BP neural network of spring layered series cement resistor and ant colony detection method

Abstract

The invention relates to the technical field of resistor detection, in particular to a BP (Back Propagation) neural network and ant colony detection method for a spring layered series cement resistor, which comprises the following steps of: acquiring a plurality of groups of current-voltage-power factor data sets of a to-be-detected spring layered series cement resistor; establishing an initial evaluation function model; based on the multiple groups of current-voltage-power factor data sets, parameters in the initial evaluation function model are iteratively corrected by using a BP neural network and an ant colony optimization algorithm, and an optimal evaluation function model is obtained; all the current-voltage-power factor data sets are classified through the optimal evaluation function model, the classified current-voltage-power factor data sets are input into a BP neural network, an evaluation result is obtained, and performance characteristics of the resistor are comprehensively reflected by collecting multi-dimensional current-voltage-power factor data; and through a nonlinear evaluation function model, a complex relationship among multiple parameters is accurately described.

Description

Technical Field

[0001] The present invention relates to the technical field of resistor detection, and particularly to a BP neural network and ant colony detection method for spring-layered series cement resistors. Background Art

[0002] As Figure 1 shown, the spring-layered series cement resistor is a resistor element with a special structure, mainly applied to industrial scenarios that require high power, high temperature resistance, and high stability, such as power plants, substations, power transmission systems, etc. Its characteristic is that it uses a spring-shaped resistance wire inside, and the resistance wire is separated into multiple segments by ceramic rings to form a series structure, and the outside is wrapped with a cement protective layer. This design can improve the heat dissipation performance, high temperature resistance characteristics, and electrical isolation performance of the resistor.

[0003] However, due to the complex internal structure of the spring-layered series cement resistor, traditional detection methods are often difficult to comprehensively evaluate its quality and performance. The existing technologies mainly have the following deficiencies:

[0004] 1. Traditional resistance value measurement methods are difficult to reflect the integrity of the internal structure of the spring-layered series cement resistor;

[0005] 2. Single-parameter detection cannot comprehensively evaluate the dynamic response characteristics of the resistor under different working conditions;

[0006] 3. Linear evaluation models are difficult to accurately describe the complex non-linear relationships between multiple parameters;

[0007] 4. Manual detection relies on empirical judgment and has problems of subjectivity and inconsistency;

[0008] 5. Existing detection methods lack adaptability and are difficult to adapt to resistors of different batches and different specifications.

[0009] Therefore, there is an urgent need for a detection method that can comprehensively, accurately, and adaptively evaluate the quality of spring-layered series cement resistors. Summary of the Invention

[0010] The purpose of the present invention is to provide a BP neural network and ant colony detection method for spring-layered series cement resistors, aiming to solve the problem that traditional detection methods are difficult to comprehensively evaluate the quality and performance of spring-layered series cement resistors.

[0011] The present invention proposes a BP neural network and ant colony detection method for spring-layered series cement resistors, including:

[0012] An acquisition step, including acquiring a multi-group of current-voltage-power factor data sets of the spring-layered series cement resistor to be detected;

[0013] Processing steps, including: establishing an initial evaluation function model; based on the multi-group current-voltage-power factor data sets, using a BP neural network and an ant colony optimization algorithm to iteratively correct the parameters in the initial evaluation function model to obtain an optimal evaluation function model;

[0014] Output steps, including: classifying all current-voltage-power factor data sets using the optimal evaluation function model, and inputting the classified current-voltage-power factor data sets into a BP neural network to obtain an evaluation result.

[0015] Preferably, the obtaining step specifically includes: obtaining multi-group current-voltage-power factor data sets of the spring-layered series cement resistor to be detected according to the multi-group current-voltage-power factor data of the standard sample resistor and combining with the measured current-voltage-power factor data of the actual sample resistor.

[0016] Preferably, the current-voltage-power factor data sets are as follows:

[0017] μ i =(μ i1 ,μ i2 ,...,μ ij ,...,μ in ),

[0018] wherein, μ ij represents the current-voltage-power factor data of the jth phase in the ith test, i = 1, 2, 3,..., N, j = A, B, C, and N represents the total number of test schemes.

[0019] Preferably, the initial evaluation function model adopts the form of the following non-linear model function:

[0020] wherein, i = 1, 2,..., N, N is the number of test schemes, a0, a1,..., a m are the initial linear evaluation weight coefficients, b j , b j c j are the initial non-linear evaluation weight coefficients, m represents the number of initial models, d represents the initial exponential coefficient, and μ i represents the current-voltage-power factor data of the ith test.

[0021] Preferably, the processing steps specifically include:

[0022] (1) Setting the initial parameters of the BP neural network;

[0023] (2) Inputting the current-voltage-power factor data sets into the BP neural network and outputting the results;

[0024] (3) Optimize the BP neural network using the ant colony algorithm: Use the ant colony algorithm to optimize the weights and thresholds of the BP network;

[0025] (4) Determine whether the current iteration count has reached the limit: If so, stop the optimization and obtain the optimal evaluation function model; otherwise, reset the network parameters according to the optimization results and return to step (2).

[0026] Preferably, in step (3), randomly generate the initial positions of the ant colony in the search space, map the values corresponding to the optimal results of the ant colony algorithm iteration to the structure of the BP neural network, and use the results as the initial parameters of the BP neural network.

[0027] Preferably, in step (3), the search space of the ant colony algorithm is as follows:

[0028] x = (a0, a1, a2,..., a n , b1c1d,..., b m c m d),

[0029] where a i represents the initial linear evaluation weight coefficient in the initial evaluation function model, and b j , b j c j represent the initial non-linear evaluation weight coefficients, and x i represents the parameter to be adjusted;

[0030] In step (3), the output of the BP neural network is as follows:

[0031]

[0032] where a i , b j , b j c j represent the parameters to be optimized.

[0033] Preferably, the following method is used to obtain the optimal result of the ant colony algorithm iteration:

[0034] (3-1) According to the measured current-voltage-power factor data set of the resistor to be tested, obtain the predicted output result through the BP neural network;

[0035] (3-2) Take the average of all the prediction results and output it as the current result;

[0036] (3-3) When the evolution termination condition is satisfied, stop the evolution and obtain the current result as the optimal result; when the evolution termination condition is not satisfied, randomly move the ant to the next layer and repeat steps (3-1) and (3-2).

[0037] Preferably, the BP neural network adopts a multi-layer structure, including:

[0038] Input layer: Receive the current-voltage-power factor data;

[0039] Multiple hidden layers: Adopt a variable structure design and adaptively adjust the number of neurons according to the data complexity;

[0040] Output layer: Generate an evaluation result;

[0041] Among them, the BP neural network introduces a dynamic learning rate adjustment mechanism, adopts an improved activation function, and designs a special error backpropagation path.

[0042] Preferably, the ant colony optimization algorithm introduces the following mechanisms:

[0043] Pheromone double-channel mechanism: Optimize linear and non-linear parameters simultaneously;

[0044] Parameter sensitivity perception mechanism: Optimize the parameters with greater influence first;

[0045] Multi-scale co-evolution strategy: Avoid premature convergence;

[0046] Among them, it also includes a performance fluctuation monitor for real-time evaluation of the optimization effect and adjustment of the ant colony search range and the BP network learning rate according to the monitoring results.

[0047] By combining the BP neural network and the ant colony optimization algorithm, the present invention establishes an adaptive and high-precision quality evaluation system, which has the following beneficial effects:

[0048] 1. By collecting multi-dimensional current-voltage-power factor data, comprehensively reflect the performance characteristics of the resistor;

[0049] 2. Through the non-linear evaluation function model, accurately describe the complex relationship between multiple parameters;

[0050] 3. Through the collaborative optimization of the BP neural network and the ant colony algorithm, improve the optimization efficiency and accuracy of the model parameters;

[0051] 4. Have an adaptive ability and can adapt to the detection requirements of resistors of different batches and different specifications;

[0052] 5. The detection accuracy is significantly improved, the misjudgment rate is greatly reduced, and the determination of resistors in the boundary state is more accurate especially;

[0053] 6. The possibility of predictive maintenance is realized, and the future performance change trend of the resistor can be predicted based on historical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 FIG. is a schematic structural diagram of the spring-layered series cement resistor of the present invention;

[0055] Figure 2 FIG. is a schematic flow diagram of the BP neural network and ant colony optimization detection method of the present invention;

[0056] Figure 3 FIG. is a schematic structural diagram of the BP neural network of the present invention;

[0057] Figure 4 FIG. is a schematic flow diagram of the ant colony algorithm optimization of the present invention;

[0058] Figure 5 FIG. is a schematic diagram of the pheromone double-channel mechanism of the present invention;

[0059] Figure 6 FIG. is a schematic diagram of the parameter sensitivity perception mechanism of the present invention;

[0060] Figure 7 FIG. is a schematic diagram of the multi-scale co-evolution strategy of the present invention;

[0061] Figure 8 FIG. is a schematic diagram of the current-voltage-power factor data acquisition of the present invention;

[0062] Figure 9 FIG. is a comparison chart of the test results of the embodiments of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0063] Please refer to the attached Figures 1-9 , and the present invention will be further described in detail below with reference to the drawings and embodiments. However, the embodiments of the present invention are not limited thereto.

[0064] Refer to Figure 1 , Embodiment 1: Basic Detection Method

[0065] Refer to Figure 2 , the BP neural network and ant colony detection method of the spring-layered series cement resistor provided by the present invention includes an acquisition step, a processing step, and an output step.

[0066] The acquisition step includes acquiring multiple groups of current-voltage-power factor data sets of the spring-layered series cement resistor to be detected. Specifically, multiple groups of current-voltage-power factor data of the spring-layered series cement resistor to be detected can be obtained by combining the current-voltage-power factor data of the standard sample resistor with the measured current-voltage-power factor data of the actual sample resistor.

[0067] In this embodiment, the current-voltage-power factor data set is as follows:

[0068] μ i =(μ i1 , μ i2 ,..., μ ij ,..., μ in ),

[0069] where μ ij represents the current-voltage-power factor data of the j-th phase in the i-th test, i = 1, 2, 3,..., N, j = A, B, C, and N represents the total number of test schemes.

[0070] Preferably, a dedicated test device can be used for data acquisition. This device can simultaneously measure the voltage value and power factor of the resistor at different currents. For example, measurements can be made in the range of current from 0.1 A to 10 A with a step size of 0.5 A to obtain 20 groups of test data. At the same time, three-phase measurements of A, B, and C are performed on each group of data to comprehensively evaluate the performance characteristics of the resistor.

[0071] The processing steps include establishing an initial evaluation function model, and based on multiple groups of current-voltage-power factor data sets, using the BP neural network and ant colony optimization algorithm to iteratively correct the parameters in the initial evaluation function model to obtain the optimal evaluation function model.

[0072] In this embodiment, the initial evaluation function model adopts the form of the following non-linear model function:

[0073]

[0074] where i = 1, 2,..., N, N is the number of test schemes, a0, a1,..., a m are the initial linear evaluation weight coefficients, b j , b j c j are the initial non-linear evaluation weight coefficients, m represents the number of initial models, d represents the initial exponential coefficient, and μ i represents the current-voltage-power factor data of the i-th test.

[0075] Preferably, the initial linear evaluation weight coefficients a0, a1,..., a m can determine the initial values based on historical detection data through the least squares method. For example, a0 can be set to 0.5, and a1 to a m can be uniformly distributed in the range of [-1, 1]. The initial non-linear evaluation weight coefficients b j , b j c jIt can be set within the range of [0.1, 0.9], the initial exponential coefficient d can be set to 2. The initial number of models m can be set according to the data complexity, and a value between 3 and 5 is usually selected.

[0076] The processing steps specifically include:

[0077] 1. Setting the initial parameters of the BP neural network;

[0078] 2. Inputting the current-voltage-power factor data set of the BP neural network and outputting the results;

[0079] 3. Optimizing the BP neural network using the ant colony algorithm: using the ant colony algorithm to optimize the weights and thresholds of the BP network;

[0080] 4. Judging whether the current iteration number reaches the limit number: if so, stop the optimization and obtain the best evaluation function model; otherwise, reset the network parameters according to the optimization results and return to step 2.

[0081] In this embodiment, setting the initial parameters of the BP neural network includes setting the number of network layers, the number of neurons in each layer, the learning rate, the momentum factor, etc. Preferably, a three-layer network structure can be adopted. The number of neurons in the input layer is the same as the dimension of the input data, the number of neurons in the hidden layer is 1.5 times that of the input layer, and the number of neurons in the output layer is 1. The learning rate can be set to a value between 0.01 and 0.1, and the momentum factor can be set to 0.9.

[0082] In the step of optimizing the BP neural network using the ant colony algorithm, randomly generate the initial positions of the ant colony in the search space, map the values corresponding to the optimal results of the ant colony algorithm iteration to the structure of the BP neural network, and use the results as the initial parameters of the BP neural network.

[0083] The search space of the ant colony algorithm is as follows:

[0084] x = (a0, a1, a2,..., a n , b1c1d,..., b m c m d),

[0085] where, a i represents the initial linear evaluation weight coefficient in the initial evaluation function model, b j , b j c j represent the initial non-linear evaluation weight coefficients, and x i represents the parameter to be adjusted.

[0086] The output of the BP neural network is as follows:

[0087]

[0088] Among them, a i , b j , b j c j represent the parameters to be optimized.

[0089] The optimal result of the ant colony algorithm iteration is obtained in the following way:

[0090] 1. According to the current - voltage - power factor data set of the resistor to be detected in the sample, the predicted output result is obtained through the BP neural network;

[0091] 2. Take the average value of all the predicted results as the current result for output;

[0092] 3. When the evolution termination condition is met, stop the evolution and obtain the current result as the optimal result; when the evolution termination condition is not met, randomly move the ant to the next layer and repeat steps 1 and 2.

[0093] In this embodiment, the evolution termination condition can be set as a combination of the following situations:

[0094] Reach the maximum number of iterations, such as 1000 times;

[0095] The change in the iteration result is less than the preset threshold for 50 consecutive times, such as 0.001;

[0096] Reach the target accuracy, such as the prediction error is less than 1%.

[0097] The output step includes classifying all the current - voltage - power factor data sets by using the best evaluation function model, inputting the classified current - voltage - power factor data sets into the BP neural network, and obtaining the evaluation result.

[0098] Preferably, the evaluation result can include the quality - grade classification of the resistor (such as excellent, good, qualified, unqualified), and the specific performance index scores (such as stability, reliability, life prediction, etc.).

[0099] Embodiment 2: Optimization of the BP neural network structure

[0100] Based on Embodiment 1, this embodiment further optimizes the BP neural network structure. Referring to Figure 3 , the BP neural network adopts a multi - layer structure, including:

[0101] Input layer: Receives the current - voltage - power factor data;

[0102] Multiple hidden layers: Adopt a variable - structure design and adaptively adjust the number of neurons according to the data complexity;

[0103] Output layer: Generates the evaluation result.

[0104] Among them, the BP neural network introduces a dynamic learning rate adjustment mechanism, adopts an improved activation function, and designs a special error back propagation path.

[0105] Preferably, the dynamic learning rate adjustment mechanism can be expressed as:

[0106]

[0107] Among them, η(t) represents the learning rate of the tth iteration, η0 represents the initial learning rate, which is usually set to 0.1, α is the learning rate fluctuation coefficient, which is usually set to 0.5, T is the learning rate adjustment period, which is usually set to 100, and β is the attenuation coefficient, which is usually set to 0.1. The improved activation function uses the Swish function, which is in the form of:

[0108] f(x)=x·σ(γx),

[0109] Among them, σ represents the sigmoid function, and γ is an adjustable parameter, which is usually set to 1.0. Compared with the traditional sigmoid function and ReLU function, the Swish function has better sensitivity in weak signal areas and can better capture small changes in resistor parameters.

[0110] The special error back-propagation path design includes skip connections and residual connections, which can effectively solve the gradient vanishing problem in deep network training and accelerate the convergence process.

[0111] Example 3: Ant Colony Optimization Algorithm Innovation Mechanism

[0112] Based on Example 1 and Example 2, this example further optimizes the ant colony algorithm. Figure 4 , Figure 5 , Figure 6 and Figure 7 , the ant colony optimization algorithm introduces the following mechanisms:

[0113] 1. Pheromone dual-channel mechanism: simultaneous optimization of linear and nonlinear parameters;

[0114] 2. Parameter sensitivity perception mechanism: prioritize optimization of parameters with greater impact;

[0115] 3. Multi-scale co-evolution strategy: avoiding premature convergence.

[0116] In addition, it also includes a performance fluctuation monitor for real-time evaluation of the optimization effect and adjustment of the ant colony search range and BP network learning rate according to the monitoring results.

[0117] Specifically, the pheromone dual-channel mechanism divides the parameter space into a linear parameter channel and a nonlinear parameter channel. The two channels use different pheromone update rules:

[0118] Pheromone update rule for the linear parameter channel:

[0119] τ ij (t + 1) = (1 - ρ)·τ ij (t) + ρ·Δτ ij ,

[0120] Pheromone update rule for the non - linear parameter channel:

[0121] τ ij (t + 1) = (1 - ρ 2 )·τ ij (t) + ρ 2 ·Δτ ij ,

[0122] where τ ij (t) represents the pheromone concentration on the path (i, j) at time t, ρ represents the pheromone evaporation coefficient, usually set to 0.5, and Δτ ij represents the pheromone increment. The parameter sensitivity perception mechanism optimizes by calculating the influence degree of each parameter on the evaluation result and allocating different numbers of ants. The parameter sensitivity calculation formula is

[0123]

[0124] where S i represents the sensitivity of the parameter x i , E represents the error of the evaluation function, and Δx i represents the change of the parameter x i . The higher the sensitivity of the parameter, the more ants are allocated and the greater the optimization intensity.

[0125] The multi - scale co - evolution strategy divides the ant colony into multiple subgroups, and each subgroup is responsible for searching in the parameter space of different scales. The scale factor is defined as:

[0126] λ k = λ0·γ k ,

[0127] where λ k represents the scale factor of the k - th subgroup, λ0 represents the initial scale factor, usually set to 1.0, γ represents the scale decay coefficient, usually set to 0.8, and k represents the subgroup number. The performance fluctuation monitor judges whether the optimization process falls into a local optimum by calculating the performance standard deviation of continuous n iterations:

[0128]

[0129] where σ represents the performance fluctuation standard deviation, and E i represents the evaluation error of the i - th iteration. It represents the average evaluation error of n iterations. When σ is less than a preset threshold (e.g., 0.001), it indicates that the optimization process may fall into a local optimum. At this time, it is necessary to increase the random search probability of the ants and expand the search range.

[0130] Through the synergistic effect of the above mechanisms, the ant colony optimization algorithm can more efficiently find the optimal parameters of the BP neural network and improve the accuracy and adaptability of the detection method.

[0131] Example 4: Example of the complete detection process

[0132] The following combines a specific example to elaborate in detail on the complete detection process of the present invention.

[0133] First, prepare 10 standard sample resistors (with known quality grades) and 100 sample resistors to be measured. For each resistor, within the range of 0.1 A to 10 A, 20 groups of three-phase current-voltage-power factor data are measured at a step size of 0.5 A.

[0134] Then, establish an initial evaluation function model and set the initial parameters: a0 = 0.5, a1 to a5 are uniformly distributed within [-1, 1], b j 、b j c j is uniformly distributed within [0.1, 0.9], d = 2, m = 5.

[0135] Next, set the parameters of the BP neural network: a three-layer network structure, 60 neurons in the input layer (corresponding to 20 groups × 3-phase data), 90 neurons in the hidden layer, and 1 neuron in the output layer. The initial learning rate is set to 0.05, and the momentum factor is set to 0.9.

[0136] Then, set the parameters of the ant colony algorithm: the number of ants is 50, the maximum number of iterations is 1000, the pheromone evaporation coefficient is 0.5, and the initial pheromone concentration is 0.1. The ants are divided into 5 sub-groups, with 10 ants in each sub-group. The initial scale factor is 1.0, and the scale attenuation coefficient is 0.8.

[0137] Start the detection process:

[0138] 1. Input the current-voltage-power factor data of the standard sample resistors into the BP neural network for initial training;

[0139] 2. Use the ant colony algorithm to optimize the parameters of the BP neural network;

[0140] 3. When the termination condition is reached (1000 iterations or the standard deviation of performance fluctuations is less than 0.001 for 50 consecutive times), obtain the best evaluation function model;

[0141] 4. Classify the current-voltage-power factor data of the resistor to be measured using the optimal evaluation function model;

[0142] 5. Input the classified data into the optimized BP neural network to obtain the quality evaluation results of each resistor.

[0143] The experimental results show that compared with the traditional detection method, the detection method of the present invention has an accuracy improvement of 35%, and the misjudgment rate is reduced to less than 5%. In particular, the determination of resistors in the boundary state is more accurate. At the same time, the detection efficiency is increased by 50%, and it can meet the detection requirements of resistors of different batches and different specifications.

[0144] Through the detailed description of the above embodiments, those skilled in the art should understand that the present invention can have various modifications and variations without departing from the spirit and scope of the present invention.

[0145] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. BP neural network and ant colony detection method for spring layered series cement resistors, characterized in that: include: The acquisition step includes acquiring multiple sets of current-voltage-power factor data sets of the spring layered series cement resistor to be detected; The processing steps include: establishing an initial evaluation function model; based on the multiple sets of current-voltage-power factor data sets, using BP neural network and ant colony optimization algorithm to iteratively correct the parameters in the initial evaluation function model to obtain the best evaluation function model; The output step includes: using the optimal evaluation function model to classify all current-voltage-power factor data sets, inputting the classified current-voltage-power factor data sets into the BP neural network, and obtaining evaluation results.

2. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 1 is characterized in that: The acquisition step specifically includes: acquiring multiple sets of current-voltage-power factor data sets of the spring layered series cement resistor to be tested based on multiple sets of current-voltage-power factor data of the standard sample resistor and combining them with the measured current-voltage-power factor data of the actual sample resistor.

3. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 1 is characterized in that: The current-voltage-power factor data set is as follows: m i =(μ i1 ,m i2 ,...,m ij ,...,m in ), Among them, μ ij Represents the current-voltage-power factor data of the jth phase in the i-th test, i=1,2,3,…,N, j=A,B,C, N represents the total number of test schemes.

4. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 1, characterized in that: The initial evaluation function model adopts the form of the following nonlinear model function: Where i = 1, 2, ..., N, N is the number of test solutions, a0, a1, ..., a m is the initial linear evaluation weight coefficient, b j , b j c j is the initial nonlinear evaluation weight coefficient, m represents the number of initial models, d represents the initial exponential coefficient, μ i Represents the current-voltage-power factor data of the i-th test.

5. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 1, characterized in that: The processing steps specifically include: (1) Initial parameter setting of BP neural network; (2) The BP neural network inputs the current-voltage-power factor data set and outputs the result; (3) Ant colony algorithm to optimize BP neural network: Use ant colony algorithm to optimize the weights and thresholds of BP network; (4) Determine whether the current number of iterations has reached the limit: if so, stop the optimization and obtain the best evaluation function model; otherwise, reset the network parameters according to the optimization results and return to step (2).

6. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 5, characterized in that: In the step (3), the initial ant colony position is randomly generated in the search space, the value corresponding to the optimal result of the ant colony algorithm iteration is mapped to the structure of the BP neural network, and the result is used as the initial parameter of the BP neural network.

7. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 5, characterized in that: In step (3), the search space of the ant colony algorithm is as follows: x=(a0,a1,a2,...,a n ,b1c1d,...,b m c m d), Among them, a i represents the initial linear evaluation weight coefficient in the initial evaluation function model, b j , b j c j represents the initial nonlinear evaluation weight coefficient, x i Indicates the parameter to be adjusted; In step (3), the BP neural network output is as follows: Among them, a i ,b j ,b j c j Indicates the parameters to be optimized.

8. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 5, characterized in that: The optimal result of the ant colony algorithm iteration is obtained in the following way: (3-1) obtaining a prediction output result through a BP neural network according to a set of current-voltage-power factor data of the sample resistor to be tested; (3-2) Take the average of all prediction results and output it as the current result; (3-3) When the evolution termination condition is met, the evolution is stopped and the current result is taken as the optimal result; when the evolution termination condition is not met, the ants are randomly moved to the next layer and steps (3-1) and (3-2) are repeated.

9. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 1, characterized in that: The BP neural network adopts a multi-layer structure, including: Input layer: receiving the current-voltage-power factor data; Multiple hidden layers: adopt variable structure design to adaptively adjust the number of neurons according to data complexity; Output layer: generates evaluation results; The BP neural network introduces a dynamic learning rate adjustment mechanism, adopts an improved activation function, and designs a special error back propagation path.

10. The BP neural network and ant colony detection method of spring layered series cement resistor according to claim 1, characterized in that: The ant colony optimization algorithm introduces the following mechanisms: Pheromone dual-channel mechanism: simultaneous optimization of linear and nonlinear parameters; Parameter sensitivity perception mechanism: prioritize optimization of parameters with greater impact; Multi-scale co-evolution strategy: avoiding premature convergence; It also includes a performance fluctuation monitor, which is used to evaluate the optimization effect in real time and adjust the ant colony search range and BP network learning rate according to the monitoring results.

Images