Multi-objective optimization method and system based on neural network model and genetic algorithm

Abstract

The invention relates to the technical field of motor optimization, in particular to a multi-objective optimization method and system based on a neural network model and a genetic algorithm, and the method comprises the steps: defining a six-dimensional input parameter structure space of motor structure parameters; sample points are generated by adopting Latin hypercube sampling, torque pulsation and output power of the sample points are obtained through finite element simulation, and a data set is constructed; a BP neural network model is initialized based on the data set, and a three-layer feed-forward network architecture is constructed as the BP neural network model and trained; a first-layer genetic algorithm is adopted to optimize the weight and bias terms of the BP neural network model; and taking the optimized BP neural network model as a dual-target fitness function, and searching an optimally output structure parameter through a second-layer genetic algorithm. Through the neural network model and in combination with the two-layer genetic algorithm, staged optimization is performed, the optimization efficiency is improved, the calculation cost is reduced, and the motor performance stability and reliability are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motor optimization, in particular to a multi-objective optimization method and system based on a neural network model and a genetic algorithm. BACKGROUND

[0002] In the field of reluctance motor design optimization, traditional methods mainly rely on finite element simulation or physical prototype testing. However, finite element simulation has the problems of repeated iterative calculation, time-consuming for a single simulation of several hours or even longer, and difficulty in multi-objective trade-off.

[0003] Existing optimization methods include constructing an approximate model to replace time-consuming finite element calculation to improve efficiency, but such methods are prone to model fitting accuracy being affected by the jumping nature of discrete variables when dealing with continuous-discrete mixed variables, and it is difficult to effectively embed engineering constraints. In addition, optimization methods such as particle swarm algorithm and genetic algorithm have the problems of long calculation time and large consumption of computing resources, especially in multi-objective optimization, there is a lack of efficient and accurate collaborative solution. Therefore, the existing methods generally have the problem of efficiency and accuracy trade-off in multi-objective collaborative optimization. Full finite element simulation directly optimizes with high accuracy but extremely low efficiency, and simple proxy models improve efficiency but cannot guarantee the accuracy of multi-objective optimization.

[0004] Therefore, it is necessary to provide a new multi-objective optimization method and system based on a neural network model and a genetic algorithm. SUMMARY

[0005] Based on the above problems existing in the prior art, the purpose of the embodiments of the present application is to provide a multi-objective optimization method and system based on a neural network model and a genetic algorithm, which improves optimization efficiency, reduces calculation cost, and guarantees motor performance stability and reliability through neural network model and combined double-layer genetic algorithm stage optimization.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is: a multi-objective optimization method based on a neural network model and a genetic algorithm, comprising: S1, defining a 6-dimensional input parameter structure space of motor structure parameters, including stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, length and armature winding turns; S2, generating sample points using Latin hypercube sampling, obtaining torque ripple and output power of the sample points through finite element simulation, and constructing a data set; S3, initializing a BP neural network model based on the data set, including data normalization, dividing the training set and the test set, constructing a three-layer feedforward network architecture as the BP neural network model and training; S4, using the first layer genetic algorithm to optimize the weights and bias terms of the BP neural network model, and then obtaining the optimal BP neural network model; S5, the optimized BP neural network model is used as a double target fitness function, and a second layer genetic algorithm is used to find the optimal output structure parameter; S6, the globally optimized structure parameter combination is applied, and the actual optimization effect of the torque pulsation and the output power is verified through finite element analysis.

[0007] Further, in S1, the 6-dimensional input parameter structure space of the motor structure parameters is defined, and the structure parameters include stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, length and armature winding turns. Step S11, according to the motor power size equation and the torque equation, six structure parameters are selected as optimization variables; Step S12, define the continuous optimization interval of the six structure parameters; Step S13, construct a 6-dimensional input parameter structure space.

[0008] Further, the stator yoke height wherein, is the lower limit of the stator yoke height, is the upper limit of the stator yoke height; the stator pole arc coefficient wherein, is the lower limit of the stator pole arc coefficient, is the upper limit of the stator pole arc coefficient; the rotor yoke height wherein, is the lower limit of the rotor yoke height, is the upper limit of the rotor yoke height; the rotor pole arc coefficient wherein, is the lower limit of the rotor pole arc coefficient, is the upper limit of the rotor pole arc coefficient; the motor length wherein, is the lower limit of the motor length, is the upper limit of the motor length; the armature winding turns wherein, is the lower limit of the armature winding turns, is the upper limit of the armature winding turns.

[0009] Further, in S2, the sample points are generated by Latin hypercube sampling, the torque pulsation and the output power of the sample points are obtained by finite element simulation, and the data set is constructed. Step S21, 200 sample points are randomly generated in the 6-dimensional input parameter structure space by using Latin hypercube sampling, and the parameters are independently and uniformly distributed, the maximum minimum distance criterion is met, and the correlation coefficient constraint is met; Step S22, for the 200 sample points, the finite element simulation is called by MATLAB to calculate the torque fluctuation and output power of each sample point; Step S23, constructing a data set , =[ , , , , , ], is the 6-dimensional structure parameter vector of the i-th sample point, is the stator yoke height of the i-th sample point, is the stator pole arc coefficient of the i-th sample point, is the rotor yoke height of the i-th sample point, is the rotor pole arc coefficient of the i-th sample point, is the length of the i-th sample point, is the number of turns of the armature winding of the i-th sample point, is the maximum difference value of the torque of a single sample point, is the average output power calculated for a single sample point.

[0010] Further, in S3, the BP neural network model is initialized based on the data set, including data normalization, division of training set and test set, construction of three-layer feedforward network architecture as BP neural network model and training, including: Step S31, extracting input data and output data according to the data set; Step S32, normalizing the input and output data to the [0, 1] interval; Step S33, dividing the data set into training set and test set in the ratio of 8:2; Step S34, using a three-layer feedforward network architecture as the BP neural network model, the three-layer feedforward network architecture includes an input layer, a first hidden layer, a second hidden layer and an output layer, and the network parameters are initialized; Step S35, training the BP neural network model by using the Levenberg-Marquardt back propagation algorithm; Step S36, denormalizing the output and saving the parameters and normalization parameters of the BP neural network model.

[0011] Further, the data set , =[ ， ， ， ， ， ], is the 6-dimensional structure parameter vector of the i-th sample point; The input data is X=[ ， ， ， ， ，N ]; The output data is Y=[ ]; wherein, is the stator yoke height, is the stator pole arc coefficient, is the rotor yoke height, is the rotor pole arc coefficient, is the motor length, is the number of turns of the armature winding, is the maximum torque difference, is the output power.

[0012] Further, in S4, the first layer genetic algorithm is used to optimize the weight and bias term of the BP neural network model, and then the optimal BP neural network model is obtained, which includes: Step S41, taking the weight matrix and bias term of the BP neural network model as the optimization variable, and the total parameter number is:

[0013] wherein, is the number of weights input to the hidden layer, is the number of hidden layer bias, is the number of weights from the hidden layer to the output layer, is the number of output layer bias; Step S42, minimizing the mean square error of the predicted output value and the true output value as the optimization target; Step S43, using the first layer genetic algorithm for iterative optimization; Step S44, decoding the optimal parameters output by the first layer genetic algorithm into the weight matrix and bias term, and then updating the BP neural network model parameters to obtain the optimized BP neural network model.

[0014] Further, the minimizing the mean square error of the predicted output value and the true output value as an optimization target comprises: calculating the mean square error (MSE) as an original error index by comparing the predicted output value and the true output value, taking the minimum root mean square error as the optimization target, and the smaller the root mean square error is, the higher the fitness is; the calculation formula of the mean square error is:

[0015] wherein, is the optimal target, is the minimum mean square error, and n is the number of samples, is the true output value of the i th sample, is the predicted output value of the i th sample by the model.

[0016] Further, in S5, the optimized BP neural network model is taken as a double-target fitness function, and a second layer genetic algorithm is used to find the optimal output structure parameter, comprising: Step S51, determining the input variable range and performing even number constraint processing on the armature winding number N; Step S52, establishing a double-target optimization function as wherein, is the torque difference value, is the actual output power, is the target output power, , is a weighting factor; Step S53, using a second layer genetic algorithm for configuration and iterative optimization; Step S54, restoring the armature winding number, outputting the optimal parameter, and inputting the optimal parameter into the optimized BP neural network model to output the optimal torque ripple and output power.

[0017] The multi-objective optimization system based on the neural network model and the genetic algorithm is applied to the multi-objective optimization method based on the neural network model and the genetic algorithm, and the system comprises: A motor parameter definition module is configured to define a 6-dimensional input parameter structure space of motor structure parameters, and the structure parameters include stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, length of revolution, and armature winding number. A data set construction module is configured to generate sample points by using Latin hypercube sampling, to obtain torque ripple and output power of the sample points by finite element simulation, and to construct a data set. A modeling module is configured to initialize a BP neural network model based on the data set, including data normalization, division of training set and test set, construction of a three-layer feedforward network architecture as the BP neural network model, and training. a network model parameter optimization module, configured to optimize the weight and bias term of the BP neural network model by using a first genetic algorithm, and then obtain an optimal BP neural network model; a motor structure parameter optimization module, configured to take the optimized BP neural network model as a double-target fitness function, and find the optimal output structure parameter by using a second genetic algorithm; an optimization result verification module, configured to apply the globally optimized structure parameter combination, verify the actual optimization effect of the torque ripple and output power by using finite element analysis.

[0018] The beneficial effects of the present application are as follows: the multi-objective optimization method based on the neural network model and the genetic algorithm includes the following steps: defining a 6-dimensional input parameter structure space of the motor structure parameters, the structure parameters including the stator yoke height, the stator pole arc coefficient, the rotor yoke height, the rotor pole arc coefficient, the length of the rotor and the number of turns of the armature winding; generating sample points by using Latin hypercube sampling, obtaining the torque ripple and the output power of the sample points by using finite element simulation, and constructing a data set; initializing the BP neural network model based on the data set, including data normalization, dividing the training set and the test set, constructing a three-layer feedforward network architecture as the BP neural network model and training; optimizing the weight and bias term of the BP neural network model by using the first genetic algorithm, and then obtaining the optimal BP neural network model; taking the optimized BP neural network model as a double-target fitness function, and finding the optimal output structure parameter by using the second genetic algorithm; applying the globally optimized structure parameter combination, verifying the actual optimization effect of the torque ripple and the output power by using finite element analysis. The multi-objective optimization method based on the neural network model and the genetic algorithm of the present application replaces the time-consuming finite element calculation by using the neural network proxy model, and combines the double-layer genetic algorithm for stage optimization, thereby improving the optimization efficiency, significantly reducing the consumption of computing resources and time cost, achieving low torque ripple while controlling the output power within the deviation range of ±5% of the target value, solving the multi-objective trade-off problem between low torque ripple and specified output power, and balancing the stability and reliability of the motor performance; the hierarchical optimization strategy of the double-layer genetic algorithm automatically satisfies the engineering constraints such as the number of turns of the winding, and reduces the number of manual corrections. BRIEF DESCRIPTION OF DRAWINGS

[0019] The present application will be further described below in combination with the drawings and examples.

[0020] In the drawings: Figure 1 The flowchart of the multi-objective optimization method based on the neural network model and the genetic algorithm provided for the first embodiment of the present application; Figure 2 The schematic diagram of the BP neural network model provided for the first embodiment of the present application; Figure 3A double-layer genetic algorithm schematic diagram provided for the first embodiment of the present application; Figure 4 A module schematic diagram of the multi-objective optimization system based on the neural network model and the genetic algorithm provided for the second embodiment of the present application; Figure 5 A structure schematic diagram of a network side server provided according to the third embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0022] First Embodiment The first embodiment of the present application provides a multi-objective optimization method based on a neural network model and a genetic algorithm, comprising: defining a 6-dimensional input parameter structure space of motor structure parameters, the structure parameters including stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, length of overhang and armature winding number of turns; generating sample points by Latin hypercube sampling, obtaining torque ripple and output power of the sample points by finite element simulation, and constructing a data set; initializing a BP neural network model based on the data set, including data normalization, dividing a training set and a test set, constructing a three-layer feedforward network architecture as the BP neural network model and training; optimizing the weights and bias terms of the BP neural network model by a first layer of genetic algorithm, and then obtaining an optimal BP neural network model; taking the optimized BP neural network model as a double-objective fitness function, and finding optimal output structure parameters by a second layer of genetic algorithm; combining the globally optimized structure parameters for application, verifying the actual optimization effect of torque ripple and output power by finite element analysis. The multi-objective optimization method based on the neural network model and the genetic algorithm of the present application replaces time-consuming finite element calculation by a neural network surrogate model, and improves optimization efficiency by combining with a double-layer genetic algorithm for stage-by-stage optimization, significantly reduces the consumption of computing resources and time cost, realizes low torque ripple while controlling the output power within a deviation range of ±5% of the target value, solves the multi-objective trade-off problem between low torque ripple and specified output power, and takes into account the stability and reliability of motor performance; through the hierarchical optimization strategy of the double-layer genetic algorithm, engineering constraints such as winding number of turns are automatically met, and the number of manual corrections is reduced.

[0023] The implementation details of the multi-objective optimization method based on the neural network model and the genetic algorithm of the present embodiment will be specifically described below. The following details are provided for the convenience of understanding, and are not essential for implementing the present solution. The specific process of the present embodiment is as shown in Figure 1 The present embodiment is applied to a multi-objective optimization system based on a neural network model and a genetic algorithm.

[0024] Step S1, define a 6-dimensional input parameter structure space of motor structure parameters, including stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, slot length, and armature winding turns.

[0025] Specifically, the definition of the 6-dimensional input parameter structure space of the motor structure parameters, including the stator yoke height, the stator pole arc coefficient, the rotor yoke height, the rotor pole arc coefficient, the slot length, and the armature winding turns includes the following steps: Step S11, select six structure parameters as optimization variables according to the motor power size equation and the torque equation.

[0026] Specifically, according to the power size equation and the torque equation, the key influencing factors of motor output power and torque ripple are analyzed, and the motor structure parameters are determined as the main design variables. The six selected structure parameters specifically include: stator yoke height , stator pole arc coefficient , rotor yoke height , rotor pole arc coefficient , slot length , and armature winding turns .

[0027] Step S12, define the continuous optimization interval of the six structure parameters.

[0028] Specifically, the stator yoke height , wherein is the lower limit of the stator yoke height, is the upper limit of the stator yoke height.

[0029] The stator pole arc coefficient , wherein is the lower limit of the stator pole arc coefficient, is the upper limit of the stator pole arc coefficient.

[0030] The rotor yoke height , wherein is the lower limit of the rotor yoke height, is the upper limit of the rotor yoke height.

[0031] The rotor pole arc coefficient , wherein is the lower limit of the rotor pole arc coefficient, is the upper limit of the rotor pole arc coefficient.

[0032] motor length wherein, motor length lower limit, motor length upper limit.

[0033] armature winding turns wherein, armature winding turns lower limit, armature winding turns upper limit.

[0034] Step S13, construct a 6-dimensional input parameter structure space.

[0035] Specifically, six structure parameters are combined into a six-dimensional input space X, and the mathematical expression is: X=[ ， ， ， ， ，N ].

[0036] Each dimension corresponds to a structure parameter and its optimization interval, forming a complete parameter search space.

[0037] Step S2, generate sample points using Latin hypercube sampling, obtain torque ripple and output power of sample points through finite element simulation, and construct a data set.

[0038] Specifically, the Latin hypercube sampling is used to generate sample points, and the torque ripple and output power of the sample points are obtained through finite element simulation to construct a data set, which includes the following steps: Step S21, generate 200 sample points in the 6-dimensional input parameter structure space using Latin hypercube sampling, and simultaneously satisfy the independent uniform distribution of parameters, the maximum minimum distance criterion and the correlation coefficient constraint.

[0039] Specifically, the independent uniform distribution of parameters is that each dimension parameter is uniformly divided into 200 equal-width subintervals in its value interval.

[0040] As an example, the stator yoke height For example, take 200 sample points and divide them into 200 equal intervals, and each interval has a width of:

[0041] For the first sample, the corresponding interval is Then, a point is randomly selected from each interval as a sample value.

[0042] The maximum-minimum distance criterion realizes the uniform distribution of data points in the whole space by maximizing the minimum distance between any two data points, thereby ensuring the space filling.

[0043] As an example, the stator yoke height is taken as For example, let the sample set be wherein is a sample point in the m-dimensional space, the objective function of the maximum-minimum distance criterion is:

[0044] wherein, is the sample set in the stator yoke height is the minimum value of the distance between sample points in all cases where m≠n, is the distance between two sample points and .

[0045] The correlation coefficient constraint is introduced to avoid strong correlation between parameters. There may be a certain correlation between different parameters, which will weaken the representativeness of the sample. Therefore, the correlation of the sample points needs to be constrained to ensure that the correlation coefficient between parameters does not exceed a certain threshold.

[0046] The correlation coefficient matrix of the sample points is defined as:

[0047] wherein, is a random variable corresponding to each parameter dimension, is the covariance of and , is the standard deviation of , is the standard deviation of . By limiting the absolute value of the non-diagonal elements of the correlation coefficient matrix to not exceed the threshold, the strong correlation between parameters is avoided.

[0048] As an example, the threshold is set to 0.25.

[0049]

[0050] Step S22, for 200 sample points, MATLAB is called to calculate the torque ripple and output power of each sample point by finite element simulation.

[0051] Specifically, the calculation formula of the torque ripple is: ​​​​

[0052] wherein, is the maximum torque value calculated for a single sample point, is the minimum torque value calculated for a single sample point, is the maximum torque difference value calculated for a single sample point, is the number of samples.

[0053] The calculation formula of the output power is:

[0054] wherein, is the output power value calculated for a single sample point, is the total output power sampling length, is the DSEM rotor position angle, is the average output power calculated for a single sample point.

[0055] Step S23, constructing a data set wherein, [ , , , , , ], is the 6-dimensional structure parameter vector of the i-th sample point, is the stator yoke height of the i-th sample point, is the stator pole arc coefficient of the i-th sample point, is the rotor yoke height of the i-th sample point, is the rotor pole arc coefficient of the i-th sample point, is the length of the i-th sample point, is the number of turns of the armature winding of the i-th sample point, is the maximum torque difference value calculated for a single sample point, is the average output power calculated for a single sample point.

[0056] Step S3, initializing the BP neural network model based on the data set, including data normalization, dividing the training set and the test set, constructing a three-layer feedforward network architecture as the BP neural network model and training.

[0057] Specifically, the BP neural network model based on the data set is initialized, including data normalization, dividing the training set and the test set, constructing a three-layer feedforward network architecture as the BP neural network model and training includes the following steps: Step S31, extracting input data and output data according to the data set.

[0058] Specifically, the data set wherein, =[ ， ， ， ， ， ], is the 6-dimensional structural parameter vector of the i-th sample point.

[0059] The input data is X=[ ， ， ， ， ，N ].

[0060] The output data is Y=[ ].

[0061] in, is the stator yoke height, is the stator pole arc coefficient, is the rotor yoke height, is the rotor pole arc coefficient, For the motor stack length, is the number of turns of the armature winding, is the maximum torque difference, is the output power.

[0062] Step S32: normalize the input and output data and scale them to the range [0, 1].

[0063] Specifically, the normalization formula is:

[0064] in, For the dataset A sample point of a single input or output in and The data sets The global minimum and maximum values ​​of a single input or output in , is the normalized data value.

[0065] Step S33: Divide the data set into a training set and a test set in a ratio of 8:2.

[0066] Specifically, let the total number of samples be n and the number of training samples be , the number of test samples is .

[0067] As an example, the sample number is 200, the training sample number is 160, and the test sample number is 40.

[0068] Step S34, a three-layer feedforward network architecture is used as the BP neural network model, the three-layer feedforward network architecture includes an input layer, a first hidden layer, a second hidden layer and an output layer, and the network parameters are initialized.

[0069] Specifically, the node number of the input layer is , has 6 neurons, and strictly corresponds to the 6 dimensions of the motor input parameters .

[0070] The node number of the first hidden layer is , has 10 neurons, and the activation function is the hyperbolic tangent function, and the expression is

[0071] The node number of the second hidden layer is , has 10 neurons, and the activation function is the hyperbolic tangent function, which is used for secondary extraction of nonlinear features.

[0072] The node number of the output layer is , has 2 neurons, corresponds to the torque ripple and output power of the motor, and the activation function is a linear function, and the expression is

[0073] The Nguyen-Widrow initialization algorithm is used to initialize the network parameters to assign weights and biases.

[0074] The weight matrix is initialized to a random value in the interval .

[0075] Wherein, , is the number of hidden layer neurons.

[0076] The bias term is initialized to a random value in the interval .

[0077] The BP neural network model is shown in Figure 2 , and the complete formula chain is

[0078] Wherein, is the reference starting value of each output variable in the original data, that is, the data set of each output variable The minimum value in , The actual variation range of each output variable in the original data set, that is, The maximum value in Subtract the minimum value .

[0079] Step S35: Using the Levenberg-Marquardt back propagation algorithm to train the BP neural network model.

[0080] Specifically, the maximum number of iterations is set to 1000, and the termination condition is enforced to prevent overfitting.

[0081] Set the target error to , early termination is performed when the mean square error (MSE) is lower than a threshold.

[0082] The formula for mean square error is:

[0083] in, For the The true value of the sample , For the The predicted value of samples , is the total number of samples.

[0084] The learning rate is adaptive, and the step size is dynamically adjusted according to the curvature of the error surface. The update formula is:

[0085] in, is the Jacobian matrix of error versus weight (m is the number of samples, n is the total number of parameters), is the output error vector, is the damping factor, is the identity matrix.

[0086] The training process involves normalizing the training set Input network, forward propagation calculation prediction output .

[0087] Calculate the output layer error .

[0088] The hidden layer error gradient is calculated layer by layer using the chain rule. The calculation formula is:

[0089] in, For the Layer error term, is the Hadamard product, is the derivative of the activation function, is the weighted input of the i-th layer (i.e., the input of the activation function).

[0090] is to align the error term of the next layer to the current layer, is to calculate the contribution of the current layer to the total error. is to introduce the derivative of the activation function of the current layer, adjust the gradient direction. Through the error gradient, the network parameters are accurately corrected, and the optimal mapping relationship is approximated.

[0091] Step S36, the output is de-normalized and the parameters and normalization parameters of the BP neural network model are saved Specifically, the de-normalization restores the network output to the original dimension, and the calculation formula is:

[0092] wherein, is the original data value after de-normalization, is the normalized data value, is the maximum value in the original data set, is the minimum value in the original data set.

[0093] Step S4, the first layer genetic algorithm is used to optimize the weight and bias term of the BP neural network model, and then the optimal BP neural network model is obtained.

[0094] Specifically, as shown in Figure 3 , the first layer genetic algorithm is used to optimize the weight and bias term of the BP neural network model, and then the optimal BP neural network model is obtained. The steps include: Step S41, taking the weight matrix and bias term of the BP neural network model as the optimization variable, the total parameter number is:

[0095] wherein, is the number of weights input to the hidden layer, is the number of hidden layer bias, is the number of hidden layer to output layer weights, is the number of output layer bias.

[0096] Step S42, the mean square error between the predicted output value and the true output value is minimized as the optimization target.

[0097] ​Specifically, by comparing the predicted output value with the true output value, the mean square error (MSE) is calculated as the original error indicator, and the minimum root mean square error is taken as the optimization target. The smaller the root mean square error, the higher the fitness. The calculation formula of the mean square error is:

[0098] wherein, is the optimal target, is the minimum mean square error, and n is the number of samples, is the true output value of the i-th sample, is the predicted output value of the i-th sample by the model.

[0099] Step S43, the first layer genetic algorithm is used for iterative optimization.

[0100] Specifically, the parameters are spliced into binary chromosome coding, and roulette selection is used to select individuals with higher fitness from the current population as parents. The probability of each individual entering the next generation is equal to the ratio of its fitness value to the sum of the fitness values of the individuals in the entire population. The number of individuals in the last generation are reserved to the next generation at a ratio of .

[0101] The selected parent individuals are subjected to a crossover operation to exchange part of the chromosome genes with a certain crossover probability, generating new individuals. A crossover point is randomly set in the coding string of the adjacent two individuals, and then part of the chromosome of the two paired individuals is exchanged at the point.

[0102] The chromosome genes of the new individuals are subjected to a mutation operation with a certain mutation probability to increase the diversity of the population. The bits specified randomly in the individual coding string are only flipped, i.e., from 0 to 1 or from 1 to 0, with a mutation probability.

[0103] Step S44, the optimal parameters output by the first layer genetic algorithm are decoded into a weight matrix and a bias term, and then the BP neural network model parameters are updated to obtain an optimized BP neural network model.

[0104] Specifically, the optimal chromosome output by the first layer genetic algorithm is decoded into a weight matrix and a bias; then the BP neural network parameters are updated to form an optimized BP neural network model.

[0105] Step S5, the optimized BP neural network model is taken as a double-objective fitness function, and the second layer genetic algorithm is used to find the optimal output structure parameters.

[0106] Specifically, the optimized BP neural network model is used as a double-target fitness function, and a second layer genetic algorithm is used to find the optimal output structure parameter, including the following steps: Step S51, determine the input variable range and perform even number constraint processing on the armature winding number N.

[0107] Specifically, the input variable range is determined according to the normalized range of the training data, , . for a set of input parameters, , .

[0108] The armature winding number N is forced to be even, and after converting the range to an integer and multiplying by 2: N = 2 * int (N) .

[0109] Initialize population generation (use binary encoding): , uniformly distributed.

[0110] Step S52, establish a double-target optimization function as , wherein is the torque difference, is the actual output power, is the target output power, , is the weighting factor.

[0111] Further, the weighting factor is assigned according to the weight of the optimization target.

[0112] The motor performance index prediction formula based on BP neural network is:

[0113] wherein, is the torque difference predicted by the neural network, is the absolute difference between the predicted output power and the target power, is the optimized BP neural network model, is the normalized motor structure parameter.

[0114] Step S53, use a second layer genetic algorithm for configuration and iterative optimization.

[0115] Specifically, the second layer genetic algorithm configuration is: population size is 50, maximum number of generations is 100, fitness tolerance is , and the integer constraint is that the sixth variable (i.e. the armature winding number N) is an integer.

[0116] The roulette wheel selection is used to select the individuals with higher fitness from the current population as the parents. The probability of each individual entering the next generation is equal to the ratio of its fitness value to the sum of the fitness values of all individuals in the population. The individuals in the last generation are retained to the next generation in proportion to .

[0117] The selected parent individuals are subjected to the crossover operation to exchange part of the chromosome genes with a certain crossover probability to generate new individuals. A crossover point is randomly set in the coding string of two adjacent individuals, and then part of the chromosome of the two paired individuals is exchanged at the point.

[0118] The chromosome genes of the new individuals are subjected to the mutation operation with a certain mutation probability to increase the diversity of the population. The bits of the individual coding string randomly specified with the mutation probability are only bit inversion, i.e., from 0 to 1 or from 1 to 0 Step S54, the armature winding number is restored, the optimal parameters are output, and the optimal parameters are input to the optimized BP neural network model to output the optimal torque ripple and output power.

[0119] Specifically, the armature winding number (i.e., the sixth variable) is restored to the actual value. is the result after the sixth variable is restored to the actual value, is the intermediate value obtained after the sixth variable is optimized.

[0120] The output optimal parameters are: [ ].Among them, is the optimal stator yoke height, is the optimal stator pole arc coefficient, is the optimal rotor yoke height, is the optimal rotor pole arc coefficient, is the optimal stack length, is the optimal armature winding number.

[0121] Input to the optimized BP neural network model to output the optimal torque ripple and the output power .

[0122] Step S6, the globally optimized structure parameter combination is applied, and the actual optimization effect of the torque ripple and the output power is verified through finite element analysis.

[0123] The globally optimized structure parameter combination is applied to the actual design of the DSEM (dual salient pole electrically excited motor), and the actual optimization effect of the torque ripple and the output power is verified through finite element analysis.

[0124] The first embodiment of the present application provides a multi-objective optimization method based on a neural network model and a genetic algorithm, comprising: defining a 6-dimensional input parameter structure space of motor structure parameters, the structure parameters including stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, length of overhang and armature winding number of turns; generating sample points by using Latin hypercube sampling, obtaining torque ripple and output power of the sample points by finite element simulation, and constructing a data set; initializing a BP neural network model based on the data set, including data normalization, dividing a training set and a test set, constructing a three-layer feedforward network architecture as the BP neural network model and training; using a first genetic algorithm to optimize the weights and bias terms of the BP neural network model, and then obtaining an optimal BP neural network model; using the optimized BP neural network model as a double-objective fitness function, and finding optimal output structure parameters by using a second genetic algorithm; combining the globally optimized structure parameters for application, verifying the actual optimization effect of torque ripple and output power by finite element analysis. The multi-objective optimization method based on the neural network model and the genetic algorithm of the present application replaces time-consuming finite element calculation by using a neural network proxy model, and combines a double-layer genetic algorithm for stage-by-stage optimization, thereby improving optimization efficiency, significantly reducing calculation resource consumption and time cost, achieving low torque ripple, controlling output power within a deviation range of ±5% of a target value, solving the multi-objective trade-off problem between low torque ripple and specified output power, and taking into account the stability and reliability of motor performance; through the hierarchical optimization strategy of the double-layer genetic algorithm, engineering constraints such as winding number of turns are automatically met, and the number of manual corrections is reduced.

[0125] The step division of the above various methods is only for the purpose of clear description, and in implementation, one step can be combined or some steps can be split and decomposed into multiple steps, as long as the same logical relationship is included, and all are within the protection scope of the present patent; adding insignificant modifications or introducing insignificant designs in the algorithm or flow, but not changing the core design of the algorithm and flow, are within the protection scope of the present patent.

[0126] Second embodiment: As shown in Figure 4 The second embodiment of the present application provides a multi-objective optimization system based on a neural network model and a genetic algorithm, the system comprising: a motor parameter definition module 201, a data set construction module 202, a modeling module 203, a network model parameter optimization module 204, a motor structure parameter optimization module 205, and an optimization result verification module 206.

[0127] Specifically, the motor parameter definition module 201 is used to define the 6-dimensional input parameter structure space of the motor structural parameters, and the structural parameters include stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, stack length and armature winding turns; the data set construction module 202 is used to generate sample points using Latin hypercube sampling, obtain the torque pulsation and output power of the sample points through finite element simulation, and construct them into a data set; the modeling module 203 is used to initialize the BP neural network model based on the data set, including data normalization, dividing the training set and test set, constructing a three-layer feedforward network architecture as the BP neural network model and training it; the network model parameter optimization module 204 is used to optimize the weights and bias terms of the BP neural network model using the first-layer genetic algorithm to obtain the optimal BP neural network model; the motor structural parameter optimization module 205 is used to use the optimized BP neural network model as a dual-objective fitness function and find the structural parameters with the optimal output through the second-layer genetic algorithm; the optimization result verification module 206 is used to apply the globally optimized structural parameter combination and verify the actual optimization effect of torque pulsation and output power through finite element analysis.

[0128] It is not difficult to find that this embodiment is a system embodiment corresponding to the first embodiment, and this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment are still valid in this embodiment, and to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the first embodiment.

[0129] It is worth noting that all modules involved in this embodiment are logical modules. In actual applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovations of the present invention, this embodiment does not include units that are not closely related to solving the technical problems proposed by the present invention. However, this does not mean that other units do not exist in this embodiment.

[0130] The third embodiment of the present invention relates to a network side server, such as Figure 5 As shown, it includes at least one processor 302; and a memory 301 that is communicatively connected to the at least one processor 302; wherein the memory 301 stores instructions that can be executed by the at least one processor 302, and the instructions are executed by the at least one processor 302 to enable the at least one processor 302 to execute the above-mentioned data processing method. The memory 301 and the processor 302 are connected in a bus manner, the bus can include any number of interconnected buses and bridges, the bus connects one or more processors 302 and various circuits of the memory 301 together. The bus can also connect various other circuits such as peripheral devices, voltage stabilizers and power management circuits together, which are well known in the art, therefore, they will not be further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements such as multiple receivers and transmitters, which provide a unit for communicating with various other devices on the transmission medium. The data processed by the processor 302 is transmitted on the wireless medium through the antenna, further, the antenna also receives data and transmits the data to the processor 302.

[0131] The processor 302 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interface, voltage regulation, power management and other control functions. And the memory 301 can be used to store the data used by the processor 302 in the execution operation.

[0132] The fourth embodiment of the application relates to a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the multi-objective optimization method based on the neural network model and the genetic algorithm in the first embodiment.

[0133] That is, those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment methods can be completed by programs instructing related hardware, the programs are stored in a storage medium, and the programs include a plurality of instructions for making a device (which can be a single-chip microcomputer, a chip, etc.) or a processor execute all or part of the steps of the method described in various embodiments of the application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk and various program code storage media.

[0134] The above-mentioned are only embodiments of the present application, and the common knowledge of the specific structure and characteristics in the scheme is not described too much, the ordinary skilled in the art knows all the ordinary technical knowledge in the field of the present application before the application date or the priority date, can know all the prior art in the field, and has the ability to apply the conventional experimental means before the date, the ordinary skilled in the art can improve and implement the present scheme under the inspiration given by the present application, some typical known structure or known method should not become the obstacle for the ordinary skilled in the art to implement the present application. It should be pointed out that for the skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should be considered as the protection scope of the present application, which will not affect the effect and practicality of the present application. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

[0135] The above-mentioned are only embodiments of the present application, and the common knowledge of the specific structure and characteristics in the scheme is not described too much, the ordinary skilled in the art knows all the ordinary technical knowledge in the field of the present application before the application date or the priority date, can know all the prior art in the field, and has the ability to apply the conventional experimental means before the date, the ordinary skilled in the art can improve and implement the present scheme under the inspiration given by the present application, some typical known structure or known method should not become the obstacle for the ordinary skilled in the art to implement the present application. It should be pointed out that for the skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should be considered as the protection scope of the present application, which will not affect the effect and practicality of the present application. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. A multi-objective optimization method based on a neural network model and a genetic algorithm, characterized in that: include: S1, defines the 6-dimensional input parameter structure space of the motor structural parameters, including stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, stack length and number of armature winding turns; S2, Latin hypercube sampling is used to generate sample points, and the torque ripple and output power of the sample points are obtained through finite element simulation to construct a data set; S3, initialize the BP neural network model based on the data set, including data normalization, dividing the training set and test set, building a three-layer feedforward network architecture as the BP neural network model and training; S4, using the first-layer genetic algorithm to optimize the weights and bias items of the BP neural network model, and then obtaining the optimal BP neural network model; S5, using the optimized BP neural network model as a dual-objective fitness function, and searching for the optimal output structural parameters through the second-layer genetic algorithm; S6, applies the globally optimized structural parameter combination and verifies the actual optimization effect of torque pulsation and output power through finite element analysis.

2. The multi-objective optimization method based on a neural network model and a genetic algorithm according to claim 1, characterized in that: In S1, the six-dimensional input parameter structure space defining the motor structural parameters, including stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, stack length and armature winding turns, includes: Step S11, selecting six structural parameters as optimization variables based on the motor power size equation and torque equation; Step S12, defining continuous optimization intervals of six structural parameters; Step S13: construct a 6-dimensional input parameter structure space.

3. The multi-objective optimization method based on a neural network model and a genetic algorithm according to claim 2, characterized in that: The stator yoke height ,in, is the upper and lower limits of the stator yoke, The upper limit of the stator yoke height; The stator pole arc coefficient ,in, is the lower limit of the stator pole arc coefficient, is the upper limit of the stator pole arc coefficient; The rotor yoke height ,in, is the lower limit of the rotor yoke height, The upper limit of the rotor yoke height; The rotor pole arc coefficient ,in, is the lower limit of the rotor pole arc coefficient, is the upper limit of the rotor pole arc coefficient; The motor stack length ,in, is the lower limit of the motor stack length, The upper limit of the motor stack length; The number of turns of the armature winding ,in, is the lower limit of the number of turns of the armature winding, The upper limit of the number of turns of the armature winding.

4. The multi-objective optimization method based on a neural network model and a genetic algorithm according to claim 1, characterized in that: In S2, Latin hypercube sampling is used to generate sample points, and the torque ripple and output power of the sample points are obtained through finite element simulation to construct a data set including: Step S21, using Latin hypercube sampling to randomly generate 200 sample points in the 6-dimensional input parameter structure space, while satisfying the independent uniform distribution of parameters, the maximum and minimum distance criteria, and the correlation coefficient constraint; Step S22: For the 200 groups of sample points, finite element simulation is performed using MATLAB to calculate the torque ripple and output power of each sample point. Step S23: Building a data set ,in, =[ , , , , , ], is the 6-dimensional structural parameter vector of the i-th sample point, is the stator yoke height of the i-th sample point, is the stator pole arc coefficient of the i-th sample point, is the rotor yoke height of the i-th sample point, is the rotor pole arc coefficient of the i-th sample point, is the stack length of the i-th sample point, is the number of turns of the armature winding at the i-th sample point, is the maximum torque difference of a single sample point, The average output power calculated for a single sample point.

5. The multi-objective optimization method based on a neural network model and a genetic algorithm according to claim 1, characterized in that: In S3, the BP neural network model is initialized based on the data set, including data normalization, dividing the training set and the test set, building a three-layer feedforward network architecture as the BP neural network model and training it, including: Step S31, extracting input data and output data according to the data set; Step S32, normalize the input and output data and scale them to the interval [0,1]; Step S33, dividing the data set into a training set and a test set in a ratio of 8:2; Step S34, using a three-layer feedforward network architecture as a BP neural network model, the three-layer feedforward network architecture including an input layer, a first hidden layer, a second hidden layer, and an output layer, and initializing the network parameters; Step S35, using the Levenberg-Marquardt back propagation algorithm to train the BP neural network model; Step S36: Denormalize and output and save the parameters and normalized parameters of the BP neural network model.

6. The multi-objective optimization method based on a neural network model and a genetic algorithm according to claim 1, characterized in that: The dataset ,in, =[ ， ， ， ， ， ], is the 6-dimensional structural parameter vector of the i-th sample point; The input data is X=[ ， ， ， ， ，N ]; The output data is Y=[ ]; in, is the stator yoke height, is the stator pole arc coefficient, is the rotor yoke height, is the rotor pole arc coefficient, For the motor stack length, is the number of turns of the armature winding, is the maximum torque difference, is the output power.

7. The multi-objective optimization method based on a neural network model and a genetic algorithm according to claim 1, characterized in that: In S4, the first-layer genetic algorithm is used to optimize the weights and bias items of the BP neural network model to obtain the optimal BP neural network model, which includes: Step S41, using the weight matrix and bias term of the BP neural network model as optimization variables, the total number of parameters is: in, is the number of weights input to the hidden layer, is the number of hidden layer biases, is the number of weights from the hidden layer to the output layer, is the number of output layer biases; Step S42, minimizing the mean square error between the predicted output value and the true output value as the optimization goal; Step S43, using the first-level genetic algorithm to perform iterative optimization; Step S44: decoding the optimal parameters output by the first-layer genetic algorithm into a weight matrix and a bias term, and then updating the BP neural network model parameters to obtain an optimized BP neural network model.

8. The multi-objective optimization method based on a neural network model and a genetic algorithm according to claim 7, characterized in that: The optimization goal of minimizing the mean square error between the predicted output value and the true output value includes: calculating the mean square error (MSE) as the original error indicator by comparing the predicted output value with the true output value, and taking the minimum root mean square error as the optimization goal. The smaller the root mean square error, the higher the fitness. The calculation formula of the mean square error is: in, For the optimal goal, is the minimum mean square error, n is the number of samples, is the true output value of the i-th sample, is the predicted output value of the model for the i-th sample.

9. The multi-objective optimization method based on a neural network model and a genetic algorithm according to claim 7, characterized in that: In S5, the optimized BP neural network model is used as a dual-objective fitness function, and the structural parameters of the optimal output are found through the second-layer genetic algorithm, including: Step S51, determining the input variable range and performing even number constraint processing on the armature winding turns N; Step S52: Establish a dual-objective optimization function: ,in, is the torque difference, is the actual output power, is the target output power, 、 is the weighting factor; Step S53, using the second-layer genetic algorithm for configuration and iterative optimization; Step S54: restore the armature winding turns, output the optimal parameters, and input the optimal parameters into the optimized BP neural network model to output the optimal torque ripple and output power.

10. A multi-objective optimization system based on a neural network model and a genetic algorithm, characterized by: The multi-objective optimization method based on a neural network model and a genetic algorithm as described in claim 1, wherein the system comprises: The motor parameter definition module is used to define the six-dimensional input parameter structure space of the motor structural parameters, including stator yoke height, stator pole arc coefficient, rotor yoke height, rotor pole arc coefficient, stack length, and number of armature winding turns; A data set construction module is used to generate sample points using Latin hypercube sampling, obtain the torque ripple and output power of the sample points through finite element simulation, and construct a data set; The modeling module is used to initialize the BP neural network model based on the data set, including data normalization, dividing the training set and test set, building a three-layer feedforward network architecture as the BP neural network model and training it; The network model parameter optimization module is used to optimize the weights and bias items of the BP neural network model using the first-layer genetic algorithm to obtain the optimal BP neural network model; The motor structural parameter optimization module is used to use the optimized BP neural network model as a dual-objective fitness function to find the optimal output structural parameters through the second-layer genetic algorithm; The optimization result verification module is used to apply the globally optimized structural parameter combination and verify the actual optimization effect of torque pulsation and output power through finite element analysis.

Images