Optimization method of impeller for micro-turbojet engine based on BP neural network

Abstract

The invention relates to the technical field of turbojet engines, and in particular to an impeller optimization method for a micro-sized turbojet engine based on a BP neural network. The method comprises the following steps: S1, constructing a sample library, wherein the sample library comprises a training sample library and a test sample library; S2, using the training sample library to train an initial BP network model, and using the test sample library to verify the trained initial BP network model, wherein an initial weight and an initial threshold are provided in the initial BP network model; S3, using a PSO-GA phase coupling algorithm to optimize the weight and threshold in the initial BP network model; S4, updating the BP network model to obtain a final BP network model; S5, inputting an optimization target into the final BP network model, outputting variables to be optimized, and completing the structural design of the impeller according to the variables to be optimized, thereby realizing efficient, accurate and low-complexity impeller optimization design.

Description

Technical Field

[0001] The invention relates to the technical field of turbojet engines, and in particular to an impeller optimization method for a micro-sized turbojet engine based on a BP neural network. Background Art

[0002] As an important component of a micro turbojet engine, a micro centrifugal compressor is widely used in military and energy fields due to its compact structure, high single-stage pressure ratio, small flow rate and axial exhaust. The impeller is the core component of a centrifugal compressor that rotates at high speed. The quality of the impeller's structural design is directly related to the service life of the impeller, the accuracy of the generated sinusoidal waveform, and the consumption of electrical energy. Therefore, in the design of modern high-performance micro centrifugal compressors, it is of great significance to optimize the structure of the centrifugal impeller. The design scheme of the main structural dimensions of the compressor impeller, such as the inlet hub radius, rim radius, outlet radius and height, can not only optimize the impeller's own design process, but also provide convenience for the overall machine coordination design.

[0003] The GA algorithm adopts the evolutionary principle of survival of the fittest, and continuously evolves the group containing possible solutions. At the same time, it uses global parallel search technology to search for the best individual in the group to obtain the optimal solution or quasi-optimal solution that meets the requirements. However, in practical applications, there are still problems such as premature maturity and poor local search ability.

[0004] PSO algorithm is an evolutionary computing technology based on swarm intelligence. Similar to GA algorithm, it is also a group-based optimization tool. It randomly initializes particle swarms in a given solution space, and iteratively optimizes by giving individual position and velocity information to the population. Each individual continuously updates and optimizes during the iteration process by tracking the extreme values ​​of position and velocity. It has fast convergence speed, is easy to implement, and has strong local optimization ability.

[0005] The PSO-GA hybrid algorithm has gradually attracted everyone's attention in the field of impeller optimization. It uses their respective advantages and makes up for each other's shortcomings to complete impeller optimization, but the specific algorithm implementation process is closely related to the quality of the optimization results.

[0006] Therefore, there is an urgent need to provide a micro-turbojet engine impeller optimization method based on BP neural network, which can optimize the impeller design efficiently, accurately and with low complexity compared with the existing technology. Summary of the invention

[0007] The present invention solves the technical problems existing in the prior art and provides a method for optimizing the impeller of a micro-sized turbojet engine based on a BP neural network.

[0008] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0009] The impeller optimization method of a micro-turbojet engine based on BP neural network includes the following steps:

[0010] S1. Build a sample library, which includes a training sample library and a test sample library;

[0011] S2, using the training sample library to train the initial BP network model, and using the test sample library to verify the trained initial BP network model, wherein the initial BP network model is provided with initial weights and initial thresholds;

[0012] S3, using the PSO-GA coupled algorithm to optimize the weights and thresholds in the initial BP network model; obtaining the current optimal weights and the current optimal threshold;

[0013] S4, inputting the current optimal weight and the current optimal threshold into the initial BP network model, updating the BP network model, and obtaining the final BP network model;

[0014] S5. Input the optimization target into the final BP network model, output the variables to be optimized, and complete the structural design of the impeller according to the variables to be optimized.

[0015] Furthermore, S3 specifically includes the following steps:

[0016] S31, initialize the population, complete the population segmentation, and obtain sub-population A and sub-population B;

[0017] S32, performing gene recombination and gene mutation on individuals in subpopulation A and subpopulation B respectively;

[0018] S33, calculating the fitness of the individuals in the subpopulation A and the subpopulation B after step S32, and performing local optimization of the fitness. When the constraint conditions are met, the subpopulation A and the subpopulation B are clustered to form a total population. Otherwise, the step S32 is repeated;

[0019] S34, the total population after clustering in step S33 is subjected to gene recombination and gene mutation again, and then global optimization of fitness is performed to obtain the current optimal weight and the current optimal threshold.

[0020] Furthermore, the gene recombination method of the individual in step S32 is: the individual is decoded by real value, and a weight coefficient associated with its fitness is assigned to each individual by a normalization method. The individual after gene recombination is expressed as:

[0021]

[0022] In the above formula, represents the velocity vector of the reorganized individual, Represents the position vector of the reorganized individual, w bestIt represents the weight value corresponding to the best individual in the group of individuals before recombination. represents the velocity vector of the individual before recombination, Represents the velocity vector corresponding to the best individual in the group before recombination, Represents the position vector of the individual before recombination.

[0023] Furthermore, the gene mutation mode of an individual is: the individual has a probability p m The following gene mutations occur:

[0024]

[0025] In the above formula, Δx represents the interval A uniformly distributed random number, represents the fitness value of the individual to be mutated, f avg represents the average fitness value of the population of the individual to be mutated, P m1 =0.1, P m2 =0.001, f max Indicates the maximum fitness value in the group of individuals to be mutated.

[0026] Furthermore, S31 specifically includes the following steps:

[0027] S311, randomly generate a primary population, perform binary coding on all individuals in the primary population, and perform an XOR operation without replacement on every two individuals in the primary population;

[0028] S312, set a critical value, divide the individuals in the primary population whose number of different bits is greater than the critical value into subpopulation A, and divide the remaining individuals into subpopulation B, calculate the error norm of each individual in subpopulation A and subpopulation B, input each error range into the fitness function, and obtain the fitness of each individual;

[0029] S313, performing elite preservation gene selection operations on the individuals in sub-population A and sub-population B respectively, performing GA algorithm selection operations on sub-population A and sub-population B respectively according to the roulette wheel selection method, and then retaining the individuals with the highest fitness in sub-population A and sub-population B to the next generation.

[0030] Furthermore, step S4 also includes: iterating steps S3-S4, and when the termination condition is met, outputting the current optimal weight and the current optimum; in step S4, setting the weight error setting value and the threshold error setting value, calculating the difference between the current optimal weight and the initial weight, and the difference between the current optimal threshold and the initial threshold, setting the difference between the current optimal weight and the initial weight as the weight error, and the difference between the current optimal threshold and the initial threshold as the threshold error, and the termination condition is: the weight error is less than the weight error setting value, and the threshold error is less than the threshold error setting value.

[0031] Furthermore, the BP network model is a three-layer network structure, including an input layer, a hidden layer and an output layer, and the structure is expressed as:

[0032] M1-M2+1;

[0033] Among them, M1 represents the number of nodes in the input layer, that is, the number of variables to be optimized; M2 represents the number of nodes in the hidden layer; and the number of nodes in the output layer is the number of targets to be optimized.

[0034] Furthermore, the expression for the number of nodes in the hidden layer is:

[0035] M2=2×M1+1.

[0036] Furthermore, the expressions of weights and thresholds in the BP network model are:

[0037] M3=M1×M2+M2×2;

[0038] M4=M2+2;

[0039] In the above formula, M3 represents the weight and M4 represents the threshold.

[0040] Further, S1 comprises the following steps:

[0041] S11, using the current design software, setting different optimization target values, inputting various variables to be optimized into the current design software to perform performance prediction of the optimization target; the variables to be optimized include the impeller inlet hub radius, rim radius, outlet radius and height, blade tip clearance, blade thickness, inlet angle of attack and back bend angle; the optimization targets include impeller efficiency and total pressure ratio;

[0042] S12. In the performance prediction of each optimization target, a correlation setting value is set for each optimization target according to the correlation, and a variable to be optimized whose correlation with the optimization target is greater than the corresponding correlation setting value is selected. The selected variable to be optimized and its corresponding optimization target are taken as a sample pair, and all sample pairs constitute a sample library;

[0043] S13. Divide the sample library into a training sample library and a test sample library.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) The present invention combines BP neural network, GA algorithm and PSO algorithm, has population diversity, high fitness, and uses the principle of elite preservation to achieve the purpose of global optimization. Compared with other existing hybrid algorithms, the implementation process is simple, the complexity is low, the optimization is efficient and accurate, and when designing a micro-turbojet engine, it can avoid the high degree of installation and coordination between the impeller and other components, greatly shorten the design cycle of the turbojet engine, and improve the independence of the impeller's own optimization, which is of great significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the present invention.

[0047] Figure 2 The present invention is a flow chart of optimizing the weights and thresholds in the BP network model using the PSO-GA phase coupling algorithm. DETAILED DESCRIPTION

[0048] The technical solution of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention, and all other embodiments obtained by ordinary technicians in the field without making creative work are within the protection scope of the present invention.

[0049] like Figure 1 As shown, the present invention provides a micro-sized turbojet engine impeller optimization method based on BP neural network, comprising the following steps:

[0050] S1. Build a sample library, which includes a training sample library and a test sample library. Specifically, the following steps are included:

[0051] S11. Use the current impeller design software to set different optimization target values, and input various variables to be optimized into the current design software to complete the performance prediction of the optimization target; the variables to be optimized include the impeller's inlet hub radius, rim radius, outlet radius and height, blade tip clearance, blade thickness, inlet angle of attack and back bend angle; the optimization targets include impeller efficiency and total pressure ratio.

[0052] S12. In the performance prediction of each optimization target, based on the correlation, take at least the first 200 variables to be optimized with strong correlation to each optimization target. Each variable to be optimized and its corresponding optimization target are taken as a sample pair. The number of sample pairs is at least 200, and all sample pairs constitute a sample library.

[0053] S13. Divide the sample library into a training sample library and a test sample library. The sample pairs in the sample library need to satisfy the following requirement: the same sample pair does not exist in both the training sample and the test sample.

[0054] S2. Use the training sample library to train the initial BP network model, and use the test sample library to verify the initial BP network model.

[0055] The BP network model is a three-layer network structure, including an input layer, a hidden layer, and an output layer. The structure is as follows:

[0056] M1-M2+1;

[0057] Among them, M1 represents the number of nodes in the input layer, that is, the number of variables to be optimized; M2 represents the number of nodes in the hidden layer; the number of nodes in the output layer is the number of targets to be optimized, that is, 2; the weights and thresholds are set in the BP network model, and the expressions of the number of nodes, weights, and thresholds in the hidden layer are:

[0058] M2 = 2 × M1 + 1;

[0059] M3=M1×M2+M2×2;

[0060] M4=M2+2;

[0061] In the above formula, M3 represents the weight and M4 represents the threshold.

[0062] S3, use the PSO-GA coupled algorithm to optimize the weights and thresholds in the BP network model, such as Figure 2 As shown, the specific steps include:

[0063] S31, initialize the population and complete the population segmentation; specifically:

[0064] S311, randomly generate the first generation population, i.e., the parent generation, perform binary coding on all individuals in the first generation population, and perform binary coding without replacement on every two individuals in the first generation population. Operation, that is, XOR operation.

[0065] S312. Set a critical value. In the primary population, the individuals whose number of different positions is greater than the critical value are divided into subpopulation A, and the remaining individuals are divided into subpopulation B. Calculate the error norm of each individual in subpopulation A and subpopulation B. The error norm is the difference between the target value and the predicted value of each individual. The predicted value is the result calculated by taking values ​​within the empirical range of each parameter using the current impeller design software. Input each error norm into the fitness function to obtain the fitness of each individual.

[0066] S313, and perform elite preservation gene selection operations on sub-population A and sub-population B respectively. The specific method is: perform the selection operation of the GA algorithm on sub-population A and sub-population B respectively according to the roulette selection method, and then retain the individuals with the highest fitness in sub-population A and sub-population B to the next generation, that is, the individuals with the highest fitness in the current sub-population A and sub-population B do not participate in crossover operations and mutation operations, so as to avoid the individuals with the highest fitness in the current sub-population A and sub-population B from being destroyed.

[0067] S32. Perform gene recombination and gene mutation on sub-population A and sub-population B respectively.

[0068] S33, calculating the fitness of the individuals in subpopulation A and subpopulation B that have undergone gene recombination and gene mutation in step S32, and performing local optimization of fitness. When the constraint conditions are met, outputting the better solutions of subpopulation A and subpopulation B, clustering subpopulation A and subpopulation B to form a total population, otherwise, repeating step S32;

[0069] S34, the total population after clustering in step S33 is subjected to gene recombination and gene mutation again, and then global optimization of fitness is performed to obtain the optimization result, that is, the current optimal weight and the current optimal threshold.

[0070] The gene recombination method of individuals is: real-value decoding of individuals in subpopulation A, subpopulation B or the total population, and gene recombination, and a weight coefficient associated with its fitness is assigned to each individual using a normalization method. The recombined individuals are expressed as:

[0071]

[0072]

[0073] In the above formula, represents the velocity vector of the reorganized individual, Represents the position vector of the reorganized individual, w best It represents the weight value corresponding to the best individual in the group of individuals before recombination. represents the velocity vector of the individual before recombination, Represents the velocity vector corresponding to the best individual in the group before recombination, Represents the position vector of the individual before recombination.

[0074] The gene mutation mode of individuals is: individuals in subpopulation A, subpopulation B or the total population with probability p m The following gene mutations occur:

[0075]

[0076] In the above formula, Δx represents the interval A uniformly distributed random number, represents the fitness value of the individual to be mutated, f avg represents the average fitness value of the population of the individual to be mutated, P m1 =0.1, P m2 =0.001, f max Indicates the maximum fitness value in the group where the individual to be mutated belongs.

[0077] S4, output the optimization result to the BP network model, that is, output the current optimal weight and the current optimal threshold to the BP network model, update the BP network model, calculate the weight error between the current optimal weight and the initial weight, and the threshold error between the current optimal threshold and the initial threshold, repeat steps S3-S4 until the end condition is met, output the latest optimal weight and optimal threshold, and obtain the final BP network model; set the weight error setting value and the threshold error setting value, set the difference between the current optimal weight and the initial weight as the weight error, and the difference between the current optimal threshold and the initial threshold as the threshold error, and the end condition is: the weight error is less than the weight error setting value, and the threshold error is less than the threshold error setting value.

[0078] S5. Input the optimization target into the final BP network model obtained in step S4, output the variables to be optimized, and complete the structural design of the impeller.

[0079] The present invention combines BP neural network, GA algorithm and PSO algorithm, and proposes an intelligent optimization algorithm for optimizing the main configuration in the early stage of impeller design. It always adheres to the principles of population diversity, high fitness and elite preservation from beginning to end to achieve the purpose of global optimization. Compared with other existing hybrid algorithms, the present invention has a simple implementation process, low complexity, efficient and accurate optimization, and can avoid the high degree of installation and coordination between the impeller and other components when designing a micro-turbojet engine, greatly shortening the design cycle of the turbojet engine and improving the independence of the impeller's own optimization, which is of great significance.

[0080] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate from the essence and scope of the technical solution of the present invention.

Claims

1. A micro-turbojet engine impeller optimization method based on BP neural network, characterized in that: The following steps are involved: S1. Build a sample library, which includes a training sample library and a test sample library; S2, using the training sample library to train the initial BP network model, and using the test sample library to verify the trained initial BP network model, wherein the initial BP network model is provided with initial weights and initial thresholds; S3, using the PSO-GA coupled algorithm to optimize the weights and thresholds in the initial BP network model; obtaining the current optimal weights and the current optimal threshold; S3 specifically includes the following steps: S31, initialize the population, complete population segmentation, and obtain sub-population A and sub-population B; S31 specifically includes the following steps: S311, randomly generate a primary population, perform binary coding on all individuals in the primary population, and perform an XOR operation without replacement on every two individuals in the primary population; S312, set a critical value, divide the individuals in the primary population whose number of different bits is greater than the critical value into subpopulation A, and divide the remaining individuals into subpopulation B, calculate the error norm of each individual in subpopulation A and subpopulation B, input each error range into the fitness function, and obtain the fitness of each individual; S313, performing elite preservation gene selection operations on the individuals in subpopulation A and subpopulation B respectively, performing GA algorithm selection operations on subpopulation A and subpopulation B respectively according to the roulette wheel selection method, and then retaining the individuals with the highest fitness in subpopulation A and subpopulation B to the next generation; S32, performing gene recombination and gene mutation on individuals in subpopulation A and subpopulation B respectively; S33, calculating the fitness of the individuals in the subpopulation A and the subpopulation B after step S32, and performing local optimization of the fitness. When the constraint conditions are met, the subpopulation A and the subpopulation B are clustered to form a total population. Otherwise, the step S32 is repeated; S34, performing gene recombination and gene mutation again on the total population after clustering in step S33, and then performing global optimization of fitness to obtain the current optimal weight and the current optimal threshold; S4, inputting the current optimal weight and the current optimal threshold into the initial BP network model, updating the BP network model, and obtaining the final BP network model; S5. Input the optimization target into the final BP network model, output the variables to be optimized, and complete the structural design of the impeller according to the variables to be optimized.

2. The micro-sized turbojet engine impeller optimization method based on BP neural network according to claim 1 is characterized in that: The gene recombination method of individuals in step S32 is: real-value decoding of individuals, using the normalization method to assign a weight coefficient associated with their fitness to each individual, and the individuals after gene recombination are expressed as: In the above formula, represents the velocity vector of the reorganized individual, Represents the position vector of the reorganized individual, w best It represents the weight value corresponding to the best individual in the group of individuals before recombination. represents the velocity vector of the individual before recombination, represents the velocity vector corresponding to the best individual in the group before recombination, Represents the position vector of the individual before recombination.

3. The micro-sized turbojet engine impeller optimization method based on BP neural network according to claim 2 is characterized in that: The gene mutation mode of an individual is: the individual has a probability p m The following gene mutations occur: In the above formula, Δx represents the interval A uniformly distributed random number, represents the fitness value of the individual to be mutated, f avg represents the average fitness value of the population of the individual to be mutated, P m1 =0.1, P m2 =0.001, f max Indicates the maximum fitness value in the group where the individual to be mutated belongs.

4. The micro-sized turbojet engine impeller optimization method based on BP neural network according to claim 1 is characterized in that: Step S4 also includes: iterating steps S3-S4, and when the end condition is met, outputting the current optimal weight and the current optimal; in step S4, setting the weight error setting value and the threshold error setting value, calculating the difference between the current optimal weight and the initial weight, and the difference between the current optimal threshold and the initial threshold, setting the difference between the current optimal weight and the initial weight as the weight error, and the difference between the current optimal threshold and the initial threshold as the threshold error, and the end condition is: the weight error is less than the weight error setting value, and the threshold error is less than the threshold error setting value.

5. The micro-sized turbojet engine impeller optimization method based on BP neural network according to claim 1 is characterized in that: The BP network model is a three-layer network structure, including input layer, hidden layer and output layer. The structure is expressed as: M1-M2+1; Among them, M1 represents the number of nodes in the input layer, that is, the number of variables to be optimized; M2 represents the number of nodes in the hidden layer; and the number of nodes in the output layer is the number of targets to be optimized.

6. The micro-sized turbojet engine impeller optimization method based on BP neural network according to claim 5 is characterized in that: The expression for the number of nodes in the hidden layer is: M2=2×M1+1.

7. The micro-sized turbojet engine impeller optimization method based on BP neural network according to claim 6 is characterized in that: The expressions of weight and threshold in BP network model are: M3=M1×M2+M2×2; M4=M2+2; In the above formula, M3 represents the weight and M4 represents the threshold.

8. The micro-sized turbojet engine impeller optimization method based on BP neural network according to claim 1 is characterized in that: S1 includes the following steps: S11, using the current design software, setting different optimization target values, and inputting each variable to be optimized into the current design software to perform performance prediction to achieve the optimization target; The variables to be optimized include the impeller's inlet hub radius, rim radius, outlet radius and height, blade tip clearance, blade thickness, inlet angle of attack and back bend angle; the optimization objectives include impeller efficiency and total pressure ratio; S12. In the performance prediction of each optimization target, at least the first 200 variables to be optimized with strong correlation with each optimization target are selected. Each variable to be optimized and its corresponding optimization target are taken as a sample pair, and all sample pairs constitute a sample library. S13. Divide the sample library into a training sample library and a test sample library.

Images