the neuron. The J shown is the mean square error index, which can be used as the fitness function of the neural network training. The fitness value represents the error threshold of the neural network. The smaller the error, the better the performance of the particle in the search. When the fitness function reaches the minimum, the searched value is the optimal weight and threshold of the neural network.

[0041] Preferably, in the seventh step, the optimal weights and threshold parameters of the BP neural network are determined as follows:

[0042] X=(x1,x2,........,xm) is the input vector of the input layer of the BP neural network, Y=(y1,y2,........,yn) is the output vector of the network, and the input and output relationship of the established BP neural network is Y=f(X) is unknown, training sample,

[0043] The number of input layer nodes of the selected test network is 5; the number of output layer nodes is 5; the maximum number of training steps is 10000; the convergence accuracy is 10 -3 ; Set the number of neurons to 5, 6, 7... and conduct 5 simulation results experiments to test the influence of the number of neurons in the middle layer on the convergence accuracy, time and number of training steps. The test results are shown in Table 2.

[0044] Table 2 Test results of different numbers of neurons in the middle layer

[0045]

[0046] The results show that when the number of neurons is 6, 7..., although the convergence accuracy cannot be achieved, the overall convergence performance of the network is good and the time is within 3s; when the number of neurons is 5, the convergence performance is poor and it is difficult to reach the global minimum.

[0047] The beneficial effects that can be achieved by the present invention using the above technical solution are:

[0048] This method improves the accuracy of transformer fault diagnosis and reduces the network error precision. The simulation and test results show that the convergence speed and diagnostic accuracy of the system have been significantly improved, and the reliability and accuracy of the system have been improved by collecting a large number of representative samples to train the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a structural diagram of the BP neural network diagnosis model in the present invention;

[0050] Figure 2It is the topological structure diagram of the BP neural network in the present invention;

[0051] Figure 3 It is the flow chart of particle swarm optimization BP algorithm in the present invention;

[0052] Figure 4 This is a transformer fault diagnosis structure diagram of the particle swarm optimization BP neural network in the present invention;

[0053] Figure 5 This is a transformer fault diagnosis training error curve diagram in the present invention. DETAILED DESCRIPTION

[0054] The power transformer fault diagnosis method based on particle swarm algorithm has the following steps:

[0055] The first step is to design the BP neural network model for transformer fault diagnosis.

[0056] When diagnosing transformer faults, the occurrence of internal transformer faults has a certain relationship with the dissolved gas in the transformer oil. The BP neural network is used to describe the existing mapping relationship, find the nature of the transformer fault and the relationship between the dissolved gas in the oil, and use it as the basis for fault diagnosis. After collecting transformer operation data, fault characteristics and dissolved gas data samples, the changes in the characteristic gas dissolved in the transformer oil are detected to judge the transformer fault or predict the fault. A BP neural network fault diagnosis model for transformer fault diagnosis is established, such as Figure 1 shown.

[0057] The second step is to select the number of particle network layers of the BP neural network

[0058] To use BP neural network for transformer fault diagnosis, the algorithm and the established diagnostic model need to be trained. According to the characteristics of BP neural network, the transformer operation state particle characteristics are used as samples as input, and the transformer fault characteristics are used as output. The relationship between input and output is unknown. It can be simply processed into three levels and parameter weights in BP neural network to determine the relationship between the input layer and output layer of BP neural network, such as Figure 2 shown.

[0059] The third step is to determine the number of nodes in the input layer of the BP neural network.

[0060] Sample data is generally input through the input layer, and the number of its nodes is equal to the dimension of the space vector. When the transformer is operating without fault, the gases dissolved in the oil are mainly oxygen and nitrogen. However, with the long-term operation of the transformer and the occurrence of faults, the gas components in the oil will gradually increase to hydrogen, low molecular hydrocarbon gases, carbon monoxide and carbon dioxide. The content of dissolved carbon monoxide and carbon dioxide can be used as a solid insulation fault in the transformer, but due to the large amplitude of carbon monoxide and carbon dioxide gases. If carbon monoxide and carbon dioxide gases are added to the input layer, the training error will increase and the breaking capacity will be relatively low. In order to balance the analysis of the oil and gas content in the transformer and the network training error, the ratios of hydrogen / hydrogen hydrocarbons, methane / total hydrocarbons, ethane / total hydrocarbons, ethylene / total hydrocarbons, and acetylene / total hydrocarbons are usually used as input vectors, so the number of input layer nodes is 5.

[0061] The fourth step is to determine the number of nodes in the output layer of the BP neural network.

[0062] Transformer fault diagnosis is to judge the fault mode according to the sample input by the network, and output the fault type according to the corresponding relationship between the fault and the phenomenon. From the relationship between the fault and the corresponding characteristic gas in the transformer oil, it can be seen that in the case of low-temperature overheating and medium-temperature overheating, there is no obvious difference between the gases dissolved in the oil of the corresponding transformers. Therefore, they are analyzed as a unified fault during fault diagnosis, which is called the medium-low temperature overheating fault type; for the same reason, low-energy discharge and overheating and low-energy discharge are combined into low-energy discharge fault type, and arc discharge and overheating in transformer fault types are combined into high-energy discharge fault type. There are 5 types of fault types for processing data transformers and 5 output layer nodes.

[0063] Step 5: Sample selection and processing of BP neural network

[0064] During training, data samples need to be preprocessed to improve the quality of sample training and quickly determine the relationship between the input layer and the output layer. Linear normalization is performed on the training sample data. After processing, the training sample data is distributed in the interval [0, 1]. This processing method can ensure the information in the original sample data while increasing the difference in the sample data. It has a wide range of applications in data processing in algorithm and model training.

[0065] The calculation formula applied to the input and processed data is:

[0066]

[0067] In formula (1), X jmax and X jmin They represent the training sample data X respectively. jThe maximum and minimum values ​​in X j The processed data is obtained. The dissolved gases in the transformer oil are classified, and the contents of each gas are calculated according to their respective data characteristics, including hydrogen / hydrogen hydrocarbons, methane / total hydrocarbons, ethane / total hydrocarbons, ethane / total hydrocarbons, ethylene / total hydrocarbons, and acetylene / total hydrocarbons. The data samples are expressed by using the content. This data processing method improves the utilization rate of sample data while ensuring the validity of the sample data, and can truly reflect the characteristics of the original sample data.

[0068] For the output quantity, the number of output layer nodes is 5. The corresponding relationship between the expected values ​​of the output layer output of the transformer fault type network is shown in Table 1 below.

[0069] Table 1 Fault types and their corresponding output values

[0070]

[0071]

[0072] Step 6: Particle Swarm Optimization Algorithm Improves BP Neural Network Strategy Development

[0073] For a 3-layer BP network with a topological structure of nkm, the network has n*k+k*m weights and k+m thresholds. From the above, we can conclude that n=5 and m=5. The weights and thresholds of the particle swarm BP network are initially optimized, that is, in the entire search space of the particle swarm, the optimal solution that can make the BP network reach a certain convergence accuracy is found from the random solution, so as to form the initial weights and thresholds of the BP network, and then the BP network searches in the negative gradient direction to achieve the training goal. Figure 3 As shown. The new improved particle swarm algorithm model is:

[0074]

[0075] In equations (2) and (3), x represents the particle position, v represents the particle velocity, p represents the extreme value of the individual particle, g represents the global optimal solution, c represents the learning acceleration factor, c1 adjusts the step length of the particle flying to the optimal individual position, and c2 adjusts the step length of the particle flying to the global optimal position; r is a random number in the interval [0,1]; k represents the number of iterations of the current particle. The inertia weight ω is adaptively adjusted using the following formula to balance the global and local search capabilities of the particle swarm algorithm, namely:

[0076]

[0077] In formula (4): k and k max They represent the current number of iterations and the maximum number of iterations of the particle swarm algorithm.max represents the maximum inertia weight, ω min It represents the minimum value in the inertia weight. When searching for each particle, the iterative formula is simple, so the calculation speed is much faster than the gradient descent method, and the inertia weight is adjusted adaptively. It can jump out of the local minimum value and prevent the entire algorithm from converging prematurely. The calculation formula of the mean square error index J is:

[0078]

[0079] In the above formula (5), N is the total number of training samples, C is the number of output neurons, is the ideal output value of the neuron, y i,j is the actual output value of the neuron. The mean square error index shown for each sample can be used as the fitness function of the neural network training. The fitness value represents the error threshold of the neural network. The smaller the error, the better the performance of the particle in the search. In this way, when the fitness function reaches the minimum, the searched out is the optimal weight and threshold of the neural network.

[0080] Step 7: Determination of the optimal weights and threshold parameters of the BP neural network

[0081] X=(x1,x2,........,x m ) is the input vector of the input layer of the BP neural network, Y=(y1,y2,........,y n ) is the output vector of the network, and the input and output relationship of the established BP neural network is Y = f(X) is unknown, and the training sample. Particle swarm optimization BP neural network transformer fault diagnosis structure, such as Figure 4 shown.

[0082] The number of input layer nodes of the selected test network is 5; the number of output layer nodes is 5; the maximum number of training steps is 10000; the convergence accuracy is 10 -3 ; Set the number of neurons to 5, 6, 7... and conduct 5 simulation results experiments to test the influence of the number of neurons in the middle layer on the convergence accuracy, time and number of training steps. The test results are shown in Table 2 below

[0083] Table 2 Test results of different numbers of neurons in the middle layer

[0084]

[0085] The results show that when the number of neurons is 6, 7..., although the convergence accuracy cannot be achieved, the overall convergence performance of the network is good and the time is within 3s; when the number of neurons is 5, the convergence performance is poor and it is difficult to reach the global minimum.

[0086] Simulation experiment of transformer fault diagnosis based on particle swarm optimization BP neural network

[0087] Set the maximum number of iterations to 200, the maximum speed to 1.0, the swarm size N to 40, the neural network structure to 5-14-5, the particle dimension D to 159 (dimensionality calculation = 5*14+14*5+14+5,), the maximum value of the inertia weight is 0.7, the minimum value is 0.3, and the acceleration factor C1=C2=2.0. The transfer function of the input layer and the hidden layer in the particle swarm optimization BP neural network uses the commonly used logsig type function, the transfer function of the hidden layer and the output layer of the network uses the purelin type function, the BP neural network initial network model uses the newff function, and Matlab is used for data processing.

[0088] A random set of 28 test samples was used to conduct transformer fault diagnosis tests, and the effectiveness of the particle swarm optimized BP neural network fault diagnosis was determined based on the test results.

[0089] Table 3 Test results of different fault types

[0090]

[0091]

[0092] From the diagnosis results in Table 3 above, it can be seen that when the particle swarm optimization BP neural network is used for transformer fault diagnosis, 2 groups of wrong fault diagnoses appear in a total of 28 test samples, and the fault diagnosis error rate is 7.14%. Using the same sample data and the standard BP neural network for transformer fault diagnosis, 5 groups of wrong fault diagnoses appear, and the fault diagnosis error rate is 17.86%. By comparison, the fault diagnosis error rate of the particle swarm optimization BP neural network is lower.

[0093] from Figure 5 It can be found that in the particle swarm optimization BP neural network, the particle swarm optimization BP neural network training error converges faster, and can quickly get a better value in the first 25 steps or so. It can reach equilibrium after 140 iterations, and the error is about 0.02. In the standard BP neural network, the error at the final equilibrium is 0.17. By comparison, it is known that the mean square error of the standard BP neural network is 5.67 times that of the particle swarm optimization BP neural network, and the optimization error is larger.

[0094] Combined with the actual operation of the transformer, the diagnostic accuracy of the designed transformer fault diagnosis method is judged. The particle swarm optimization BP neural network algorithm transformer fault diagnosis method collects 100 representative samples from the transformer insulation oil chromatography analysis report of the Electric Power Research Institute and related literature. 75 groups of samples are used to train the particle swarm optimization BP neural network transformer fault diagnosis. After determining the weights and thresholds of the diagnostic network, 25 groups of test samples are used for fault diagnosis. It is found that only two groups of samples have different output results when diagnosed, and the accuracy rate reaches 92%. It further proves that the proposed transformer fault diagnosis method also has a high diagnostic accuracy after changing the data of the training samples.

Claims

1. A power transformer fault diagnosis method based on particle swarm algorithm, characterized in that: The steps are as follows: The first step is to design the BP neural network model for transformer fault diagnosis; The second step is to select the number of particle network layers of the BP neural network; The third step is to determine the number of nodes in the input layer of the BP neural network; The fourth step is to determine the number of nodes in the output layer of the BP neural network; Step 5: Sample selection and processing of BP neural network; Step 6: Particle swarm optimization algorithm improves BP neural network strategy formulation; The seventh step is to determine the optimal weights and threshold parameters of the BP neural network.

2. The power transformer fault diagnosis method based on particle swarm algorithm according to claim 1 is characterized in that: In the first step, the BP neural network model for transformer fault diagnosis is designed as follows: The BP neural network is used to describe the existing mapping relationship, and the nature of the transformer fault and the relationship between the dissolved gas in the oil are found as the basis for fault diagnosis. After collecting transformer operation data, fault characteristics and dissolved gas data samples, the changes in the characteristic gas dissolved in the transformer oil are detected to judge the transformer fault or predict the fault. A BP neural network fault diagnosis model for transformer fault diagnosis is established.

3. The power transformer fault diagnosis method based on particle swarm algorithm according to claim 1 is characterized in that: In the second step, the number of particle network layers of the BP neural network is selected as follows: According to the characteristics of the BP neural network, the particle characteristics of the transformer operating status are used as samples as input, and the transformer fault characteristics are used as output. The relationship between the input and output is unknown and is simply processed into three levels and parameter weights in the BP neural network to determine the relationship between the input layer and the output layer of the BP neural network.

4. The power transformer fault diagnosis method based on particle swarm algorithm according to claim 1 is characterized in that: In the third step, the number of nodes in the input layer of the BP neural network is determined as follows: The sample data is input through the input layer, and the number of its nodes is equal to the dimension of the space vector. When the transformer is operating without fault, the gases dissolved in the oil are mainly oxygen and nitrogen. With the long-term operation of the transformer and the occurrence of faults, the gas components in the oil will gradually increase to hydrogen, low molecular hydrocarbon gases, carbon monoxide and carbon dioxide. The content of dissolved carbon monoxide and carbon dioxide is used as the basis for judging whether the transformer has a solid insulation fault. The ratios of hydrogen / hydrogen hydrocarbons, methane / total hydrocarbons, ethane / total hydrocarbons, ethylene / total hydrocarbons, and acetylene / total hydrocarbons are used as input vectors, so the number of input layer nodes is 5.

5. The power transformer fault diagnosis method based on particle swarm algorithm according to claim 1 is characterized in that: In the fourth step, the number of nodes in the output layer of the BP neural network is determined as follows: The fault diagnosis of the transformer is based on the fault mode corresponding to the sample input by the network. According to the correspondence between the fault and the phenomenon, the fault type is output, which is divided into: medium and low temperature overheating fault type, low energy discharge fault type and high energy discharge fault type. There are 5 types of fault types for processing data transformers, and the number of output layer nodes is 5.

6. The power transformer fault diagnosis method based on particle swarm algorithm according to claim 1 is characterized in that: In the fifth step, the sample selection and processing of the BP neural network are as follows: The training sample data is linearly normalized. After processing, the training sample data is distributed in the interval [0, 1]. The calculation formula applied to the input and processed data is: In formula (1), X jmax and X jmin They represent the training sample data X respectively. j The maximum and minimum values ​​in X j After the processed data is obtained, the dissolved gases in the transformer oil are classified, and the content of each gas is calculated according to the characteristics of each data, including hydrogen / hydrogen hydrocarbons, methane / total hydrocarbons, ethane / total hydrocarbons, ethane / total hydrocarbons, ethylene / total hydrocarbons, and acetylene / total hydrocarbons. For the output quantity, the number of output layer nodes is 5, and the corresponding relationship between the expected values ​​of the transformer fault type network output layer output is shown in Table 1 below Table 1 Fault types and their corresponding output values 7. The power transformer fault diagnosis method based on particle swarm algorithm according to claim 1 is characterized in that: In the sixth step, the particle swarm optimization algorithm improves the BP neural network strategy formulation as follows: For a 3-layer BP network with a topological structure of nkm, the network has n*k+k*m weights and k+m thresholds, so n=5, m=5. The weights and thresholds of the particle swarm BP network are initially optimized, that is, in the entire search space of the particle swarm, the optimal solution that can make the BP network reach a certain convergence accuracy is found from the random solution, so as to form the initial weights and thresholds of the BP network, and then the BP network searches in the negative gradient direction to achieve the training goal. The new improved particle swarm algorithm model is: In equations (2) and (3), x represents the particle position, v represents the particle velocity, p represents the extreme value of the individual particle, g represents the global optimal solution, c represents the learning acceleration factor, c1 adjusts the step size of the particle flying to the optimal individual position, and c2 adjusts the step size of the particle flying to the global optimal position; r is a random number in the interval [0,1]; k represents the number of iterations of the current particle, and the inertia weight ω is adaptively adjusted using the following formula to balance the global and local search capabilities of the particle swarm algorithm, namely: In formula (4): k and k max They represent the current number of iterations and the maximum number of iterations of the particle swarm algorithm, ω max represents the maximum inertia weight, ω min represents the minimum value of the inertia weight. The calculation formula of the mean square error index J is: In the above formula (5), N is the total number of training samples, C is the number of output neurons number, which is the ideal input of the neuron i,j The output value is the actual output value of the neuron. The mean square error index shown can be used as the fitness function of the neural network training. The fitness value represents the error threshold of the neural network. The smaller the error, the better the performance of the particle in the search. When the fitness function reaches the minimum, the searched value is the optimal weight and threshold of the neural network.

8. The power transformer fault diagnosis method based on particle swarm algorithm according to claim 1 is characterized in that: In the seventh step, the optimal weights and threshold parameters of the BP neural network are determined as follows: X=(x1,x2,........,xm) is the input vector of the input layer of the BP neural network, Y=(y1,y2,........,yn) is the output vector of the network, and the input and output relationship of the established BP neural network is Y=f(X) is unknown, training sample, The number of input layer nodes of the selected test network is 5; the number of output layer nodes is 5; the maximum number of training steps is 10000; the convergence accuracy is 10 -3 ; Set the number of neurons to 5, 6, 7... and conduct 5 simulation results experiments to test the influence of the number of neurons in the middle layer on the convergence accuracy, time and number of training steps. The test results are shown in Table 2 Table 2 Test results of different numbers of neurons in the middle layer The results show that when the number of neurons is 6, 7..., although the convergence accuracy cannot be achieved, the overall convergence performance of the network is good and the time is within 3s; when the number of neurons is 5, the convergence performance is poor and it is difficult to reach the global minimum.

Images