A noise prediction method for automotive axial cooling fans based on PSO-BP neural network

Abstract

This invention discloses a noise prediction method for automotive axial-flow cooling fans based on PSO-BP neural networks, belonging to the field of automotive noise prediction technology. The invention includes at least the following steps: Step 1: Building an automotive axial-flow cooling test fan and simulation model, verifying the rationality of the simulation model to confirm its ability to accurately simulate actual fan behavior, and then performing simulation analysis through the simulation model after rationality verification to obtain the fan noise level corresponding to different fan parameter configurations. This invention integrates particle swarm optimization (PSO) with BP neural networks, aiming to combine the advantages of both to improve prediction efficiency and accuracy. This invention not only retains the advantages of the BP algorithm in precise optimization but also leverages the global optimization capability of the PSO algorithm to accelerate the optimization process, avoid local optimum traps, and thus significantly improve the accuracy and efficiency of lifetime prediction.

Description

Technical Field

[0001] This invention relates to the field of automotive noise prediction technology, specifically a method for predicting the noise of automotive axial flow cooling fans based on PSO-BP neural networks. Background Technology

[0002] In recent years, with consumers' increasing demands for automotive ride comfort and vibration and noise control standards, and the enrichment of in-vehicle features such as the widespread adoption of air conditioning systems and exhaust purification devices, the temperature inside the engine compartment has risen significantly. This, in turn, increases the workload of the cooling fan, making noise problems more prominent. Therefore, more stringent and specific requirements have been placed on the design and performance of engine cooling fans, aiming to effectively reduce noise and ensure efficient heat dissipation.

[0003] Fans are mainly divided into two types: centrifugal and axial. Centrifugal fans are characterized by fluid entering the fan along the axial direction and then being discharged laterally under the centrifugal force generated by the rotating blades. In contrast, axial fans guide the fluid to flow continuously along their axial direction. Due to their smaller space requirements and lower noise levels, they are more commonly used in vehicles. In the noise composition of automotive axial cooling fans, aerodynamic noise dominates, significantly exceeding the magnitude of mechanical and electromagnetic noise, and thus becoming the primary source of noise in automotive axial cooling fans.

[0004] Existing backpropagation (BP) neural networks are known for their excellent nonlinear mapping capabilities, but they are often limited by the difficulty in precisely setting the number of neurons in the hidden layer. For example, too many neurons increase the computational burden and slow down convergence, while too few neurons affect the prediction accuracy and are prone to getting trapped in local optima and have slow convergence speed.

[0005] Aeroacoustics and fluid mechanics are closely intertwined. To gain a deeper understanding of aerodynamic noise fields, it is essential to accurately capture the flow field characteristics of the fluid. Traditionally, the study of fan flow field and sound field characteristics relies on both experimental measurements and numerical simulations. However, traditional experimental methods are limited by high costs, lengthy cycles, complex operations, and the non-universality of models, making large-scale industrial applications difficult. In contrast, numerical simulations significantly reduce computational costs and provide more detailed flow field information than experiments. However, even with numerical simulations, models with complex geometries and high Reynolds number flow characteristics still face enormous grid volumes and complex computational challenges, consuming substantial computational resources.

[0006] Therefore, a new solution is needed to address the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a noise prediction method for automotive axial flow cooling fans based on PSO-BP neural networks. Taking the noise of automotive axial flow cooling fans as the research object, after selecting appropriate neural network feature input and output quantities, the obtained feature data is substituted into the neural network for training. The trained neural network shows good prediction accuracy and generalization on the test set. This method greatly reduces the time for calculating axial flow cooling fan noise, thereby solving the technical problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the noise of automotive axial flow cooling fans based on PSO-BP neural networks, comprising at least the following steps:

[0009] Step 1: Build a test fan and simulation model for automotive axial flow cooling, and verify the rationality of the simulation model to confirm that it can accurately simulate the actual fan behavior. After the rationality verification, conduct simulation analysis through the simulation model to obtain the fan noise level corresponding to different fan parameter configurations.

[0010] Step 2: Collect parameter data and noise data of the automotive axial flow cooling test fan or simulation model. The parameter data includes rotational speed, blade tip clearance, blade installation angle, blade width, radius of curvature, number of blades, and hub ratio. Divide the parameter data into two subsets: training data and test data. The training data is used for model training and learning, and the test data is used to evaluate the performance and accuracy of the trained model.

[0011] Step 3: Based on the given set of neural network input and output samples, a topology of a BP neural network model was constructed, and the neural network was trained.

[0012] Step 4: Determine the Particle Swarm Optimization (PSO) algorithm. In the PSO algorithm, the dimension D of the position and velocity vector of each particle is determined by the number of parameters of the neural network. Specifically, D equals the number of connection weights from the input layer to the hidden layer, plus the number of connection weights from the hidden layer to the output layer, plus the number of thresholds for all neurons in the hidden layer, and finally plus the number of thresholds for all neurons in the output layer. These parameters together constitute the position and velocity vector of the particle in the search space, which is used to find the optimal neural network parameter configuration during the iteration process.

[0013] Step 5: Initialize the particle swarm optimization algorithm;

[0014] Step 6: Calculate the fitness of each particle based on the input and output samples;

[0015] Step 7: After calculating the fitness values ​​of all particles, by comparing the magnitudes of these fitness values, the individual extreme point of each particle can be determined. At the same time, by further comparing the individual extreme points of all particles, the global optimal extreme point can be determined.

[0016] Step 8: According to the particle position and velocity update formula defined in the particle swarm optimization algorithm, iteratively update the position and velocity of each particle. When the updated value exceeds the limit range, appropriate adjustments or constraints need to be made to ensure the effectiveness and rationality of the algorithm.

[0017] Step 9: Calculate the algorithm error;

[0018] Step 10: Evaluate whether the current error has reached the preset accuracy requirement, or whether the number of iterations has exceeded the set maximum value. If the error meets the preset accuracy standard, the algorithm is considered to have converged. At this time, the weights and thresholds of each dimension in the global optimal solution Pg obtained in the last iteration are the optimal parameter configuration for solving the problem. If the number of iterations has not reached the maximum value and the error still does not meet the accuracy requirement, the algorithm will return to step 6 and continue to execute the iteration process to find a better solution. If the number of iterations has reached the maximum value but the preset accuracy has not been reached, the algorithm will terminate execution and give an approximate solution to the problem based on the current global optimal solution.

[0019] Step 11: Optimize BP neural network parameters using particle swarm optimization algorithm: Use particle swarm optimization algorithm to optimize the initial weights and thresholds of BP neural network to improve the BP neural network's ability to predict sample data, thus achieving the application of predicting noise of automotive axial flow cooling fans.

[0020] Furthermore, the construction of the topology of the BP neural network model includes at least the following steps:

[0021] First, determine the initial structure of the BP neural network. When balancing model performance, training efficiency, and generalization ability, choosing a three-layer network structure is a reasonable starting point.

[0022] The three-layer network includes an input layer, a hidden layer, and an output layer;

[0023] The neurons between the input layer, hidden layer, and output layer are interconnected by weights and thresholds. These weights and thresholds are encoded as real number vectors, each of which represents an individual particle in the swarm optimization algorithm. In this way, the optimization algorithm is used to find the optimal combination of weights and thresholds to optimize the performance of the neural network.

[0024] The input parameters of the BP neural network input layer are the rotational speed, tip clearance, blade mounting angle, blade width, radius of curvature, number of blades, and hub ratio of the automotive axial flow cooling fan. The output parameter of the output layer is the noise of the automotive axial flow cooling fan. Therefore, the number of nodes m in the input layer is 7; the number of nodes n in the output layer is 1; and the number of nodes in the hidden layer is determined according to an empirical formula for calculating the number of hidden layer elements. (a is a constant from 1 to 10) Determine the range, select appropriate number of iterations, learning rate, etc., and repeatedly compare through trial and error to determine the optimal number of hidden layer neurons in the model;

[0025] The initial connection weights between the input layer neurons and the hidden layer neurons, and between the hidden layer neurons and the output layer neurons, are W, respectively. i j O j k The threshold biases of the hidden layer neurons and the output layer neurons are b, respectively. j , Where i = 1, 2, 3...m, j = 1, 2, 3...q, k = 1, 2, 3...n, and the activation function is f(x) = 1 / (1+e^(-j / q)). -x );

[0026] To facilitate calculation and comparison, the neuron data of the input and output layers need to be normalized. The normalization formula is usually as follows:

[0027]

[0028] In the formula: x * It is the normalized value, x i It is training data, x min x max These are the minimum and maximum values ​​in the data, respectively. This processing method can scale the data to a range of 0 to 1, which facilitates subsequent calculations and analysis.

[0029] Furthermore, the training of the neural network includes at least the following steps:

[0030] The first stage is the forward propagation stage. In this stage, the input data is passed forward layer by layer through the neural network. After the calculation and activation of each layer node, the output value of the network is finally obtained.

[0031] Next comes the backpropagation phase, in which the network adjusts the connection weights between layers and the bias terms of each neuron from back to front, based on the difference between the output value and the target value, in order to reduce this error.

[0032] This process repeats itself, and each combination of forward and backward propagation is considered to be a training iteration.

[0033] Through multiple iterations of training, until the network error is reduced to below the preset target threshold, the network weights and biases reach a stable state, marking the completion of model training and the successful construction of an optimized neural network model.

[0034] Furthermore, step 5 includes at least the following steps:

[0035] First, the size of the particle swarm, i.e. the total number of particles, is set. Then, each particle is randomly assigned an initial position and initial velocity, which represent a set of potential neural network parameter configurations.

[0036] At the same time, the inertia factor w of the algorithm is determined, which is used to adjust the degree of influence of the current velocity in the particle velocity update;

[0037] Set learning factors c1 and c2, which control the tendency of the particle to move towards its own historical best position and global best position, respectively;

[0038] Define the maximum number of iterations as the termination condition for the algorithm;

[0039] In addition, the individual extreme value of each particle is initialized to the fitness value corresponding to its initial position, and it is assumed that the global optimum is the best individual extreme value in the initial particle swarm.

[0040] As the algorithm iterates, these values ​​will be updated based on the particle's fitness.

[0041] Furthermore, step 6 includes at least the following steps:

[0042] First, for each particle in the particle swarm, perform the following operations in sequence;

[0043] For each sample in the sample set, the forward computation method of the BP neural network is used to obtain the output value of the neural network represented by the particle for that sample, and the error of that sample is calculated accordingly.

[0044] This process will iterate through all samples, accumulating the error of all samples for the current particle;

[0045] Then, the mean square error of these errors is calculated to obtain the fitness value of the particle;

[0046] After calculating the fitness of one particle, return to the previous step and continue performing the same operation on the next particle in the swarm until the fitness of all particles has been calculated.

[0047] Furthermore, the optimization of BP neural network parameters using the particle swarm optimization algorithm includes at least the following steps:

[0048] (3) Initialize the BP neural network structure:

[0049] The initialization of the BP neural network structure includes setting the number of neurons in the input layer, hidden layer, and output layer, the learning rate α, β, and the input and output of the training samples.

[0050] (2) Initialize the particle swarm:

[0051] When applying particle swarm optimization to optimize the initial weights and thresholds of a backpropagation neural network, it is necessary to set the column parameters in detail to control the operation of the algorithm.

[0052] The parameters include the particle size N, the position and velocity vectors of each particle, the individual extrema and global optimum of each particle, the iteration error precision ε, constant coefficients c1 and c2, and the maximum inertia weight w. max Minimum inertia weight w min Maximum speed v max and maximum number of iterations;

[0053] (4) Iteratively update position and velocity;

[0054] In each iteration, the new velocity of each particle is first calculated using the following velocity update formula;

[0055] Then, check whether the updated speed exceeds the preset maximum speed v. max ;

[0056] If the updated speed is greater than v max Then the velocity of the particle is adjusted to v. max To ensure it does not exceed the allowed range; if the updated speed is less than or equal to v max If so, then the speed remains unchanged;

[0057] Finally, the new position of each particle is calculated using the updated velocity and position update formulas;

[0058] The specific formula is as follows:

[0059] V ij (t+1)=w·v ij (t)+c1r1(t)[p ij (t)-x ij (t)]+c2r2(t)[P gj (t)-x ij (t)]

[0060] X ij (t+1)=xij (t)+v ij (t+1)

[0061] In the formula: i = 1, 2...N, where N is the number of particles;

[0062] j = 1, 2...D, where D is the dimension of the particle;

[0063] v ij (t), x ij (t) represents the velocity and position of particle i in dimension j at time t;

[0064] v ij (t+1), x ij (t+1) represents the velocity and position of particle i in dimension j at time (t+1);

[0065] p ij This is the optimal position where particle i has appeared so far;

[0066] p gj It is the optimal position where all particles have appeared so far;

[0067] c1 and c2 are learning factors, usually set to 2; r1 and r2 are random numbers in [0,1].

[0068] (4) Calculate the fitness value of each particle using the following fitness function:

[0069]

[0070] Where: n t O is the number of training samples; ir T ir These represent the network weights and thresholds determined at the position of the i-th particle in the training sample r, respectively, and the actual and expected network outputs.

[0071] (5) Calculate the global minimum fitness f of the particle swarm. g =min{f1, f2, ..., f N};

[0072] If the current iteration count reaches the preset maximum iteration count, or the global minimum value f g If the value is less than the given precision threshold ∈, the algorithm stops iterating;

[0073] Otherwise, the algorithm will continue, first updating the individual extreme value P of each particle. i and global extremum P g Then, based on these extreme values, the particle's velocity and position are updated in order to continue searching for better solutions in the solution space;

[0074] (6) When the algorithm reaches the stopping condition, the global extreme value P g The position represents a set of optimal network parameters, which include the weights and thresholds of the network. The output is the optimal network parameters determined by the global extremum Pg, which includes the weights and thresholds of all layers. The optimal network parameters and output parameters will be used as the configuration of the trained neural network.

[0075] (7) Verify the accuracy of the model to ensure that its predictive ability in practical applications meets the expected standard, and then output the predicted value.

[0076] Furthermore, the accuracy verification of the model to ensure that its predictive ability in practical applications meets the expected standard, and the output of predicted values, includes at least the following steps:

[0077] After optimization by the particle swarm optimization algorithm, the obtained optimal weights and thresholds are reconfigured for the BP neural network.

[0078] Subsequently, the neural network was tested using a set of independent test samples, and the performance of the model was evaluated by calculating the error between the output of the neural network and the corresponding real result.

[0079] If the calculated error meets the preset accuracy requirements or error threshold, it indicates that the model has been successfully established.

[0080] If the error requirement is not met, the training process needs to be repeated.

[0081] Ultimately, when the model meets the design requirements, the BP neural network is considered the optimal model after optimization by the particle swarm optimization algorithm, and its output is the predicted value when used for subsequent prediction of noisy data.

[0082] Compared with the prior art, the beneficial effects of the present invention are:

[0083] 1. This invention integrates particle swarm optimization (PSO) with backpropagation (BP) neural network (BNN) to combine the advantages of both and improve prediction efficiency and accuracy. This invention not only retains the advantages of the BP algorithm in precise optimization, but also accelerates the optimization process by leveraging the global optimization capability of the PSO algorithm, avoiding local optimum traps, thereby significantly improving the accuracy and efficiency of lifetime prediction.

[0084] 2. Compared with traditional experimental measurement methods and numerical simulation methods, this model can save a lot of manpower, material resources, and huge computing resources, and can be applied to different models, making it widely applicable in industry.

[0085] 3. This invention further optimizes the model in terms of sample size, variable factors, and structural algorithms, and can provide a more accurate noise prediction method for automotive axial flow cooling fans. Attached Figure Description

[0086] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0087] Figure 1 This is a flowchart of the PSO-BP algorithm of the present invention;

[0088] Figure 2 This is a diagram of the BP neural network structure of the present invention. Detailed Implementation

[0089] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0090] See Figure 1 - Figure 2 A method for predicting noise in automotive axial cooling fans based on PSO-BP neural networks, comprising at least the following steps:

[0091] Step 1: Build a test fan and simulation model for automotive axial flow cooling, and verify the rationality of the simulation model to confirm that it can accurately simulate the actual fan behavior. After the rationality verification, conduct simulation analysis through the simulation model to obtain the fan noise level corresponding to different fan parameter configurations.

[0092] Step 2: Collect parameter data and noise data of the automotive axial flow cooling test fan or simulation model. The parameter data includes rotational speed, blade tip clearance, blade installation angle, blade width, radius of curvature, number of blades, and hub ratio. Divide the parameter data into two subsets: training data and test data. The training data is used for model training and learning, and the test data is used to evaluate the performance and accuracy of the trained model.

[0093] Step 3: Based on the given set of neural network input and output samples, a topology of a BP neural network model was constructed, and the neural network was trained.

[0094] Step 4: Determine the Particle Swarm Optimization (PSO) algorithm. In the PSO algorithm, the dimension D of the position and velocity vector of each particle is determined by the number of parameters of the neural network. Specifically, D equals the number of connection weights from the input layer to the hidden layer, plus the number of connection weights from the hidden layer to the output layer, plus the number of thresholds for all neurons in the hidden layer, and finally plus the number of thresholds for all neurons in the output layer. These parameters together constitute the position and velocity vector of the particle in the search space, which is used to find the optimal neural network parameter configuration during the iteration process.

[0095] Step 5: Initialize the particle swarm optimization algorithm;

[0096] Step 6: Calculate the fitness of each particle based on the input and output samples;

[0097] Step 7: After calculating the fitness values ​​of all particles, by comparing the magnitudes of these fitness values, the individual extreme point of each particle can be determined. At the same time, by further comparing the individual extreme points of all particles, the global optimal extreme point can be determined.

[0098] Step 8: According to the particle position and velocity update formula defined in the particle swarm optimization algorithm, iteratively update the position and velocity of each particle. When the updated value exceeds the limit range, appropriate adjustments or constraints need to be made to ensure the effectiveness and rationality of the algorithm.

[0099] Step 9: Calculate the algorithm error;

[0100] Step 10: Evaluate whether the current error has reached the preset accuracy requirement, or whether the number of iterations has exceeded the set maximum value. If the error meets the preset accuracy standard, the algorithm is considered to have converged. At this time, the weights and thresholds of each dimension in the global optimal solution Pg obtained in the last iteration are the optimal parameter configuration for solving the problem. If the number of iterations has not reached the maximum value and the error still does not meet the accuracy requirement, the algorithm will return to step 6 and continue to execute the iteration process to find a better solution. If the number of iterations has reached the maximum value but the preset accuracy has not been reached, the algorithm will terminate execution and give an approximate solution to the problem based on the current global optimal solution.

[0101] Step 11: Optimize BP neural network parameters using particle swarm optimization algorithm: Use particle swarm optimization algorithm to optimize the initial weights and thresholds of BP neural network to improve the BP neural network's ability to predict sample data, thus achieving the application of predicting noise of automotive axial flow cooling fans.

[0102] Furthermore, the construction of the topology of the BP neural network model includes at least the following steps:

[0103] First, determine the initial structure of the BP neural network. When designing a BP neural network, a three-layer network structure with a single hidden layer is usually preferred because this design can maintain a certain level of computational accuracy while keeping the network structure relatively simple. Although increasing the number of hidden layers can help further reduce computational errors and improve model accuracy, it also significantly increases network complexity, thereby prolonging training time and increasing the risk of model overfitting. Therefore, when balancing model performance, training efficiency, and generalization ability, choosing a three-layer network structure as an initial step is a reasonable starting point.

[0104] The three-layer network includes an input layer, a hidden layer, and an output layer;

[0105] The neurons between the input layer, hidden layer, and output layer are interconnected by weights and thresholds. These weights and thresholds are encoded as real number vectors, each of which represents an individual particle in the swarm optimization algorithm. In this way, the optimization algorithm is used to find the optimal combination of weights and thresholds to optimize the performance of the neural network.

[0106] The input parameters of the BP neural network input layer are the rotational speed, tip clearance, blade mounting angle, blade width, radius of curvature, number of blades, and hub ratio of the automotive axial flow cooling fan. The output parameter of the output layer is the noise of the automotive axial flow cooling fan. Therefore, the number of nodes m in the input layer is 7; the number of nodes n in the output layer is 1; and the number of nodes in the hidden layer is determined according to an empirical formula for calculating the number of hidden layer elements. (a is a constant from 1 to 10) Determine the range, select appropriate number of iterations, learning rate, etc., and repeatedly compare through trial and error to determine the optimal number of hidden layer neurons in the model;

[0107] The initial connection weights between the input layer neurons and the hidden layer neurons, and between the hidden layer neurons and the output layer neurons, are W, respectively. ij O jk The threshold biases of the hidden layer neurons and the output layer neurons are b, respectively. j , Where i = 1, 2, 3...m, j = 1, 2, 3...q, k = 1, 2, 3...n, and the activation function is f(x) = 1 / (1+e^(-j / q)). -x );

[0108] To facilitate calculation and comparison, the neuron data of the input and output layers need to be normalized. The normalization formula is usually as follows:

[0109]

[0110] In the formula: x *It is the normalized value, x i It is training data, x min x max These are the minimum and maximum values ​​in the data, respectively. This processing method can scale the data to a range of 0 to 1, which facilitates subsequent calculations and analysis.

[0111] Furthermore, the training of the neural network includes at least the following steps:

[0112] The first stage is the forward propagation stage. In this stage, the input data is passed forward layer by layer through the neural network. After the calculation and activation of each layer node, the output value of the network is finally obtained.

[0113] Next comes the backpropagation phase, in which the network adjusts the connection weights between layers and the bias terms of each neuron from back to front, based on the difference between the output value and the target value (i.e., the error), in order to reduce this error.

[0114] This process repeats itself, and each combination of forward and backward propagation is considered to be a training iteration.

[0115] Through multiple iterations of training, until the network error is reduced to below the preset target threshold, the network weights and biases reach a stable state, marking the completion of model training and the successful construction of an optimized neural network model.

[0116] Furthermore, step 5 includes at least the following steps:

[0117] First, the size of the particle swarm, i.e. the total number of particles, is set. Then, each particle is randomly assigned an initial position and initial velocity, which represent a set of potential neural network parameter configurations.

[0118] At the same time, the inertia factor w of the algorithm is determined, which is used to adjust the degree of influence of the current velocity in the particle velocity update;

[0119] Set learning factors c1 and c2, which control the tendency of the particle to move towards its own historical best position and global best position, respectively;

[0120] Define the maximum number of iterations as the termination condition for the algorithm;

[0121] In addition, the individual extreme value of each particle is initialized to the fitness value corresponding to its initial position (i.e., the neural network performance evaluation index), and it is assumed that the global optimum is the best individual extreme value in the initial particle swarm.

[0122] As the algorithm iterates, these values ​​will be updated based on the particle's fitness.

[0123] Furthermore, step 6 includes at least the following steps:

[0124] First, for each particle in the particle swarm, perform the following operations in sequence;

[0125] For each sample in the sample set, the forward computation method of the BP neural network is used to obtain the output value of the neural network represented by the particle for that sample, and the error of that sample is calculated accordingly.

[0126] This process will iterate through all samples, accumulating the error of all samples for the current particle;

[0127] Then, the mean square error of these errors is calculated to obtain the fitness value of the particle;

[0128] After calculating the fitness of one particle, return to the previous step and continue performing the same operation on the next particle in the swarm until the fitness of all particles has been calculated.

[0129] Furthermore, the optimization of BP neural network parameters using the particle swarm optimization algorithm includes at least the following steps:

[0130] (5) Initialize the BP neural network structure:

[0131] The initialization of the BP neural network structure includes setting the number of neurons in the input layer, hidden layer, and output layer, the learning rate α, β, and the input and output of the training samples.

[0132] (2) Initialize the particle swarm:

[0133] When applying particle swarm optimization to optimize the initial weights and thresholds of a backpropagation neural network, it is necessary to set the column parameters in detail to control the operation of the algorithm.

[0134] The parameters include the particle size N, the position and velocity vectors of each particle, the individual extrema and global optimum of each particle, the iteration error precision ε, constant coefficients c1 and c2, and the maximum inertia weight w. max Minimum inertia weight w min Maximum speed v max and maximum number of iterations;

[0135] (6) Iteratively update position and velocity;

[0136] In each iteration, the new velocity of each particle is first calculated using the following velocity update formula;

[0137] The formula typically includes parameters such as current velocity, the difference between the individual's optimal position and the current position, the difference between the global optimal position and the current position, as well as inertia weight, individual learning factor, and social learning factor.

[0138] Then, check whether the updated speed exceeds the preset maximum speed v. max ;

[0139] If the updated speed is greater than v max Then the velocity of the particle is adjusted to v. max To ensure it does not exceed the allowed range; if the updated speed is less than or equal to v max If so, then the speed remains unchanged;

[0140] Finally, the new position of each particle is calculated using the updated velocity and position update formulas;

[0141] The specific formula is as follows:

[0142] V ij (t+1)=w·v ij (t)+c1r1(t)[p ij (t)-x ij (t)]+c2r2(t)[P gj (t)-x ij (t)]

[0143] X ij (t+1)=x ij (t)+v ij (t+1)

[0144] In the formula: i = 1, 2...N, where N is the number of particles;

[0145] j = 1, 2...D, where D is the dimension of the particle;

[0146] v ij (t), x ij (t) represents the velocity and position of particle i in dimension j at time t;

[0147] v ij (t+1), x ij (t+1) represents the velocity and position of particle i in dimension j at time (t+1);

[0148] p ij It is the best position (individual extreme value) that particle i has appeared so far;

[0149] p gj It is the optimal position (population extreme value) where all particles have appeared so far;

[0150] c1 and c2 are learning factors, usually set to 2; r1 and r2 are random numbers in [0,1].

[0151] (4) Calculate the fitness value of each particle using the following fitness function:

[0152]

[0153] Where: n t O is the number of training samples; ir T ir These represent the network weights and thresholds determined at the position of the i-th particle in the training sample r, respectively, and the actual and expected network outputs.

[0154] (5) Calculate the global minimum fitness f of the particle swarm. g =min{f1, f2, ..., f N};

[0155] If the current iteration count reaches the preset maximum iteration count, or the global minimum value f g If the error is less than the given accuracy threshold (which indicates that the network's training error has reached the required accuracy), the algorithm stops iterating.

[0156] Otherwise, the algorithm will continue, first updating the individual extreme value P of each particle. i (i.e., the optimal position experienced by each particle) and the global extremum P g (i.e., the best position experienced by the entire particle swarm), and then update the particle velocity and position based on these extreme values ​​in order to continue searching for better solutions in the solution space;

[0157] (6) When the algorithm reaches the stopping condition, the global extreme value P g The position represents a set of optimal network parameters, which include the weights and thresholds of the network. The output is the optimal network parameters determined by the global extremum Pg, which includes the weights and thresholds of all layers. The optimal network parameters and output parameters will be used as the configuration of the trained neural network.

[0158] (7) Verify the accuracy of the model to ensure that its predictive ability in practical applications meets the expected standard, and then output the predicted value.

[0159] Furthermore, the accuracy verification of the model to ensure that its predictive ability in practical applications meets the expected standard, and the output of predicted values, includes at least the following steps:

[0160] After optimization by the particle swarm optimization algorithm, the obtained optimal weights and thresholds are reconfigured for the BP neural network.

[0161] Subsequently, the neural network was tested using a set of independent test samples, and the performance of the model was evaluated by calculating the error between the output of the neural network and the corresponding real result.

[0162] If the calculated error meets the preset accuracy requirements or error threshold, it indicates that the model has been successfully established.

[0163] If the error requirement is not met, the training process needs to be repeated, including using particle swarm optimization to optimize the weights and thresholds of the neural network until a satisfactory prediction accuracy is achieved.

[0164] Ultimately, when the model meets the design requirements, the BP neural network is considered the optimal model after optimization by the particle swarm optimization algorithm, and its output is the predicted value when used for subsequent prediction of noisy data.

[0165] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A noise prediction method for automotive axial cooling fans based on PSO-BP neural networks, characterized in that: At least the following steps are included: Step 1: Build a test fan and simulation model for automotive axial flow cooling, and verify the rationality of the simulation model to confirm that it can accurately simulate the actual fan behavior. After the rationality verification, conduct simulation analysis through the simulation model to obtain the fan noise level corresponding to different fan parameter configurations. Step 2: Collect parameter data and noise data of the automotive axial flow cooling test fan or simulation model. The parameter data includes rotational speed, blade tip clearance, blade installation angle, blade width, radius of curvature, number of blades, and hub ratio. Divide the parameter data into two subsets: training data and test data. The training data is used for model training and learning, and the test data is used to evaluate the performance and accuracy of the trained model. Step 3: Based on the given set of neural network input and output samples, a topology of a BP neural network model was constructed, and the neural network was trained. Step 4: Determine the Particle Swarm Optimization (PSO) algorithm. In the PSO algorithm, the dimension D of the position and velocity vector of each particle is determined by the number of parameters of the neural network. Specifically, D equals the number of connection weights from the input layer to the hidden layer, plus the number of connection weights from the hidden layer to the output layer, plus the number of thresholds for all neurons in the hidden layer, and finally plus the number of thresholds for all neurons in the output layer. These parameters together constitute the position and velocity vector of the particle in the search space, which is used to find the optimal neural network parameter configuration during the iteration process. Step 5: Initialize the particle swarm optimization algorithm; Step 6: Calculate the fitness of each particle based on the input and output samples; Step 7: After calculating the fitness values ​​of all particles, by comparing the magnitudes of these fitness values, the individual extreme point of each particle can be determined. At the same time, by further comparing the individual extreme points of all particles, the global optimal extreme point can be determined. Step 8: According to the particle position and velocity update formula defined in the particle swarm optimization algorithm, iteratively update the position and velocity of each particle. When the updated value exceeds the limit range, appropriate adjustments or constraints need to be made to ensure the effectiveness and rationality of the algorithm. Step 9: Calculate the algorithm error; Step 10: Evaluate whether the current error has reached the preset accuracy requirement, or whether the number of iterations has exceeded the set maximum value. If the error meets the preset accuracy standard, the algorithm is considered to have converged. At this time, the weights and thresholds of each dimension in the global optimal solution Pg obtained in the last iteration are the optimal parameter configuration for solving the problem. If the number of iterations has not reached the maximum value and the error still does not meet the accuracy requirement, the algorithm will return to step 6 and continue to execute the iteration process to find a better solution. If the number of iterations has reached the maximum value but the preset accuracy has not been reached, the algorithm will terminate execution and give an approximate solution to the problem based on the current global optimal solution. Step 11: Optimize BP neural network parameters using particle swarm optimization algorithm: Use particle swarm optimization algorithm to optimize the initial weights and thresholds of BP neural network to improve the BP neural network's ability to predict sample data, thus achieving the application of predicting noise of automotive axial flow cooling fans.

2. The method for predicting noise of automotive axial cooling fans based on PSO-BP neural networks according to claim 1, characterized in that: The construction of the topology of a BP neural network model includes at least the following steps: First, determine the initial structure of the BP neural network. When balancing model performance, training efficiency, and generalization ability, choosing a three-layer network structure is a reasonable starting point. The three-layer network includes an input layer, a hidden layer, and an output layer; The neurons between the input layer, hidden layer, and output layer are interconnected by weights and thresholds. These weights and thresholds are encoded as real number vectors, each of which represents an individual particle in the swarm optimization algorithm. In this way, the optimization algorithm is used to find the optimal combination of weights and thresholds to optimize the performance of the neural network. The input parameters of the BP neural network input layer are the rotational speed, tip clearance, blade mounting angle, blade width, radius of curvature, number of blades, and hub ratio of the automotive axial flow cooling fan. The output parameter of the output layer is the noise of the automotive axial flow cooling fan. Therefore, the number of nodes m in the input layer is 7; the number of nodes n in the output layer is 1; and the number of nodes in the hidden layer is determined according to an empirical formula for calculating the number of hidden layer elements. , Using a constant from 1 to 10, select an appropriate number of iterations and learning rate, and repeatedly compare and determine the optimal number of neurons in the hidden layer of the model through trial and error. The initial connection weights between the input layer neurons and the hidden layer neurons, and between the hidden layer neurons and the output layer neurons, are W, respectively. ij O j k The threshold biases of the hidden layer neurons and the output layer neurons are b, respectively. j , Where i = 1, 2, 3...m, j = 1, 2, 3...q, k = 1, 2, 3...n, and the activation function is f(x) = 1 / (1+e^(-j / q)). -x ); To facilitate calculation and comparison, the neuron data of the input and output layers need to be normalized. The normalization formula is usually as follows: ； In the formula: These are normalized values. It is training data. These are the minimum and maximum values ​​in the data, respectively. This processing method can scale the data to a range of 0 to 1, which facilitates subsequent calculations and analysis.

3. The noise prediction method for automotive axial flow cooling fans based on PSO-BP neural networks according to claim 1, characterized in that: The training of the neural network includes at least the following steps: The first stage is the forward propagation stage. In this stage, the input data is passed forward layer by layer through the neural network. After the calculation and activation of each layer node, the output value of the network is finally obtained. Next comes the backpropagation phase, in which the network adjusts the connection weights between layers and the bias terms of each neuron from back to front, based on the difference between the output value and the target value, in order to reduce this error. This process repeats itself, and each combination of forward and backward propagation is considered to be a training iteration. Through multiple iterations of training, until the network error is reduced to below the preset target threshold, the network weights and biases reach a stable state, marking the completion of model training and the successful construction of an optimized neural network model.

4. The method for predicting noise of automotive axial flow cooling fans based on PSO-BP neural networks according to claim 1, characterized in that: Step 5 includes at least the following steps: First, the size of the particle swarm, i.e. the total number of particles, is set. Then, each particle is randomly assigned an initial position and initial velocity, which represent a set of potential neural network parameter configurations. At the same time, the inertial weight w of the algorithm is determined, which is used to adjust the degree of influence of the current velocity in the particle velocity update; Set learning factors c1 and c2, which control the tendency of the particle to move towards its own historical best position and global best position, respectively; Define the maximum number of iterations as the termination condition for the algorithm; In addition, the individual extreme value of each particle is initialized to the fitness value corresponding to its initial position, and it is assumed that the global optimum is the best individual extreme value in the initial particle swarm. As the algorithm iterates, these values ​​will be updated based on the particle's fitness.

5. The method for predicting noise of automotive axial cooling fans based on PSO-BP neural networks according to claim 1, characterized in that: Step 6 includes at least the following steps: First, for each particle in the particle swarm, perform the following operations in sequence; For each sample in the sample set, the forward computation method of the BP neural network is used to obtain the output value of the neural network represented by the particle for that sample, and the error of that sample is calculated accordingly. This process will iterate through all samples, accumulating the error of all samples for the current particle; Then, the mean square error of these errors is calculated to obtain the fitness value of the particle; After calculating the fitness of one particle, return to the previous step and continue performing the same operation on the next particle in the swarm until the fitness of all particles has been calculated.

6. The noise prediction method for automotive axial flow cooling fans based on PSO-BP neural networks according to claim 1, characterized in that: The optimization of BP neural network parameters using the particle swarm optimization algorithm includes at least the following steps: (1) Initialize the BP neural network structure: The initialization of the BP neural network structure includes setting the number of neurons in the input layer, hidden layer, and output layer of the network, the learning rate α, β, and the input and output of the training samples; (2) Initialize the particle swarm: When applying particle swarm optimization to optimize the initial weights and thresholds of a backpropagation neural network, it is necessary to set the column parameters in detail to control the operation of the algorithm. The parameters include the particle size N, the position and velocity vectors of each particle, the individual extrema and global optimum of each particle, the iteration error precision ε, constant coefficients c1 and c2, and the maximum inertia weight w. max Minimum inertia weight w min Maximum speed v max and maximum number of iterations; (2) Iteratively update position and velocity; In each iteration, the new velocity of each particle is first calculated using the following velocity update formula; Then, check whether the updated speed exceeds the preset maximum speed v. max ; If the updated speed is greater than v max Then the particle's velocity is adjusted to v max To ensure it does not exceed the allowed range; if the updated speed is less than or equal to v max If so, then the speed remains unchanged; Finally, the new position of each particle is calculated using the updated velocity and position update formulas; The specific formula is as follows: V i j (t+1)=w·v i j (t)+c1 r1 (t)[p i j (t)-x i j (t)]+c2 r2(t)[P g j (t)-x i j (t)] X i j (t+1)=x i j (t)+v i j (t+1) In the formula: i = 1, 2...N, where N is the number of particles; j=1,2……D, where D is the dimension of the particle; t is the number of iterations; v i j (t), x i j (t) represents the velocity and position of particle i in dimension j at time t; v i j (t+1), x i j (t+1) represents the velocity and position of particle i in dimension j at time (t+1); p i j (t) represents the optimal position of the individual in the t-th iteration; P g j (t) represents the globally optimal position at the t-th iteration; c1 and c2 are learning factors, which are generally set to 2; r1(t) and r2(t) are random numbers in the range [0,1]. w is the inertia weight; (4) Calculate the fitness value of each particle using the following fitness function: ； Where: n t The number of training samples; , These represent the network weights and thresholds determined at the position of the i-th particle in the training sample r, respectively, and the actual and expected network outputs. (5) Calculate the global minimum fitness f of the particle swarm. g =min{f1, f2, ..., f N }; If the current iteration count reaches the preset maximum iteration count, or the global minimum value f g Less than a given precision threshold If the algorithm stops iterating, then the algorithm stops iterating. Otherwise, the algorithm will continue, first updating the individual extreme value P of each particle. i and global extremum P g Then, based on these extreme values, the particle's velocity and position are updated in order to continue searching for better solutions in the solution space; (6) When the algorithm reaches the stopping condition, the global extreme value P g The position represents a set of optimal network parameters, which include the weights and thresholds of the network. The output is the optimal network parameters determined by the global extremum Pg. The output parameters include the weights and thresholds of all layers. The optimal network parameters and output parameters will be used as the configuration of the trained neural network. (7) Verify the accuracy of the model to ensure that its predictive ability in practical applications meets the expected standard, and then output the predicted value.

7. The method for predicting noise of automotive axial flow cooling fans based on PSO-BP neural networks according to claim 6, characterized in that: The process of verifying the accuracy of the model to ensure that its predictive ability in practical applications meets the expected standards, and then outputting predicted values, includes at least the following steps: After optimization by the particle swarm optimization algorithm, the obtained optimal weights and thresholds are reconfigured for the BP neural network. Subsequently, the neural network was tested using a set of independent test samples, and the performance of the model was evaluated by calculating the error between the output of the neural network and the corresponding real result. If the calculated error meets the preset accuracy requirements or error threshold, it indicates that the model has been successfully established. If the error requirement is not met, the training process needs to be repeated. Ultimately, when the model meets the design requirements, the BP neural network is considered the optimal model after optimization by the particle swarm optimization algorithm, and its output is the predicted value when used for subsequent prediction of noisy data.

Images