Self-ventilation motor fan aerodynamic noise simulation evaluation method

Abstract

The invention relates to the technical field of noise evaluation, in particular to a self-ventilation motor fan aerodynamic noise simulation evaluation method. According to the technical scheme, the method comprises the steps of data acquisition and processing, deep learning model construction, simulation and optimization of a multi-physics coupling model and result fusion. According to the method, the hybrid deep learning model combining the convolutional neural network and the recurrent neural network is adopted, so that the noise prediction precision is remarkably improved, and the evaluation result is closer to the actual noise condition; a multi-physics coupling model is adopted to reflect the comprehensive influence of multiple factors on aerodynamic noise under actual working conditions, a noise generation mechanism is analyzed through a CFD model and an acoustic FEM model, a reliable physical basis is provided for noise evaluation, and errors are reduced; besides, the Bayesian fusion method is combined with the results of the deep learning model and the multi-physics field coupling model, the advantages of the two methods are comprehensively considered, the limitation of a single method is made up, and the reliability and stability of the evaluation result are further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of noise assessment, and particularly to a method for simulating and evaluating the aerodynamic noise of a self-ventilated motor fan. Background Art

[0002] Self-ventilated motors are widely used in many fields such as industrial production, transportation, and household appliances. The aerodynamic noise generated by the fan during its operation not only causes noise pollution to the working environment and affects the physical and mental health of operators, but may also reflect problems in the motor performance. Therefore, it is of great significance to accurately evaluate the aerodynamic noise of a self-ventilated motor fan.

[0003] There are some problems in the existing methods for simulating and evaluating the aerodynamic noise of a self-ventilated motor fan. On the one hand, the traditional methods for simulating and evaluating the aerodynamic noise of a self-ventilated motor fan have insufficient ability to simulate complex flow field structures. The airflow during the operation of a self-ventilated motor fan is complex and variable, including phenomena such as turbulence and vortices. It is difficult for traditional methods to accurately depict these details, resulting in inaccurate identification and quantification of noise sources, and thus affecting the accuracy of noise assessment. On the other hand, there is a lack of comprehensive consideration of multi-factor coupling. In actual work, aerodynamic noise is related to various factors such as the rotational speed of the motor, load, shape and material of the fan blades, etc. Traditional methods cannot truly reflect the noise characteristics under the interaction of various factors. Therefore, we propose a method for simulating and evaluating the aerodynamic noise of a self-ventilated motor fan. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for simulating and evaluating the aerodynamic noise of a self-ventilated motor fan in view of the problems existing in the background art.

[0005] The technical solution of the present invention: A method for simulating and evaluating the aerodynamic noise of a self-ventilated motor fan, comprising the following steps:

[0006] Collect various data during the operation of a self-ventilated motor fan under different working conditions by using sensors, and preprocess the collected data;

[0007] Construct a deep learning model, the deep learning model adopts a hybrid model combining a convolutional neural network and a recurrent neural network, and train and optimize the hybrid model;

[0008] Simulate the flow field and the propagation and radiation characteristics of aerodynamic noise of a self-ventilated motor fan through a multi-physics field coupling model, and introduce a reinforcement learning algorithm to optimize the multi-physics field coupling model;

[0009] Adopt a Bayesian fusion method to fuse the results simulated by the deep learning model and the multi-physics field coupling model, and use the fused results to perform feedback adjustment on the model.

[0010] Optionally, the sensor includes an acoustic sensor, a pressure sensor, and a speed sensor. The collected data includes the noise signal, flow field parameters, and motor electrical parameters during the operation of the self-ventilating motor fan. The data preprocessing includes the following steps:

[0011] First, use the fast Fourier transform to convert the collected time-domain noise signal into a frequency-domain signal. The formula is as follows:

[0012]

[0013] where x(n) is the time-domain signal, X(k) is the frequency-domain signal, N is the number of sampling points, k = 0, 1, …, N - 1. Then, use a band-pass filter to remove the interference components outside the specific frequency range and retain the frequency components related to the aerodynamic noise;

[0014] For different types of data, use the min-max normalization method to unify their numerical ranges to the interval [0, 1]. For data x, the normalization formula is as follows:

[0015]

[0016] where x min and x max are the minimum and maximum values of the data, respectively.

[0017] Optionally, in the construction of the deep learning model, the convolutional neural network part consists of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer extracts the spatial features in the flow field image data through a convolutional kernel. The convolution operation formula is as follows:

[0018]

[0019] where x is the input feature map, y is the output feature map, w is the convolutional kernel, b is the bias, M and N are the sizes of the convolutional kernel. The pooling layer is used to reduce the data dimension. The max-pooling operation formula is as follows:

[0020]

[0021] where S is the pooling window and s is the pooling stride;

[0022] The recurrent neural network part is used to process data with sequential features and introduces the long short-term memory network algorithm to process long-sequence data. The long short-term memory network algorithm processes time-series data through the forget gate f t 、input gate i t 、output gate o t and cell state C t ;

[0023] For each time step t, the hidden layer of the recurrent neural network receives the input x at the current moment t and the output h of the hidden layer at the previous moment t-1 , and then calculates the output h of the hidden layer at the current moment t , and its calculation formula is as follows:

[0024] h t = σ(W xh x t + W hh h t-1 + b h )

[0025] where σ is the activation function, W xh , W hh are weight matrices, and b h is the bias vector

[0026] Optionally, the fusion of the convolutional neural network and the recurrent neural network adopts a feature interaction fusion method, specifically including:

[0027] Input the flow field image data into the convolutional neural network for partial feature extraction to obtain the intermediate feature map C 1 (i, j);

[0028] Convert the feature map C 1 (i, j) into a one-dimensional vector and use it as the initial input state of the recurrent neural network or add it as an additional input to the input sequence of the recurrent neural network. Assume that at time step t = 1, the input of the recurrent neural network is Then the hidden state at this time is updated as:

[0029]

[0030] As the time series progresses, in each time step of the recurrent neural network, the current hidden state h t is fed back to the convolutional neural network, allowing the convolutional neural network to update its feature extraction according to the current time information. Among them, h t is added as an additional input channel to the flow field image to form a new input I'(x, y) = (I(x, y), h t ), and then a convolution operation is performed on it:

[0031]

[0032] where K'(m, n) and b' are the new convolution kernel and bias term;

[0033] Multi-stage Feature Fusion: In multiple processing stages, the calculations of the convolutional neural network and the recurrent neural network are alternately performed. Specifically, in the first stage, the convolutional neural network is used to extract the initial spatial features of the flow field, and the results are input into the recurrent neural network. In subsequent stages, a part of the output of the recurrent neural network is used as feedback to adjust the input or convolutional kernel weights of the convolutional neural network;

[0034] Assume that in the nth stage, according to the output of the recurrent neural network Adjust the convolutional kernel weight K n (m,n) of the convolutional neural network, and the adjustment formula is as follows:

[0035]

[0036] where α is the learning rate, and g(·) is an adjustment function based on the output of the recurrent neural network.

[0037] Optionally, the establishment of the multi-physics field coupling model includes a CFD model and an acoustic FEM model. The CFD model is a flow field model of a self-ventilated motor fan established based on computational fluid dynamics. By solving the fluid dynamics equations, the flow field around the self-ventilated motor fan is simulated. The fluid dynamics equations include three-dimensional incompressible Reynolds-averaged Navier-Stokes equations, and the formula is as follows:

[0038]

[0039] where ρ is the fluid density, is the average velocity component, is the average pressure, μ is the dynamic viscosity, u' i and u' j are velocity fluctuation components, and the standard k-ε turbulence model is used to close the Reynolds stress term. Among them, k represents the turbulent kinetic energy, ε represents the turbulent kinetic energy dissipation rate, and the standard k-ε turbulence model describes the characteristics of turbulence through the following equations:

[0040] Turbulent kinetic energy equation:

[0041]

[0042] where P k is the generation term of the turbulent kinetic energy, μ t is the turbulent viscosity, σ k is the turbulent Prandtl number of the k equation;

[0043] Turbulent kinetic energy dissipation rate equation:

[0044]

[0045] where, and are empirical constants, and σ ∈ is the turbulent Prandtl number of the ∈ equation.

[0046] Optionally, the acoustic FEM model is used to convert the physical information of the flow field into acoustic information. Based on the wave equation, for the propagation of sound waves in a homogeneous medium, the wave equation is:

[0047]

[0048] where p is the sound pressure, c is the speed of sound, Q is the sound source term. In the application of the self-ventilated motor fan, the sound source term Q is obtained from the flow field calculated by the CFD model;

[0049] The computational domain is divided into multiple finite element cells. For the sound pressure p within each cell, it is represented using the interpolation function N, i.e., p = ∑ i N i P i where P i is the nodal sound pressure and N i is the shape function. Integrating the wave equation within each cell and applying the variational principle, the following discrete form of the system of equations is obtained:

[0050]

[0051] where [M] is the mass matrix, [C] is the damping matrix, [K] is the stiffness matrix, {p} is the nodal sound pressure vector, {F} is the equivalent nodal force vector. For the mass matrix [M], its element M ij is expressed as:

[0052] M ij = ∫ Ω ρN i N j dΩ

[0053] For the stiffness matrix [K], its element K ij is expressed as:

[0054]

[0055] For the damping matrix [C], it is calculated according to a specific damping model, including Rayleigh damping [C] = α[M] + β[K], where α and β are damping coefficients.

[0056] Optionally, the information of the CFD model and the acoustic FEM model is transferred and integrated through coupling. The coupling method includes the transfer of flow field information. The flow field result calculated by the CFD model is used as the input of the acoustic FEM model, especially as the input of the sound source term Q. According to the Lighthill acoustic analogy, the sound source term Q is expressed as:

[0057]

[0058] wherein, T ij = ρu i u j + pδ ij - τ ij , τ ij is the viscous stress tensor, and δ ij is the Kronecker symbol.

[0059] Optionally, the reinforcement learning algorithm in the optimization of the multi-physics coupling model by introducing the reinforcement learning algorithm is the deep Q-network algorithm, including constructing a deep neural network for approximating the Q function, where the Q function represents the expected cumulative reward that can be obtained after taking action a in state s. In each step, the agent selects action a t according to the current state s t , and after executing the action, a new state s t+1 and reward r t are obtained. The parameters of the Q network are updated through the following formula:

[0060]

[0061] where α is the learning rate and γ is the discount factor. The agent automatically adjusts the parameters in the model by continuously interacting with the multi-physics coupling model according to different working conditions and objectives.

[0062] Optionally, the fusion of the results simulated by the deep learning model and the multi-physics coupling model includes assuming that the prediction result of the deep learning model is y DL , and the simulation result of the multi-physics coupling is y CFD-FEM . The fused result y fused is:

[0063] y fused = w DL y DL + w CFD-FEM y CFD-FEM

[0064] where w DL and w CFD-FEM are weights, and w DL + w CFD-FEM = 1. The calculation of the weights is based on Bayes' formula:

[0065]

[0066]

[0067] Among them, P(y|D) is the probability of the prediction result y under the data D, and P(D) is the prior probability of the data.

[0068] Optionally, the model is feedback-adjusted using the result after fusing the deep learning model and the multi-physics field coupling model. According to the error between the fusion result and the actual measurement noise value, the parameters of the deep learning model are adjusted using the backpropagation algorithm. At the same time, according to the differences between the multi-physics field coupling simulation result and the actual flow field and noise characteristics, the parameters and structure of the CFD model and the acoustic FEM model are adjusted.

[0069] In summary, the present application includes at least one of the following beneficial technical effects:

[0070] 1. The present invention adopts a hybrid deep learning model combining a convolutional neural network and a recurrent neural network, which can comprehensively capture the features related to the aerodynamic noise of a self-ventilated motor fan from both spatial and temporal dimensions. Among them, the convolutional neural network is used to extract the spatial features of the flow field image, and the recurrent neural network is used to mine the laws of the noise time series. Compared with traditional single models, the noise prediction accuracy is significantly improved, making the evaluation result closer to the actual noise situation;

[0071] 2. The present invention adopts a multi-physics field coupling model, and the data covers noise signals, flow field parameters, and motor electrical parameters, changing the current situation of considering some factors in isolation in traditional methods and reflecting the comprehensive influence of multiple factors on aerodynamic noise under actual working conditions. At the same time, in the multi-physics field coupling simulation, the noise generation mechanism is analyzed from the physical principle level through the CFD model and the acoustic FEM model, providing a more reliable physical basis for noise evaluation and reducing the errors caused by model simplification and assumptions in traditional CFD and acoustic simulation methods;

[0072] 3. The present invention uses the Bayesian fusion method to combine the results of the deep learning model and the multi-physics field coupling model, comprehensively considering the advantages of the two methods, making up for the limitations of single methods, further improving the reliability and stability of the evaluation result, and making the noise evaluation more accurate. Description of the Drawings

[0073] Figure 1 A flowchart of a method for simulating and evaluating the aerodynamic noise of a self-ventilated motor fan according to the present invention is given. Detailed Embodiments

[0074] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0075] Embodiment

[0076] A method for simulating and evaluating the aerodynamic noise of a self-ventilated motor fan proposed by the present invention is as follows Figure 1 shown, and includes the following steps:

[0077] I. Data collection and processing

[0078] Use sensors to collect various data during the operation of the self-ventilated motor fan under different working conditions;

[0079] Among them, the sensors are preferably B&K4190 type acoustic sensors, Omega PX409 type pressure sensors, and TIF-880 type wind speed sensors. 8 groups of acoustic sensors are evenly arranged around the fan to obtain noise data at different positions. The pressure sensors are used to measure the pressure distribution at the inlet, outlet and key parts of the fan, and a total of 10 pressure sensors are arranged. The wind speed sensors are used to measure the air flow velocity around the fan, and 12 wind speed sensors are arranged. At the same time, record the electrical parameters of the motor, including voltage and current. At each working condition, collect data for a duration of 10 minutes to ensure the sufficiency of the data.

[0080] Preprocess the collected data. The data preprocessing includes the following steps:

[0081] First, use the fast Fourier transform to convert the collected time-domain noise signal into a frequency-domain signal. The formula is as follows:

[0082]

[0083] Among them, x(n) is the time-domain signal, X(k) is the frequency-domain signal, N is the number of sampling points, k = 0, 1,..., N - 1. Then use a band-pass filter to remove the frequency components below 20 Hz and above 20 kHz, and retain the frequency components related to the aerodynamic noise;

[0084] For different types of data, use the minimum-maximum normalization method to unify its numerical range to the [0, 1] interval. For data x, the normalization formula is as follows:

[0085]

[0086] where x min and x max are the minimum and maximum values of the data respectively.

[0087] In the data collection stage, multi-source data of the self-ventilated motor fan under different motor speeds, loads, environmental temperatures and humidities are comprehensively collected, covering noise signals, flow field parameters and motor electrical parameters, changing the status quo of the traditional method considering only some factors in isolation, and reflecting the comprehensive influence of multiple factors on aerodynamic noise under actual working conditions.

[0088] II. Constructing a Deep Learning Model

[0089] The deep learning model is a hybrid model combining a convolutional neural network and a recurrent neural network;

[0090] Among them, the convolutional neural network part consists of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer extracts spatial features from the flow field image data through a convolutional kernel. The convolution operation formula is as follows:

[0091]

[0092] Among them, x is the input feature map, y is the output feature map, w is the convolutional kernel, b is the bias, M and N are the sizes of the convolutional kernel. The pooling layer is used to reduce the data dimension. The max-pooling operation formula is as follows:

[0093]

[0094] Among them, S is the pooling window and s is the pooling stride;

[0095] In addition, the recurrent neural network part is used to process data with sequential features. Since the traditional recurrent neural network has problems such as vanishing gradients or exploding gradients and is difficult to process long-sequence data, the long short-term memory network algorithm is introduced to process long-sequence data. The long short-term memory network algorithm processes time series data through a forget gate f t , an input gate i t , an output gate o t and a cell state C t . The forget gate f t determines how much information in the previous cell state can be retained to the current moment. The input gate i t determines how much information in the current input can be added to the cell state. The output gate o t then determines how much information in the cell state C t will be output to the hidden layer h t , enabling the long short-term memory network algorithm to better capture long-distance dependencies in the noise signal.

[0096] For each time step t, the hidden layer of the recurrent neural network receives the input x t at the current moment and the output h t-1 of the hidden layer at the previous moment, and then calculates the output h t of the hidden layer at the current moment. Its calculation formula is: h t =σ(W xh x t +W hh h t-1 +b h), where σ is the activation function, and W xh and W hh are weight matrices, and b h is the bias vector.

[0097] The fusion of the convolutional neural network and the recurrent neural network adopts the feature interaction fusion method, specifically including inputting the flow field image data into the convolutional neural network for partial feature extraction to obtain the intermediate feature map C 1 (i, j);

[0098] Convert the feature map C 1 (i, j) into a one-dimensional vector and use it as the initial input state of the recurrent neural network or add it as an additional input to the input sequence of the recurrent neural network. Assume that at time step t = 1, the input of the recurrent neural network is Then the hidden state update at this time is:

[0099]

[0100] As the time series progresses, at each time step of the recurrent neural network, the current hidden state h t is fed back to the convolutional neural network to allow the convolutional neural network to update its feature extraction according to the current time information. Among them, h t is added as an additional input channel to the flow field image to form the new input I'(x, y) = (I(x, y), h t ), and then perform a convolution operation on it:

[0101]

[0102] where K'(m, n) and b' are the new convolution kernel and bias term.

[0103] Multi-stage feature fusion: In multiple processing stages, the calculations of the convolutional neural network and the recurrent neural network are alternately performed. Among them, in the first stage, the convolutional neural network is used to extract the initial spatial features of the flow field and input the result into the recurrent neural network. In subsequent stages, part of the output of the recurrent neural network is used as feedback to adjust the input or convolution kernel weights of the convolutional neural network;

[0104] Assume that in the nth stage, according to the output of the recurrent neural network, adjust the convolution kernel weight K n (m, n) of the convolutional neural network, and the adjustment formula is as follows:

[0105]

[0106] Among them, α is the learning rate, and g(·) is an adjustment function based on the output of the recurrent neural network.

[0107] Through the combination of a convolutional neural network and a recurrent neural network, the hybrid deep learning model can comprehensively learn and analyze the information related to the aerodynamic noise of the self-ventilated motor fan from both spatial and temporal dimensions. Compared with a single model, it can capture the characteristics and laws of the noise more accurately, providing more powerful support for the simulation and evaluation of aerodynamic noise.

[0108] The deep learning model is trained. The preprocessed data is divided into a training set, a validation set, and a test set according to a ratio of 7:2:1. The stochastic gradient descent algorithm is used as the optimizer, the learning rate is set to 0.001, and the mean squared error is used as the loss function. The formula is where y i is the actually measured noise value, is the noise value predicted by the model, and N is the number of samples;

[0109] During the training process, the performance of the model is evaluated on the validation set every 10 epochs. If the loss on the validation set does not decrease for 5 consecutive epochs, the training is stopped, and finally the prediction accuracy and error of the model are obtained on the test set.

[0110] III. Simulation and Optimization of the Multiphysics Coupling Model

[0111] The establishment of the multiphysics coupling model includes a CFD model and an acoustic FEM model;

[0112] Among them, the CFD model is a flow field model of the self-ventilated motor fan established based on computational fluid dynamics. By solving the fluid dynamics equations, the flow field around the self-ventilated motor fan is simulated, including the distribution of airflow velocity and pressure parameters, providing a flow field basis for subsequent aerodynamic noise analysis. The fluid dynamics equations include the three-dimensional incompressible Reynolds-averaged Navier-Stokes equations, which describe the motion laws of viscous fluids. For Newtonian fluids, its continuity equation and momentum equation are as follows:

[0113]

[0114] Among them, ρ is the fluid density, is the average velocity component, is the average pressure, μ is the dynamic viscosity, u' i and u' j are the velocity fluctuation components.

[0115] The standard k-ε turbulence model is used to close the Reynolds stress term. Among them, k represents the turbulent kinetic energy, ε represents the turbulent kinetic energy dissipation rate, and the standard k-ε turbulence model describes the characteristics of turbulence through the following equations:

[0116] Turbulent kinetic energy equation: where P k is the generation term of turbulent kinetic energy, μ t is the turbulent viscosity, and σ k is the turbulent Prandtl number of the k - equation;

[0117] Turbulent kinetic energy dissipation rate equation: where and are empirical constants, and σ ∈ is the turbulent Prandtl number of the ∈ - equation.

[0118] In addition, the acoustic FEM model is used to convert the physical information of the flow field into acoustic information. Based on the wave equation, for the propagation of sound waves in a homogeneous medium, its wave equation is where p is the sound pressure, c is the speed of sound, Q is the sound source term. In the application of a self - ventilated motor fan, the sound source term Q is obtained from the flow field calculated by the CFD model.

[0119] Acoustic finite - element discretization: The computational domain is divided into multiple finite - element cells. For the sound pressure p within each cell, it is represented using the interpolation function N, i.e., p = ∑ i N i P i where P i is the nodal sound pressure and N i is the shape function. Integrating the wave equation within each cell and applying the variational principle, the following discrete - form system of equations is obtained:

[0120]

[0121] where [M] is the mass matrix, [C] is the damping matrix, [K] is the stiffness matrix, {p} is the nodal sound - pressure vector, and {F} is the equivalent nodal force vector. For the mass matrix [M], its element M ij is expressed as:

[0122] M ij = ∫ Ω ρN i N j dΩ

[0123] For the stiffness matrix [K], its element K ij is expressed as:

[0124]

[0125] For the damping matrix [C], it is calculated according to a specific damping model, including Rayleigh damping [C] = α[M]+β[K], where α and β are damping coefficients.

[0126] In multi-physics coupling simulation, through the CFD model and the acoustic FEM model, the noise generation mechanism is deeply analyzed from the physical principle level, providing a more reliable physical basis for noise assessment and reducing the errors caused by model simplification and assumptions in traditional CFD and acoustic simulation methods.

[0127] Through coupling, the information of the CFD model and the acoustic FEM model is transmitted and integrated. The coupling methods include the transfer of flow field information. The flow field results calculated by the CFD model are used as the input of the acoustic FEM model, especially as the input of the sound source term Q. According to the Lighthill acoustic analogy, the sound source term Q is expressed as:

[0128]

[0129] where T ij =ρu i u j +pδ ij -τ ij , τ ij is the viscous stress tensor, and δ ij is the Kronecker symbol.

[0130] The reinforcement learning algorithm introduced to optimize the multi-physics coupling model is the deep Q-network algorithm, which includes constructing a deep neural network to approximate the Q function. The Q function represents the expected cumulative reward that can be obtained after taking action a in state s. In each step, the agent selects action a t according to the current state s t , and after executing the action, a new state s t+1 and reward r t are obtained. The parameters of the Q network are updated through the following formula:

[0131]

[0132] where α is the learning rate and γ is the discount factor. The agent continuously interacts with the multi-physics coupling model and automatically adjusts the parameters in the model according to different working conditions and objectives, realizing the intelligence of the simulation process. Compared with the traditional method that relies on manual experience and repeated trial-and-error to adjust parameters, it greatly reduces manual intervention, improves the simulation efficiency, reduces the time and computing resource costs, and also reduces the differences in evaluation results caused by human factors.

[0133] IV. Result Fusion

[0134] The Bayesian fusion method is used to fuse the results simulated by the deep learning model and the multi-physics coupling model, and the model is feedback-adjusted using the fused results.

[0135] Fusing the results simulated by the deep learning model and the multi - physical - field coupling model includes assuming that the prediction result of the deep learning model is y DL and the result of the multi - physical - field coupling simulation is y CFD-FEM The fused result y fused is:

[0136] y fused = w DL y DL + w CFD-FEM y CFD-FEM

[0137] where w DL and w CFD-fem are weights, and w DL + w CFD-FEM = 1. The calculation of the weights is based on Bayes' formula:

[0138]

[0139] where P(y|D) is the probability of the prediction result y under the data D, and P(D) is the prior probability of the data.

[0140] Using the Bayesian fusion method to combine the results of the deep learning model and the multi - physical - field coupling simulation, taking into account the advantages of both methods, makes up for the limitations of a single method, further improves the reliability and stability of the evaluation results, and makes the noise evaluation more accurate.

[0141] Using the results after fusing the deep learning model and the multi - physical - field coupling model to perform feedback adjustment on the model. According to the error between the fused result and the actual measured noise value, the parameters of the deep learning model are adjusted using the back - propagation algorithm. At the same time, according to the differences between the multi - physical - field coupling simulation results and the actual flow field and noise characteristics, the parameters and structures of the CFD model and the acoustic FEM model are adjusted, further improving the accuracy and generalization ability of the overall evaluation method.

[0142] The above - mentioned specific embodiments are only several alternative embodiments of the present invention. Based on the technical solution of the present invention and the relevant inspirations of the above - mentioned embodiments, those skilled in the art can make various alternative improvements and combinations to the above - mentioned specific embodiments.

Claims

1. A method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan, characterized in that: The following steps are involved: Sensors are used to collect various data of the self-ventilation motor fan under different working conditions, and the collected data are pre-processed; Constructing a deep learning model, wherein the deep learning model adopts a hybrid model combining a convolutional neural network and a recurrent neural network, and training and optimizing the hybrid model; The propagation and radiation characteristics of the flow field and aerodynamic noise of the self-ventilating motor fan are simulated by a multi-physics field coupling model, and a reinforcement learning algorithm is introduced to optimize the multi-physics field coupling model; The Bayesian fusion method is used to fuse the simulation results of the deep learning model and the multi-physics field coupling model, and the fused results are used to perform feedback adjustment on the model.

2. A method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 1, characterized in that: The sensors include acoustic sensors, pressure sensors, and speed sensors. The collected data include noise signals, flow field parameters, and motor electrical parameters when the self-ventilating motor fan is running. The data preprocessing includes the following steps: First, the collected time domain noise signal is converted into a frequency domain signal using fast Fourier transform. The formula is as follows: Among them, x(n) is the time domain signal, X(k) is the frequency domain signal, N is the number of sampling points, k = 0, 1, ..., N-1, and then a bandpass filter is used to remove the interference components outside the specific frequency range and retain the frequency components related to aerodynamic noise; For different types of data, the minimum-maximum normalization method is used to unify its numerical range to the [0,1] interval. For data x, the normalization formula is as follows: Among them, x min and x max are the minimum and maximum values ​​of the data respectively.

3. The method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 1, characterized in that: In the deep learning model, the convolutional neural network part is composed of multiple convolutional layers, pooling layers and fully connected layers. The convolutional layer extracts spatial features in the flow field image data through the convolution kernel. The convolution operation formula is as follows: Where x is the input feature map, y is the output feature map, w is the convolution kernel, b is the bias, M and N are the sizes of the convolution kernel, the pooling layer is used to reduce the data dimension, and the maximum pooling operation formula is as follows: Among them, S is the pooling window, s is the pooling step size; The recurrent neural network part is used to process data with sequence characteristics, and introduces a long short-term memory network algorithm to process long sequence data. The long short-term memory network algorithm uses a forget gate f t , input gate i t , output gate o t and cell state C t To process time series data; For each time step t, the hidden layer of the recurrent neural network receives the input x at the current moment t And the output h of the hidden layer at the previous moment t-1 , and then calculate the output h of the hidden layer at the current moment t , and its calculation formula is as follows: h t =σ(W xh x t +W hh h t-1 +b h ) Among them, σ is the activation function, W xh , W hh is the weight matrix, b h is the bias vector.

4. A method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 3, characterized in that: The fusion of the convolutional neural network and the recurrent neural network adopts a feature interactive fusion method, which specifically includes: The flow field image data is input into the convolutional neural network to extract some features and obtain the intermediate feature map C 1 (i,j); The feature map C 1 (i,j) is converted to a one-dimensional vector And use it as the initial input state of the recurrent neural network or add it as an additional input to the input sequence of the recurrent neural network. Assume that at time step t=1, the input of the recurrent neural network is The hidden state at this time is updated to: As the time series progresses, in each time step of the recurrent neural network, the current hidden state h t Feedback to the convolutional neural network, so that the convolutional neural network updates its feature extraction according to the current time information, wherein h t As an additional input channel, it is added to the flow field image to form a new input I'(x,y) = (I(x,y),h t ), and then perform convolution operation on it: Among them, K'(m,n) and b' are the new convolution kernel and bias term; Multi-stage feature fusion: In multiple processing stages, the convolutional neural network and the recurrent neural network are calculated alternately, wherein in the first stage, the convolutional neural network is used to extract the initial spatial features of the flow field, and the results are input into the recurrent neural network, and in the subsequent stages, part of the output of the recurrent neural network is used as feedback to adjust the input of the convolutional neural network or the convolution kernel weight; Assume that in the nth stage, according to the output of the recurrent neural network Adjust the convolution kernel weight K of the convolutional neural network n (m,n), the adjustment formula is as follows: Where α is the learning rate and g(·) is the adjustment function based on the output of the recurrent neural network.

5. The method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 1, characterized in that: The multi-physics field coupling model establishment includes a CFD model and an acoustic FEM model. The CFD model is a flow field model of a self-ventilating motor fan established based on computational fluid dynamics. The flow field around the self-ventilating motor fan is simulated by solving fluid dynamics equations. The fluid dynamics equations include three-dimensional incompressible Reynolds-averaged Navier-Stokes equations, and the formula is as follows: Where ρ is the fluid density, is the average velocity component, is the average pressure, μ is the dynamic viscosity, u' i and u' j is the velocity fluctuation component, and the Reynolds stress term is closed using the standard k-∈ turbulence model, where k represents the turbulent kinetic energy and ∈ represents the turbulent kinetic energy dissipation rate. The standard k-ε turbulence model describes the characteristics of turbulence through the following equation: Turbulent kinetic energy equation: Among them, P k is the generation term of turbulent kinetic energy, μ t is the turbulent viscosity, σ k is the turbulent Prandtl number for the k-equation; The turbulent kinetic energy dissipation rate equation is: in, and is an empirical constant, σ ∈ is the turbulent Prandtl number of the ∈ equation.

6. A method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 5, characterized in that: The acoustic FEM model is used to convert the physical information of the flow field into acoustic information. Based on the wave equation, the wave equation for the propagation of sound waves in a homogeneous medium is: Wherein, p is the sound pressure, c is the sound velocity, and Q is the sound source term. In the application of the self-ventilating motor fan, the sound source term Q is obtained from the flow field calculated by the CFD model; The computational domain is divided into multiple finite element units. The sound pressure p in each unit is expressed using the interpolation function N, that is, p = ∑ i N i P i , where P i is the node sound pressure, N i is the shape function. The wave equation is integrated in each unit and the variational principle is applied to obtain the following discrete form of equations: Where [M] is the mass matrix, [C] is the damping matrix, [K] is the stiffness matrix, {p} is the nodal acoustic pressure vector, {F} is the equivalent nodal force vector, and for the mass matrix [M], its element M ij It is expressed as: M ij =∫ Ω ρN i N j dΩ For the stiffness matrix [K], its element K ij It is expressed as: The damping matrix [C] is calculated according to a specific damping model, including Rayleigh damping [C] = α [M] + β [K], where α and β are damping coefficients.

7. A method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 6, characterized in that: The information of the CFD model and the acoustic FEM model is transmitted and integrated through coupling. The coupling method includes the transmission of flow field information. The flow field result calculated by the CFD model is used as the input of the acoustic FEM model, especially as the input of the sound source term Q. According to the Lighthill acoustic analogy, the sound source term Q is expressed as: Among them, T ij =ρu i u j +pδ ij -τ ij , τ ij is the viscous stress tensor, δ ij is the Kronecker symbol.

8. The method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 1, characterized in that: The reinforcement learning algorithm used in the optimization of the multi-physics field coupling model by introducing the reinforcement learning algorithm is a deep Q network algorithm, including constructing a deep neural network for approximating the Q function, wherein the Q function represents the expected cumulative reward that can be obtained after taking action a in state s. In each step, the agent calculates the expected cumulative reward according to the current state s. t Select action a t , after executing the action, the new state s is obtained t+1 and reward r t , update the parameters of the Q network by the following formula: Among them, α is the learning rate, γ is the discount factor, and the intelligent agent automatically adjusts the parameters in the model according to different working conditions and goals by continuously interacting with the multi-physics field coupling model.

9. The method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 1, characterized in that: The deep learning model and the multi-physics coupling model simulation results are merged, including assuming that the deep learning model prediction result is y DL , the simulation result of the physical field coupling is y CFD-FEM , the fusion result y fused for: y fused =w DL y DL +w CFD-FEM y CFD-FEM Among them, w DL and w CFD-FEM is the weight, and w DL +w CFD-FEM =1, the weight is calculated based on the Bayesian formula: Among them, P(y|D) is the probability of predicting the result y under the data D, and P(D) is the prior probability of the data.

10. A method for simulating and evaluating aerodynamic noise of a self-ventilating motor fan according to claim 9, characterized in that: The result of the fusion of the deep learning model and the multi-physics field coupling model is used to perform feedback adjustment on the model. According to the error between the fusion result and the actual measured noise value, the parameters of the deep learning model are adjusted using the back propagation algorithm. At the same time, according to the difference between the multi-physics field coupling simulation result and the actual flow field and noise characteristics, the parameters and structure of the CFD model and the acoustic FEM model are adjusted.

Images