CN109143093A

CN109143093A - Based on the battery SOC evaluation method for intersecting optimization neural network in length and breadth

Info

Publication number: CN109143093A
Application number: CN201810692602.6A
Authority: CN
Inventors: 吴杰康
Original assignee: Guangdong University of Technology
Current assignee: Guangdong University of Technology
Priority date: 2018-06-29
Filing date: 2018-06-29
Publication date: 2019-01-04

Abstract

The invention relates to a battery SOC estimation method based on a crisscross optimization neural network. The classical neural network algorithm is optimized by the crisscross algorithm, and the advantages of the crisscross algorithm's strong global search ability and fast convergence speed are combined with the strong neural network. The fitting ability is organically combined to avoid the neural network from falling into local optimum and improve its convergence speed. In addition, compared with the existing battery SOC estimation method, the present invention is suitable for a series of commonly used batteries such as lithium batteries and lead-acid batteries. Whether the battery is in a stationary state or a use state, the battery can be estimated in real time for SOC, and The accuracy is high, and the error is smaller than other methods.

Description

Based on the battery SOC evaluation method for intersecting optimization neural network in length and breadth

Technical field

The present invention relates to the technical fields of battery, more particularly to based on the battery SOC for intersecting optimization neural network in length and breadth Evaluation method.

Background technique

With the continuous development of human economic society, the energy increasingly become promote socio-economic development it is indispensable because The development of element, economic society also exacerbates demand of the mankind to the energy.It is most of from change at present in the energy used in the mankind Stone fuel.The use of fossil fuel promotes the improvement of people's living standards, while also bringing serious environmental problem.It is slow Environmental problem brought by combustion of fossil fuel is solved, new-energy automobile is being greatly developed in the whole world.Power battery is as new energy The energy storage device of source automobile is the core of new-energy automobile, is the maximum bottleneck in new-energy automobile technology and cost, is new energy A most crucial ring in the Automotive Industry Chain of source.The power in energy storage device or new-energy automobile either in generation of electricity by new energy Battery, battery all play crucial effect as energy storage equipment.And key technology of the power battery as electric car, to lotus Electricity condition (state of charge, SOC) accurately estimated and monitored, from the point of view of safety and battery availability factor all It is most important.

It accurately estimates battery SOC, the requirement of electric car is on the one hand derived from, from giving full play to cell potential and raising Two angles of safety efficiently manage battery；On the other hand, the height that batteries of electric automobile shows in use It is non-linear, make accurately to estimate that SOC has very big difficulty.Both sides combines, so that batteries of electric automobile SOC estimation method It selects particularly important.

Existing battery SOC evaluation method mainly has: discharge test method, current integration method, open circuit voltage method, measurement internal resistance Method, linear model method, neural network and Kalman filtering method.Wherein neural network is not only only capable of compared to other methods Accurately battery SOC is estimated, and it is not influenced by battery types and battery status.Currently available technology research In much battery SOC is estimated using neural network algorithm.This method is by the discharge current of battery, battery pack The input as neural network such as voltage, environment temperature and discharge capacity, SOC is as its output, to carry out to battery SOC Estimation, neural network to it regardless of that can carry out SOC estimation, and estimation precision in battery standing state or working condition It is higher compared to other methods.But traditional neural network is easily trapped into local optimum in its algorithm operational process, in this way Result in estimation inaccurate, error is larger.There are also in neural network method use optimization algorithm, but estimation precision according to So be not it is very high, there is a certain error.

Summary of the invention

It is an object of the invention to overcome the deficiencies of the prior art and provide a kind of based on intersection optimization neural network in length and breadth Battery SOC evaluation method, by the global search of crossover algorithm in length and breadth is strong and the advantage of fast convergence rate and neural network compared with Strong capability of fitting organically combines, and accelerates neural network convergence rate, and will not fall into local optimum, this method Battery SOC can be carried out in real time and accurately be estimated, not influenced by battery types and battery status, compare existing electricity The accuracy of pond SOC estimation method, the method is higher, and error is smaller.

The technical scheme of the present invention is realized as follows:

Based on the battery SOC evaluation method for intersecting optimization neural network in length and breadth, comprising the following steps:

S1: data sample is obtained, and carries out sample data normalized；

S2: the principal element of analyzing influence battery SOC constructs BP neural network structure；

S3: algorithm parameter initialization；

S4: the connection weight and every threshold value between the output valve and each layer of BP neural network are calculated；

S5: according to the input value and output valve of BP neural network, with the real output value of BP network and desired output Fitness function of the mean square error as CSO calculates the adaptation value of each CSO particle, obtains the individual optimal value of particle and complete Office's optimal value, then makes comparisons the optimal value of CSO individual with global optimum, takes adaptation value the superior as current optimal position It sets；

S6: it is optimized using weight and threshold value of the crossover algorithm in length and breadth to BP neural network；

S7: step S5 and S6 are repeated, until meeting termination condition；

S8: the parameter that CSO algorithm optimization is obtained is as the initial weight of BP neural network and threshold value, and by initial weight It substitutes into BP neural network algorithm and is trained with threshold value；If the output error value of BP neural network meets scheduled error essence Degree then stops iteration, exports result；Otherwise, step S5 is returned to, Optimized Iterative is re-started, until meeting BP neural network calculation Until the minimum allowable error of method.

Further, the analytic process in the step S2 is as follows:

S2-1: the relationship analysis of battery SOC and open-circuit voltage is carried out, the SOC of battery is estimated according to SOC-OCV curve；

S2-2: carrying out the relationship analysis of battery SOC and temperature, and in the case of obtaining different temperatures, the relationship of voltage and SOC are bent Line；

S2-3: carrying out the relationship analysis of battery SOC and discharge current, obtains voltage and capacity relationship when different multiplying electric discharge Curve.

Further, in the step S2, the discharge current I, battery voltage U and environment temperature T for choosing battery make For the input vector of the input layer of BP neural network structure, and three impact factors are independent of one another, and the output vector of network is Battery SOC carries out BP neural network Construction of A Model.

Further, the step S3 algorithm parameter initialization specifically: the topological structure of input BP neural network algorithm, Minimum allowable error amount；Input CSO algorithm population scale, dimensionality of particle number, maximum number of iterations, lateral cross probability and Crossed longitudinally probability.

Further, the step S4 calculates connection weight and items between the output valve and each layer of BP neural network Detailed process is as follows for threshold value:

S4-1: k-th of input sample x (k) and corresponding desired output d are chosen₀(k):

X (k)=(x₁(k),x₂(k),...,x_n(k))；

d₀(k)=(d₁(k),d₂(k),...,d_q(k))；

S4-2: the input hi of hidden layer neuron is calculated_h(k) with output ho_h(k) and the input of output layer neuron yi_o(k) with output yo_o(k):

ho_h(k)=f (hi_h(k)) h=1,2 ..., p；

yo_o(k)=f (yi_o(k)) o=1,2 ..., q；

S4-3: inputting according to output layer desired output and reality output and output layer, calculates function e to each mind of output layer Partial derivative through member:

Error function

S4-4: according to the sensitivity δ o (k) of output layer, the input value of hidden layer connection weight w and output layer are calculated and are missed Partial derivative of the difference function to each neuron of hidden layer:

S4-5: output layer connection weight is corrected using the partial derivative in step S4-3:

S4-6: hidden layer connection weight is corrected using the partial derivative in step S4-4:

Further, the step S5 calculate each CSO particle adaptation value formula it is as follows:

In formula, y_oiIt (i) is the real output value of neural network；d_oiFor the desired output of neural network；E (i) is nerve The mean square error of network real output value and desired output.

Further, the mapping between COS particle and the weight and threshold value of BP neural network is established in the step S6, i.e., The weight of neural network and threshold coding are indicated into the individual in population at real vector, the group of vector is randomly generated into； Specific Optimization Steps are as follows:

S6-1: lateral operation is carried out to population:

Lateral cross is in a kind of arithmetic crossover carried out between the identical dimension of two Different Individual particles in population；Assuming that father Lateral cross is carried out for the d dimension of individual particles X (i) and X (j), then the formula of their generation filial generations is as follows:

MS_hc(i, d)=r₁×X(i,d)+(1-r₁)×X(j,d)+c₁×(X(i,d)-X(j,d))；

MS_hc(j, d)=r₂×X(j,d)+(1-r₂)×X(i,d)+c₂×(X(j,d)-X(i,d))；

In formula: r₁, r₂For the random number between [0,1]；c₁, c₂For the random number between [- 1,1]；X (i, d), X (j, d) The d of individual particles X (i) and X (j) is tieed up respectively in parent population；MS_hc(i, d) and MS_hc(j, d) is respectively X (i, d) and X (j, d) ties up filial generation by the d that lateral cross generates；

The wherein r in first formula₁× X (i, d) is the memory term of particle X (i), is the current optimal value of particle itself； (1-r₁) × X (j, d) is the group cognition item of particle X (i) and X (j), indicates that difference is interparticle and influences each other；This two logical Cross inertia weight factor r₁Preferably it is combined together；c₁For Studying factors, Section 3 c₁× (X (i, d)-X (j, d)) can increase Search space, in edge optimizing；After the completion of lateral cross operation, obtained golden mean of the Confucian school solution MS_hc(i, d), MS_hc(j, d) must distinguish Compare with the fitness of parent particle X (i), X (j), the only better golden mean of the Confucian school solution of fitness can just remain, and become and be dominant Solve DS_hc, participate in next iteration；

S6-2: crossed longitudinally operation is carried out to population:

The one kind carried out between crossed longitudinally two different dimensions for a particle in population counts intersection；It is assumed that particle The d of X (i)₁Peacekeeping d₂Dimension generates golden mean of the Confucian school solution MS to participate in always wanting to intersect, according to following formula_vc(i,d₁):

MS_vc(i,d₁)=rX (i, d₁)+(1-r)·X(i,d₂)

i∈N(1,M),d₁,d₂∈N(1,D)

In formula: i ∈ [0,1]；MS_vc(i,d₁) be individual particles X (i) d₁Peacekeeping d₂Dimension passes through crossed longitudinally generation D₁Tie up offspring；First item is the d of particle X (i)₁The memory term of dimension, Section 2 are the d of particle X (i)₁Peacekeeping d₂Dimension It influences each other, is combined together by inertia weight factor r；Obtained golden mean of the Confucian school solution MS_vc(i,d₁) comprising parent particle X (i) D₁The information of dimension and certain probability contain the d of X (i)₂Information is tieed up, and the d of X (i) will not be destroyed₂Tie up information；The golden mean of the Confucian school Solve MS_vc(i,d₁) fitness compared with parent particle X (i), it preferably remains and solves DS as being dominant_vc, changed next time Generation；

New population is generated by the contention operation of filial generation and parent；If new adaptive value is optimal better than current individual, Replace current individual optimal with the adaptive value: if updated individual optimal value is better than current global optimum, with the individual Optimal value replaces current global optimum, to complete the optimization to network items weight and threshold value.

Compared with prior art, this programme principle and advantage is as follows:

Classical neural network algorithm is optimized by crossover algorithm in length and breadth, by the global search of crossover algorithm in length and breadth The advantages of sexuality is strong and fast convergence rate organically combines with the stronger capability of fitting of neural network, avoids neural network Local optimum is fallen into, and improves its convergence rate.In addition, comparing existing battery SOC evaluation method, this programme is suitable for lithium A series of common batteries such as battery, lead-acid battery, either battery are in standing or use state, can be in real time to electricity Pond carries out SOC estimation, and accuracy is high, smaller compared to other methods error.

Detailed description of the invention

Fig. 1 is that the present invention is based on the flow charts for the battery SOC evaluation method for intersecting optimization neural network in length and breadth；

Fig. 2 is the relation curve schematic diagram of battery SOC and open-circuit voltage U；

Fig. 3 is the T relationship curve synoptic diagram of battery SOC and environment temperature；

Fig. 4 is the I relation curve schematic diagram of battery SOC and discharge current；

Fig. 5 is the BP neural network Construction of A Model figure of battery SOC estimation.

Specific embodiment

The present invention is further explained in the light of specific embodiments:

As shown in Figure 1, based on the battery SOC evaluation method for intersecting optimization neural network in length and breadth, comprising the following steps:

S1: all kinds of parameter monitorings of discharge test are carried out to battery pack by real-time monitoring system, obtain sample data, so Sample data is normalized afterwards；

S2: the principal element of analyzing influence battery SOC constructs BP neural network structure, detailed process are as follows:

S2-1: the relationship analysis of battery SOC and open-circuit voltage is carried out:

The electromotive force of battery is the polarizing voltage three parts structure by the open-circuit voltage of battery, the ohm voltage drop of battery and battery At.After battery is switched to static condition from charging and discharging state, the chemical reaction of inside battery tends towards stability, and battery is opened at this time Road voltage (OCV) numerically with battery end voltage it is equal be battery electromotive force；It is bent according to SOC-OCV as shown in Figure 2 Line estimates the SOC of battery；

S2-2: the relationship analysis of battery SOC and temperature is carried out:

Temperature directly affects the actually available capacity of battery；When environment temperature is higher, the chemical reaction ratio of inside battery More active, the active volume of battery is larger；And when the temperature of the surroundings is low, the utilization rate of active material is lower, and battery can be used Capacity reduces；In the case of different temperatures, the relation curve of voltage and SOC are as shown in Figure 3；

S2-3: the relationship analysis of battery SOC and discharge current is carried out:

When one timing of temperature, discharge test is carried out with 1C, 3C current versus cell, different multiplying as shown in Figure 4 is obtained and puts Voltage and capacity relationship curve when electric；

As shown in Figure 4, when discharge-rate increase when, battery discharge platform voltage is gradually reduced, when electric current is larger, battery from When operating voltage is to discharge cut-off voltage, the electricity of releasing is less；Also, in battery charge state between 20% to 90% When, the curvilinear trend of battery steady state voltage and battery SOC is relatively fixed, illustrates to have therebetween metastable non-linear Relationship, discharge voltage can with SOC reduce and gradually decrease, especially electric discharge latter stage, cell discharge voltage with SOC change Rate is larger；

By above-mentioned analysis, the discharge current I, battery voltage U and environment temperature T of battery are chosen as BP nerve net The input vector of the input layer of network structure, and three impact factors are independent of one another, and the output vector of network is battery SOC；BP neural network Construction of A Model is as shown in Figure 5.

S3: algorithm parameter initialization；The topological structure for inputting BP neural network algorithm is (including input number of layers, implicit Number of layers and output number of layers), minimum allowable error amount；The population scale of CSO algorithm is inputted, dimensionality of particle number is maximum The number of iterations, lateral cross probability and crossed longitudinally probability.

S4: the connection weight and every threshold value between the output valve and each layer of BP neural network are calculated, detailed process is such as Under:

X (k)=(x₁(k),x₂(k),...,x_n(k))；

d₀(k)=(d₁(k),d₂(k),...,d_q(k))；

ho_h(k)=f (hi_h(k)) h=1,2 ..., p；

yo_o(k)=f (yi_o(k)) o=1,2 ..., q；

Error function

S5: according to the input value and output valve of BP neural network, with the real output value of BP network and desired output Fitness function of the mean square error as CSO calculates the adaptation value of each CSO particle according to following formula, obtains of particle Then body optimal value and global optimum make comparisons the optimal value of CSO individual with global optimum, take adaptation value the superior's conduct Current optimal location；

The mapping between COS particle and the weight and threshold value of BP neural network is established, i.e., by the weight of neural network and threshold Value is encoded into real vector to indicate the individual in population, and the group of these vectors is randomly generated into.Specific optimization object master There is the connection weight of the input layer and hidden layer that calculate in step S4: w_ih, hidden layer and output layer connection weight: w_ho, it is hidden The threshold value of each neuron containing layer: b_hAnd the threshold value of each neuron of output layer: b_o, these weights and threshold value constitute CSO algorithm Initial population；Specific Optimization Steps are as follows:

S6-1: lateral operation is carried out to population:

MS_hc(i, d)=r₁×X(i,d)+(1-r₁)×X(j,d)+c₁×(X(i,d)-X(j,d))；

MS_hc(j, d)=r₂×X(j,d)+(1-r₂)×X(i,d)+c₂×(X(j,d)-X(i,d))；

S6-2: crossed longitudinally operation is carried out to population:

MS_vc(i,d₁)=rX (i, d₁)+(1-r)·X(i,d₂)

i∈N(1,M),d₁,d₂∈N(1,D)

Above-mentioned optimization operation, new population is generated by the contention operation of filial generation and parent；If new adaptive value is better than Current individual is optimal, then replaces current individual optimal with the adaptive value: if updated individual optimal value is most better than the current overall situation The figure of merit then replaces current global optimum with the individual optimal value, to complete the optimization to network items weight and threshold value.

S7: step S5 and S6 are repeated, until meeting termination condition；

The present embodiment optimizes classical neural network algorithm by crossover algorithm in length and breadth, by crossover algorithm in length and breadth The advantages of global search sexuality is strong and fast convergence rate organically combines with the stronger capability of fitting of neural network, avoids Neural network falls into local optimum, and improves its convergence rate.In addition, comparing existing battery SOC evaluation method, this implementation Example is suitable for a series of common batteries such as lithium battery, lead-acid battery, and either battery is in standing or use state, can SOC estimation is carried out to battery in real time, and accuracy is high, it is smaller compared to other methods error.

The examples of implementation of the above are only the preferred embodiments of the invention, and implementation model of the invention is not limited with this It encloses, therefore all shapes according to the present invention, changes made by principle, should all be included within the scope of protection of the present invention.

Claims

1. A battery SOC estimation method based on a cross-optimized neural network, characterized in that it comprises the following steps:

S1: Obtain data samples and normalize the sample data;

S2: Analyze the main factors affecting the battery SOC, and construct a BP neural network structure;

S3: initialization of algorithm parameters;

S4: Calculate the output value of the BP neural network and the connection weights and thresholds between the layers;

S5: According to the input value and output value of the BP neural network, the mean square error of the actual output value and the expected output value of the BP network is used as the fitness function of the CSO, and the fitness value of each CSO particle is calculated to obtain the individual optimal value of the particle. The optimal value and the global optimal value are compared, and then the optimal value of the individual CSO is compared with the global optimal value, and the one with the best adaptation value is taken as the current optimal position;

S6: Optimize the weights and thresholds of the BP neural network by using the criss-cross algorithm;

S7: Repeat steps S5 and S6 until the end condition is met;

S8: Use the parameters optimized by the CSO algorithm as the initial weights and thresholds of the BP neural network, and substitute the initial weights and thresholds into the BP neural network algorithm for training; if the output error value of the BP neural network meets the predetermined error accuracy, Then stop the iteration and output the result; otherwise, go back to step S5, and repeat the optimization iteration until the minimum allowable error of the BP neural network algorithm is satisfied.

2. The battery SOC estimation method based on the vertical and horizontal cross optimization neural network according to claim 1, wherein the analysis process in the step S2 is as follows:

S2-1: Analyze the relationship between battery SOC and open circuit voltage, and estimate battery SOC according to the SOC-OCV curve;

S2-2: Analyze the relationship between battery SOC and temperature, and obtain the relationship curve between voltage and SOC under different temperature conditions;

S2-3: Analyze the relationship between battery SOC and discharge current, and obtain the relationship curve between voltage and capacity when discharging at different rates.

3. The battery SOC estimation method based on the vertical and horizontal cross optimization neural network according to claim 1, wherein in the step S2, the discharge current I of the battery, the battery pack voltage U and the ambient temperature T are selected as the BP neural network The input vector of the input layer of the structure, and the three influencing factors are independent of each other, the output vector of the network is the battery SOC, and the BP neural network model is constructed.

4. The battery SOC estimation method based on the vertical and horizontal cross optimization neural network according to claim 1, wherein the step S3 algorithm parameter initialization is specifically: input the topology of the BP neural network algorithm, the minimum allowable error value; input The population size of the CSO algorithm, the number of particle dimensions, the maximum number of iterations, the horizontal crossover probability and the vertical crossover probability.

5. The battery SOC estimation method based on the vertical and horizontal cross optimization neural network according to claim 1, wherein the step S4 calculates the output value of the BP neural network and the connection weight between each layer and each threshold value. The specific process is as follows:

S4-1: Select the kth input sample x(k) and the corresponding expected output d ₀ (k):

_x (k) ₌ (x1(k), _x2 (k),...,xn(k));

d ₀ (k)=(d ₁ (k),d ₂ (k),...,d _q (k));

S4-2: Calculate the input hi _h (k) and output ho _h (k) of neurons in the hidden layer and the input yi _o (k) and output yo _o (k) of neurons in the output layer:

ho _h (k)=f(hi _h (k)) h=1,2,...,p;

yo _o (k)=f(yi _o (k)) o=1,2,...,q;

S4-3: According to the expected output and actual output of the output layer and the input of the output layer, calculate the partial derivative of the function e to each neuron in the output layer:

Error function

S4-4: According to the sensitivity δo(k) of the output layer, the connection weight w of the hidden layer and the input value of the output layer, calculate the partial derivative of the error function to each neuron in the hidden layer:

S4-5: Use the partial derivative in step S4-3 to correct the connection weights of the output layer:

S4-6: Use the partial derivative in step S4-4 to correct the connection weight of the hidden layer:

6. The method for estimating battery SOC based on a crisscross optimization neural network according to claim 1, wherein the formula for calculating the fitness value of each CSO particle in the step S5 is as follows:

In the formula, _yoi (i) is the actual output value of the neural network; _doi is the expected output value of the neural network; E(i) is the mean square error between the actual output value and the expected output value of the neural network.

7. The battery SOC estimation method based on the vertical and horizontal cross-optimized neural network according to claim 1, is characterized in that, in the described step S6, the mapping between the weights and the thresholds of the COS particle and the BP neural network is established, that is, the neural network. The weights and thresholds are encoded into real vectors to represent the individuals in the population, and the population of vectors is randomly generated; the specific optimization steps are as follows:

S6-1: Perform lateral operations on the population:

Horizontal crossover is an arithmetic crossover between the same dimension of two different individual particles in the population; assuming that the d-th dimension of parent individual particles X(i) and X(j) are crossed horizontally, then they produce offspring The formula is as follows:

MS _hc (i,d)=r ₁ ×X(i,d)+(1-r ₁ )×X(j,d)+c ₁ ×(X(i,d)-X(j,d)) ;

MS _hc (j,d)=r ₂ ×X(j,d)+(1-r ₂ )×X(i,d)+c ₂ ×(X(j,d)-X(i,d)) ;

In the formula: r ₁ , r ₂ are random numbers between [0, 1]; c ₁ , c ₂ are random numbers between [-1, 1]; X(i,d), X(j,d ) are the d-th dimension of the individual particles X(i) and X(j) in the parent population, respectively; MS _hc (i,d) and MS _hc (j,d) are X(i,d) and X(j, respectively , d) the d-dimensional offspring generated by horizontal crossover;

Among them, r ₁ ×X(i,d) in the first formula is the memory term of the particle X(i), which is the current optimal value of the particle itself; (1-r ₁ )×X(j,d) is the particle The group cognitive terms of X(i) and X(j) represent the mutual influence between different particles; these two terms are well combined by the inertia weight factor r ₁ ; c ₁ is the learning factor, and the third term c ₁ ×(X(i,d)-X(j,d)) can increase the search space and optimize at the edge; after the horizontal crossover operation is completed, the obtained median solution MS _hc (i, d), MS _hc (j, d) It must be compared with the fitness of parent particles X(i) and X(j) respectively. Only the moderate solution with better fitness can be retained and become the dominant solution DS _hc to participate in the next iteration;

S6-2: Perform a vertical crossover operation on the population:

Longitudinal crossover is an arithmetic crossover between two different dimensions of a particle in the population; assuming that the d _1st dimension and the d _2nd dimension of the particle X(i) are involved in the total crossover, the median solution is generated according to the following formula MS _vc (i,d ₁ ):

MS _vc (i,d ₁ )=r·X(i,d ₁ )+(1-r)·X(i,d ₂ )

i∈N(1,M),d ₁ ,d ₂ ∈N(1,D)

In the formula: i∈[0,1]; MS _vc (i,d ₁ ) is the d ₁ -dimensional descendant generated by the longitudinal intersection of the d ₁ -th dimension and the d-th ₂ -dimension of the individual particle X(i); the first term is The memory term of the d _1st dimension of the particle X(i), the second term is the interaction of the d _1st dimension and the d _2nd dimension of the particle X(i), which are combined together by the inertia weight factor r; the obtained mean solution MS _vc (i,d ₁ ) contains the d 1-dimensional information of the parent particle X(i) and the d ₂ -dimensional information of X(i) with a certain probability, and will not destroy the d ₂ _- dimensional information of X(i) ; Compare the fitness of the moderate solution MS _vc (i,d ₁ ) with the parent particle X(i), and retain the better solution as the dominant solution DS _vc for the next iteration;

A new population is generated through the competitive operation between the offspring and the parent; if the new fitness value is better than the current individual optimal value, the current individual optimal value is replaced by the fitness value: if the updated individual optimal value is better than the current global optimal value The optimal value is used to replace the current global optimal value, so as to complete the optimization of the network weights and thresholds.