Noise control method, system and equipment for micro engine and storage medium

Abstract

The invention discloses a noise control method, system and equipment for a micro engine and a storage medium, belongs to the technical field of engine noise control, and solves the technical problem that the micro engine is difficult to realize an ideal noise reduction effect. The method comprises the steps that intake / exhaust noise, mechanical noise and combustion noise data in the working process of the micro engine are collected, empirical mode decomposition is carried out, and noise signals are decomposed into a plurality of IMF components of intrinsic mode functions; constructing an LSTM neural network model; the LSTM neural network model is trained, and an EMD-EGMGWO-LSTM model is obtained; in the working process of the micro engine, noise data are collected in real time and input into the trained EMD-EGMGWO-LSTM model, and a predicted value of the noise data is obtained; and an anti-phase noise control signal which is opposite to the phase and matched with the amplitude is immediately generated to realize an efficient active noise reduction effect. The control strategy can be continuously optimized and adjusted, and the noise reduction effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of engine noise control, and more specifically, to a noise control method, system, device and storage medium for a micro engine. Background Art

[0002] Micro engines are widely used in many fields due to their small size and high power density. However, their compact design makes noise control a major problem.

[0003] Traditional physical noise reduction methods, such as sound insulation packages and mufflers, are difficult to implement effectively without affecting performance due to space limitations and weight considerations. In addition, the dynamic changes in the working state of micro-engines make the noise characteristics complex and changeable. Although the traditional active noise control (ANC) strategy based on the FxLMS algorithm relies on experience and test data, it is difficult to achieve ideal noise reduction effects due to the complexity of secondary path modeling, high algorithm complexity, and insufficient adaptability to changes in the dynamic noise characteristics of the engine. Summary of the invention

[0004] The technical problem to be solved by the present invention is to address the above-mentioned deficiencies in the prior art. One purpose of the present invention is to provide a noise control method for a micro engine.

[0005] A second objective of the present invention is to provide a noise control system for a micro-engine.

[0006] A third object of the present invention is to provide a computer device.

[0007] A fourth objective of the present invention is to provide a computer storage medium.

[0008] In order to achieve the above-mentioned first objective, the present invention provides a noise control method for a micro-engine, comprising the following steps:

[0009] Step 1. Data collection and preprocessing: The intake / exhaust noise, mechanical noise and combustion noise data of the micro-engine during operation are collected in real time through noise sensors, and empirical mode decomposition is performed to decompose the noise signal into several IMF components of intrinsic mode functions to obtain the different frequency band components of the micro-engine noise;

[0010] Step 2. Construct an LSTM neural network model according to the characteristics of micro-engine noise. The LSTM neural network model includes an input layer, a hidden layer, and an output layer. The input layer is responsible for receiving the IMF component of the intrinsic mode function or the time series fragment of the original noise data after the empirical mode decomposition preprocessing, the hidden layer extracts the noise characteristics through nonlinear transformation, and the output layer outputs the noise prediction value or control signal.

[0011] Step 3. Training and optimization: The LSTM neural network model is trained based on the micro-engine noise data under multiple working conditions. The number of neurons in the LSTM hidden layer and the learning rate are optimized through EGMGWO to obtain the EMD-EGMGWO-LSTM model capable of predicting micro-engine noise. At the same time, L2 regularization is used to prevent the EMD-EGMGWO-LSTM model from overfitting and achieve effective generalization of the EMD-EGMGWO-LSTM model.

[0012] Step 4. Realize real-time noise control. During the operation of the micro-engine, collect noise data in real time and input it into the trained EMD-EGMGWO-LSTM model to obtain the predicted value of the noise data. Based on the predicted value, generate an anti-phase noise control signal with opposite phase and matching amplitude in real time, and generate reverse sound waves or vibrations that offset the original noise waveform through the vibration suppression device, thereby achieving efficient active noise reduction effect.

[0013] As a further improvement, the intake / exhaust noise includes: exhaust pipe periodic noise, resonance noise in the engine pipe;

[0014] The fundamental frequency expression of the engine exhaust pipe periodic noise can be expressed as:

[0015]

[0016] In formula (1), P f Represents the periodic frequency of noise; N m Indicates the number of working cylinders in the engine; k n Indicates the maximum speed of the engine under ideal conditions; k Indicates the engine cylinder displacement;

[0017] The resonance noise in the engine duct can be expressed as:

[0018]

[0019] In formula (2), H f Indicates the resonance noise of engine intake / exhaust; S p Indicates the cross-sectional area of ​​the intake / exhaust duct; V p Indicates the total volume of all cylinders in the engine; L p Indicates the length of the intake / exhaust pipe.

[0020] Furthermore, engine combustion noise is the noise caused by pressure fluctuations and vibrations generated by the combustion of fuel in the cylinder when the engine is in operation, which can be described as follows:

[0021] The spectral density C(f) of combustion noise can be estimated by equation (3):

[0022] C(f)=|H(f) 2 ·P(f) (3)

[0023] Where H(f) is the transfer function of combustion noise, which is related to factors such as engine structure and combustion chamber shape; P(f) is the spectral density of combustion pressure;

[0024] Combustion noise sound pressure level L rz It can be described by formula (4):

[0025]

[0026] Among them, p rz (t) is the time domain signal of combustion noise, and T is the measurement time.

[0027] Furthermore, mechanical noise mainly comes from the friction, vibration and impact of the moving parts inside the engine. Mechanical noise includes structural dynamic noise and gear noise;

[0028] Structural dynamic noise refers to the noise excited by the friction and impact of moving parts and the noise generated by the inherent vibration of parts excited by mechanical forces. Its sound pressure level can be estimated by formula (5):

[0029]

[0030] Among them, p i (t) is the time domain signal of the ith noise source, and n is the number of noise sources;

[0031] Sound pressure level of gear noise L clz It can be estimated by formula (6):

[0032]

[0033] In the formula, p cl (t) is the time domain signal of gear noise.

[0034] Furthermore, the process of EMD processing includes:

[0035] The recursive mean value is calculated by formula (7):

[0036]

[0037] In formula (7), IMF represents the recursive mean value of the extreme value of the interpolation function; f max and f min Respectively represent the envelope of the maximum and minimum values ​​of the function;

[0038] The residual of each noise point compared to the extreme point can be obtained by averaging:

[0039] r m =x(t)-IMF (8)

[0040] In formula (8): r m represents the residual of each noise point relative to the current IMF, that is, the difference between the original signal and the IMF; x(t) represents the original noise signal; this formula is used to continuously update the signal to extract all IMF components;

[0041] By decomposing and filtering the characteristics represented by the modal classification at different frequencies, the filtering mechanism can be obtained as follows:

[0042]

[0043] In formula (9): μ sd It represents the residual ratio of the screening mechanism in any discrete sequence, which is used to determine whether the extraction of IMF components is complete; h k-1 (t) and h k (t) represent the screening components of two adjacent regions, i.e., the intermediate results in the IMF extraction process;

[0044] Under the screening mechanism, the output value of the engine noise signal can be obtained. At this time, the engine noise signal can be expressed as follows under different decomposition units:

[0045]

[0046] In formula (10), x(t) represents the original noise signal; IMF i represents the i-th IMF component, which represents the modal characteristics of the signal at different frequencies; r n Represents the final residual component, that is, the remaining signal part after all IMF components are extracted.

[0047] Furthermore, the principle of optimizing the number of neurons and learning rate of the LSTM hidden layer by EGMGWO in step 3 is to optimize the number of neurons and learning rate of the LSTM hidden layer by the GWO algorithm, and to improve the ability of the GWO algorithm to obtain the optimal parameters by introducing the dynamic adjustment of the convergence factor and the Gaussian variation local search mechanism. The process is as follows:

[0048] Step 31. Algorithm initialization, including:

[0049] 1) Population initialization, setting the population size N, search space dimension d, and maximum number of iterations T max ; Randomly generate the initial population position X i =(x i1 ,x i2 ,…,x id ), where i = 1, 2, ..., N;

[0050] 2) Parameter initialization, initialization convergence factor a init , decay constant T delay , Dynamically adjust the parameter λ min , Dynamically adjust the parameter λ max , dynamically adjust the parameter β; initialize the Gaussian mutation parameters, mean μ = 0, standard deviation σ; initialize the positions and fitness values ​​of Alpha, Beta, and Delta of the three gray wolf individuals; set the acceptance probability p of the Gaussian mutation local search accept ;

[0051] Step 32. Fitness evaluation, for each individual X i , use the LSTM hyperparameter configuration represented by it to train the LSTM neural network model, and use it to predict the engine noise data after empirical mode decomposition; calculate the mean square error (MSE) between the prediction result and the actual noise data as the fitness value f(X i ):

[0052]

[0053] Among them, y k is the real noise data, is the prediction result of the LSTM neural network model, and n is the number of data points;

[0054] Step 33. Location update, including:

[0055] 1) Calculate the coefficient vector and For each individual X i , generate random numbers r1,r2∈[0,1]; calculate the convergence factor of nonlinear attenuation:

[0056]

[0057] Compute the coefficient vector:

[0058] A i =2a(t)·r i -a(t), C i =2·sin(2πr2) (13);

[0059] 2) Calculate the dynamic adjustment factor based on the population fitness and the number of iterations:

[0060]

[0061] 3) The individual position of the gray wolf is updated. For each individual X i , according to its current position X i (t), Alpha's position Xα 、Beta's position X β 、Delta's position X δ , and the coefficient vector A i , C i And dynamically adjust the factor λ to update the individual position:

[0062]

[0063] 4) Gaussian mutation local search, for the updated individual X i Each dimension x of (t+1) ij (t+1), apply Gaussian mutation:

[0064]

[0065] Among them, N(0,1) represents the standard normal distribution;

[0066] For the newly generated candidate solution Perform fitness evaluation and obtain

[0067] According to the set acceptance probability p accept Update individual location:

[0068]

[0069] Where r is a randomly generated number, r∈[0,1];

[0070] Step 34. Boundary processing, assuming that the value range of the hyperparameters of the LSTM neural network model is [L j ,U j ], where L j and U j are the lower and upper bounds of the jth hyperparameter respectively; after each position update, the updated individual position X i (t+1)=(x i1 (t+1),x i2 (t+1),…,x id (t+1)) performs boundary checking and processing;

[0071] For each dimension j=1,2,…,d, perform the following operations:

[0072]

[0073] The individual position X after boundary processing can be realized i (t+1) satisfies the value range constraints of all hyperparameters;

[0074] Step 35. Iteration termination condition, repeat fitness evaluation, position update and boundary processing until the maximum number of iterations T is reached max ;

[0075] Step 36. Output the optimal solution, and output the optimal individual position X according to the fitness value best and its fitness value f(X best ), use X best The hyperparameter configuration of the representative LSTM neural network model trains the final EMD-EGMGWO-LSTM model.

[0076] Furthermore, the LSTM neural network model is constructed and optimized as follows:

[0077] Build a micro-engine noise suppression model and set the neural network to have three layers: one input layer, one hidden layer, and one output layer;

[0078] The input layer has 12 nodes and a linear activation function is selected;

[0079] The neural network structure used in the hidden layer is a standard LSTM layer, without adding an additional activation function, and the number of hidden layer nodes is optimized and determined according to the GWO algorithm;

[0080] The output layer has 1 node, Epoch is set to 500, and the gradient descent optimization algorithm is Adam;

[0081] The initial parameters of the GWO algorithm are: the number of wolf populations is 50, the maximum number of iterations is 500, the dimension is 3, the learning rate search range is 0.001 to 0.1, the hidden layer node search range is 10 to 100, and the L2 regularization coefficient search range is 0.001 to 0.1.

[0082] In order to achieve the above-mentioned second objective, the present invention provides a noise control system for a micro-engine, comprising:

[0083] The acquisition module is used for data acquisition and preprocessing. It uses noise sensors to collect the intake / exhaust noise, mechanical noise and combustion noise data of the micro-engine in real time during operation, and performs empirical mode decomposition to decompose the noise signal into several IMF components of intrinsic mode functions to obtain the different frequency band components of the micro-engine noise.

[0084] A model building module is used to build an LSTM neural network model according to the characteristics of micro-engine noise. The LSTM neural network model includes an input layer, a hidden layer and an output layer. The input layer is responsible for receiving the IMF component of the intrinsic mode function or the time series fragment of the original noise data after the empirical mode decomposition preprocessing, the hidden layer extracts the noise characteristics through nonlinear transformation, and the output layer outputs the noise prediction value or control signal.

[0085] The training module is used for training and optimization. The LSTM neural network model is trained based on the micro-engine noise data under multiple working conditions. The number of neurons in the LSTM hidden layer and the learning rate are optimized through EGMGWO to obtain the EMD-EGMGWO-LSTM model capable of predicting micro-engine noise. At the same time, L2 regularization is used to prevent the EMD-EGMGWO-LSTM model from overfitting and achieve effective generalization of the EMD-EGMGWO-LSTM model.

[0086] The control module is used to realize real-time noise control. During the operation of the micro-engine, the noise data is collected in real time and input into the trained EMD-EGMGWO-LSTM model to obtain the predicted value of the noise data. Based on the predicted value, an anti-phase noise control signal with opposite phase and matching amplitude is generated instantly, and a reverse sound wave or vibration is generated through a vibration suppression device to offset the original noise waveform, thereby achieving a highly efficient active noise reduction effect.

[0087] In order to achieve the third objective above, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the above-mentioned noise control method for a micro engine when executing the computer program.

[0088] In order to achieve the fourth objective above, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the noise control method of a micro-engine described above is implemented.

[0089] Beneficial Effects

[0090] Compared with the prior art, the present invention has the following advantages:

[0091] The present invention collects and processes the noise signals of the micro-engine in real time to perform intelligent analysis and control of the noise; through adaptive learning and optimization of the neural network model, the prediction and suppression capabilities of nonlinear noise are improved, and the noise control requirements under various complex working conditions can be met; through the real-time feedback mechanism, the control strategy can be continuously optimized and adjusted to improve the noise reduction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 is a flow chart of the present invention;

[0093] Figure 2 The figure is a flow chart of the improved Gaussian mutation gray wolf optimization algorithm in the present invention. DETAILED DESCRIPTION

[0094] The present invention will be further described below with reference to the specific embodiments in the accompanying drawings.

[0095] See also Figure 1-2 A noise control method for a micro engine includes the following steps 1 to 4:

[0096] Step 1. Data collection and preprocessing: The intake / exhaust noise, mechanical noise and combustion noise data of the micro-engine during operation are collected in real time through noise sensors, such as sound level meters, and empirical mode decomposition (EMD) is performed to decompose the noise signal into several IMF components of intrinsic mode functions to obtain the different frequency band components of the micro-engine noise.

[0097] Step 2. Construct an LSTM neural network model according to the characteristics of micro-engine noise. The LSTM neural network model includes an input layer, a hidden layer and an output layer. The input layer is responsible for receiving the IMF component of the intrinsic mode function or the time series fragment of the original noise data after empirical mode decomposition (EMD) preprocessing, the hidden layer extracts the noise characteristics through nonlinear transformation, and the output layer outputs the noise prediction value or control signal.

[0098] Step 3. Training and optimization: The LSTM neural network model is trained based on the micro-engine noise data under multiple working conditions. The number of neurons in the LSTM hidden layer and the learning rate are optimized through EGMGWO to obtain the EMD-EGMGWO-LSTM model capable of predicting micro-engine noise. At the same time, L2 regularization is used to prevent the EMD-EGMGWO-LSTM model from overfitting and achieve effective generalization of the EMD-EGMGWO-LSTM model.

[0099] Step 4. Realize real-time noise control. During the operation of the micro-engine, collect noise data in real time and input it into the trained EMD-EGMGWO-LSTM model to obtain the predicted value of the noise data. Based on the predicted value, generate an anti-phase noise control signal with opposite phase and matching amplitude in real time, and generate reverse sound waves or vibrations that offset the original noise waveform through the vibration suppression device, thereby achieving efficient active noise reduction effect.

[0100] The engine noise mainly comes from the impact of gas on the valve during the intake / exhaust process, mechanical noise and combustion noise.

[0101] The intake / exhaust noise includes: exhaust pipe periodic noise and resonance noise in the engine pipe. The engine intake / exhaust noise can be described by equations (1) to (2).

[0102] The fundamental frequency expression of the engine exhaust pipe periodic noise can be expressed as:

[0103]

[0104] In formula (1), Pf Represents the periodic frequency of noise; N m Indicates the number of working cylinders in the engine; k n Indicates the maximum speed of the engine under ideal conditions; k Indicates the engine cylinder displacement.

[0105] The resonance noise in the engine duct can be expressed as:

[0106]

[0107] In formula (2), H f Indicates the resonance noise of engine intake / exhaust; S p Indicates the cross-sectional area of ​​the intake / exhaust duct; V p Indicates the total volume of all cylinders in the engine; L p Indicates the length of the intake / exhaust pipe.

[0108] Engine combustion noise is the noise caused by pressure fluctuations and vibrations generated by the combustion of fuel in the cylinder when the engine is in operation. It can be described by equations (3) to (4):

[0109] The spectral density C(f) of combustion noise can be estimated by equation (3):

[0110] C(f)=|H(f) 2 ·P(f) (3)

[0111] Among them, H(f) is the transfer function of combustion noise, which is related to factors such as engine structure and combustion chamber shape; P(f) is the spectral density of combustion pressure.

[0112] Combustion noise sound pressure level L rz It can be described by formula (4):

[0113]

[0114] Among them, p rz (t) is the time domain signal of combustion noise, and T is the measurement time.

[0115] Mechanical noise mainly comes from the friction, vibration and impact of the moving parts inside the engine. Mechanical noise includes structural dynamic noise and gear noise, which can be described using equations (5) to (6):

[0116] Structural dynamic noise refers to the noise excited by the friction and impact of moving parts and the noise generated by the inherent vibration of parts excited by mechanical forces. Its sound pressure level can be estimated by formula (5):

[0117]

[0118] Among them, p i (t) is the time domain signal of the ith noise source, and n is the number of noise sources.

[0119] Sound pressure level of gear noise L clz It can be estimated by formula (6):

[0120]

[0121] In the formula, p cl (t) is the time domain signal of gear noise.

[0122] The process of EMD processing can be described using equations (7) to (10):

[0123] The recursive mean value is calculated by formula (7):

[0124]

[0125] In formula (7), IMF represents the recursive mean value of the interpolation function extreme value; f max and f min They represent the envelope of the maximum and minimum values ​​of the function respectively.

[0126] The residual of each noise point compared to the extreme point can be obtained by averaging:

[0127] r m =x(t)-IMF (8)

[0128] In formula (8): r m represents the residual amount of each noise point relative to the current IMF, that is, the difference between the original signal and the IMF; x(t) represents the original noise signal; this formula is used to continuously update the signal to extract all IMF components.

[0129] By decomposing and filtering the characteristics represented by the modal classification at different frequencies, the filtering mechanism can be obtained as follows:

[0130]

[0131] In formula (9): μ sd It represents the residual ratio of the screening mechanism in any discrete sequence, which is used to determine whether the extraction of IMF components is complete; h k-1 (t) and h k (t) represent the screening components of two adjacent regions, i.e., the intermediate results in the IMF extraction process.

[0132] Under the screening mechanism, the output value of the engine noise signal can be obtained. At this time, the engine noise signal can be expressed as follows under different decomposition units:

[0133]

[0134] In formula (10), x(t) represents the original noise signal; IMF i represents the i-th IMF component, which represents the modal characteristics of the signal at different frequencies; r n It represents the final residual component, that is, the remaining signal part after all IMF components are extracted. After EMD processing, the components of the engine noise signal under different empirical modes can be obtained.

[0135] The principle of optimizing the number of neurons and learning rate of the LSTM hidden layer through EGMGWO in step 3 is to optimize the number of neurons and learning rate of the LSTM hidden layer through the GWO algorithm, and to improve the ability of the GWO algorithm to obtain the optimal parameters by introducing the dynamic adjustment of the convergence factor and the Gaussian variation local search mechanism, so as to solve the problem of local optimality in the traditional GWO algorithm. Figure 2 As shown, the process is as follows:

[0136] Step 31. Algorithm initialization, including:

[0137] 1) Population initialization, setting the population size N, search space dimension d, and maximum number of iterations T max ; Randomly generate the initial population position X i =(x i1 ,x i2 ,…,x id ), where i = 1, 2,…, N.

[0138] 2) Parameter initialization, initialization convergence factor a init , decay constant T delay , Dynamically adjust the parameter λ min , Dynamically adjust the parameter λ max , dynamically adjust the parameter β; initialize the Gaussian mutation parameters, mean μ = 0, standard deviation σ; initialize the positions and fitness values ​​of Alpha, Beta, and Delta of the three gray wolf individuals; set the acceptance probability p of the Gaussian mutation local search accept .

[0139] Step 32. Fitness evaluation, for each individual X i , use the LSTM hyperparameter configuration represented by it to train the LSTM neural network model, and use it to predict the engine noise data processed by empirical mode decomposition (EMD). Calculate the mean square error (MSE) between the predicted result and the actual noise data as the fitness value f(X i ):

[0140]

[0141] Among them, y kis the real noise data, is the prediction result of the LSTM neural network model, and n is the number of data points.

[0142] Step 33. Location update, including:

[0143] 1) Calculate the coefficient vector and For each individual X i , generate random numbers r1,r2∈[0,1]; calculate the convergence factor of nonlinear attenuation:

[0144]

[0145] Compute the coefficient vector:

[0146] A i =2a(t)·r i -a(t), C i =2·sin(2πr2) (13);

[0147] 2) Calculate the dynamic adjustment factor based on the population fitness and the number of iterations:

[0148]

[0149] 3) The individual position of the gray wolf is updated. For each individual X i , according to its current position X i (t), Alpha's position X α 、Beta's position X β 、Delta's position X δ , and the coefficient vector A i , C i And dynamically adjust the factor λ to update the individual position.

[0150]

[0151] 4) Gaussian mutation local search, for the updated individual X i Each dimension x of (t+1) ij (t+1), apply Gaussian mutation:

[0152]

[0153] Among them, N(0,1) represents the standard normal distribution.

[0154] For the newly generated candidate solution Perform fitness evaluation and obtain

[0155] According to the set acceptance probability paccept Update individual location:

[0156]

[0157] Among them, r is a randomly generated number, r∈[0,1].

[0158] Step 34. Boundary processing, assuming that the value range of the hyperparameters of the LSTM neural network model is [L j ,U j ], where L j and U j are the lower and upper bounds of the jth hyperparameter respectively; after each position update, the updated individual position X i (t+1)=(x i1 (t+1),x i2 (t+1),…,x id (t+1)) performs boundary checking and processing.

[0159] For each dimension j=1,2,…,d, perform the following operations:

[0160]

[0161] The individual position X after boundary processing can be realized i (t+1) satisfies the value range constraints of all hyperparameters.

[0162] Step 35. Iteration termination condition, repeat fitness evaluation, position update and boundary processing until the maximum number of iterations T is reached max .

[0163] Step 36. Output the optimal solution, and output the optimal individual position X according to the fitness value best and its fitness value f(X best ), use X best The hyperparameter configuration of the representative LSTM neural network model trains the final EMD-EGMGWO-LSTM model.

[0164] The LSTM neural network model is constructed and optimized with the gray wolf as follows:

[0165] Build a micro-engine noise suppression model and set the neural network to have three layers: one input layer, one hidden layer, and one output layer;

[0166] The input layer has 12 nodes and a linear activation function is selected;

[0167] The neural network structure used in the hidden layer is a standard LSTM layer, without adding an additional activation function, and the number of hidden layer nodes is optimized and determined according to the GWO algorithm;

[0168] The output layer has 1 node, Epoch is set to 500, and the gradient descent optimization algorithm is Adam;

[0169] The initial parameters of the GWO algorithm are: the number of wolf populations is 50, the maximum number of iterations is 500, the dimension is 3, the learning rate search range is 0.001 to 0.1, the hidden layer node search range is 10 to 100, and the L2 regularization coefficient search range is 0.001 to 0.1.

[0170] A noise control system for a micro engine, comprising:

[0171] The acquisition module is used for data acquisition and preprocessing. It uses noise sensors to collect the intake / exhaust noise, mechanical noise and combustion noise data of the micro-engine in real time during operation, and performs empirical mode decomposition (EMD) to decompose the noise signal into several IMF components of intrinsic mode functions to obtain the different frequency band components of the micro-engine noise.

[0172] A model building module is used to build an LSTM neural network model according to the characteristics of micro-engine noise. The LSTM neural network model includes an input layer, a hidden layer and an output layer. The input layer is responsible for receiving the IMF component of the intrinsic mode function or the time series fragment of the original noise data after EMD preprocessing. The hidden layer extracts the noise characteristics through nonlinear transformation. The output layer outputs the noise prediction value or control signal.

[0173] The training module is used for training and optimization. The LSTM neural network model is trained based on the micro-engine noise data under multiple working conditions. The number of neurons in the LSTM hidden layer and the learning rate are optimized through EGMGWO to obtain the EMD-EGMGWO-LSTM model capable of predicting micro-engine noise. At the same time, L2 regularization is used to prevent the EMD-EGMGWO-LSTM model from overfitting and achieve effective generalization of the EMD-EGMGWO-LSTM model.

[0174] The control module is used to realize real-time noise control. During the operation of the micro-engine, the noise data is collected in real time and input into the trained EMD-EGMGWO-LSTM model to obtain the predicted value of the noise data. Based on the predicted value, an anti-phase noise control signal with opposite phase and matching amplitude is generated instantly, and a reverse sound wave or vibration is generated through a vibration suppression device to offset the original noise waveform, thereby achieving a highly efficient active noise reduction effect.

[0175] A computer device comprises a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the above-mentioned noise control method for a micro-engine is realized.

[0176] A computer-readable storage medium stores a computer program, which implements the above-mentioned noise control method for a micro-engine when executed by a processor.

[0177] The above are only preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A noise control method for a micro engine, characterized in that: The following steps are involved: Step 1. Data collection and preprocessing: The intake / exhaust noise, mechanical noise and combustion noise data of the micro-engine during operation are collected in real time by noise sensors, and empirical mode decomposition (EMD) is performed to decompose the noise signal into several IMF components of intrinsic mode functions to obtain the different frequency band components of the micro-engine noise; Step 2. According to the characteristics of micro-engine noise, an LSTM neural network model is constructed. The LSTM neural network model includes an input layer, a hidden layer and an output layer. The input layer is responsible for receiving the IMF component of the intrinsic mode function or the time series fragment of the original noise data after empirical mode decomposition (EMD) preprocessing, the hidden layer extracts the noise characteristics through nonlinear transformation, and the output layer outputs the noise prediction value or control signal. Step 3. Training and optimization: The LSTM neural network model is trained based on the micro-engine noise data under multiple working conditions. The number of neurons in the LSTM hidden layer and the learning rate are optimized through EGMGWO to obtain the EMD-EGMGWO-LSTM model capable of predicting micro-engine noise. At the same time, L2 regularization is used to prevent the EMD-EGMGWO-LSTM model from overfitting and achieve effective generalization of the EMD-EGMGWO-LSTM model. Step 4. Realize real-time noise control. During the operation of the micro-engine, collect noise data in real time and input it into the trained EMD-EGMGWO-LSTM model to obtain the predicted value of the noise data. Based on the predicted value, generate an anti-phase noise control signal with opposite phase and matching amplitude in real time, and generate reverse sound waves or vibrations that offset the original noise waveform through the vibration suppression device, thereby achieving efficient active noise reduction effect.

2. A noise control method for a micro engine according to claim 1, characterized in that: The intake / exhaust noise includes: exhaust pipe periodic noise, resonance noise in the engine pipe; The fundamental frequency expression of the engine exhaust pipe periodic noise can be expressed as: In formula (1), P f Represents the periodic frequency of noise; N m Indicates the number of working cylinders in the engine; k n Indicates the maximum speed of the engine under ideal conditions; k Indicates the engine cylinder displacement; The resonance noise in the engine duct can be expressed as: In formula (2), H f Indicates the resonance noise of engine intake / exhaust; S p Indicates the cross-sectional area of ​​the intake / exhaust duct; V p Indicates the total volume of all cylinders in the engine; L p Indicates the length of the intake / exhaust pipe.

3. The noise control method of a micro engine according to claim 1, characterized in that: Engine combustion noise is the noise caused by pressure fluctuations and vibrations generated by the combustion of fuel in the cylinder when the engine is in operation, which can be described as follows: The spectral density C(f) of combustion noise can be estimated by equation (3): C(f)=|H(f) 2 ·P(f) (3) Where H(f) is the transfer function of combustion noise, which is related to factors such as engine structure and combustion chamber shape; P(f) is the spectral density of combustion pressure; Combustion noise sound pressure level L rz It can be described by formula (4): Among them, p rz (t) is the time domain signal of combustion noise, and T is the measurement time.

4. The noise control method of a micro engine according to claim 1, characterized in that: Mechanical noise mainly comes from the friction, vibration and impact of the moving parts inside the engine. Mechanical noise includes structural dynamic noise and gear noise. Structural dynamic noise refers to the noise excited by the friction and impact of moving parts and the noise generated by the inherent vibration of parts excited by mechanical forces. Its sound pressure level can be estimated by formula (5): Among them, p i (t) is the time domain signal of the ith noise source, and n is the number of noise sources; Sound pressure level of gear noise L clz It can be estimated by formula (6): In the formula, p cl (t) is the time domain signal of gear noise.

5. The noise control method of a micro engine according to claim 1, characterized in that: The process of EMD processing includes: The recursive mean value is calculated by formula (7): In formula (7), IMF represents the recursive mean value of the interpolation function extreme value; f max and f min Respectively represent the envelope of the maximum and minimum values ​​of the function; The residual of each noise point compared to the extreme point can be obtained by averaging: r m =x(t)-IMF (8) In formula (8): r m represents the residual of each noise point relative to the current IMF, that is, the difference between the original signal and the IMF; x(t) represents the original noise signal; this formula is used to continuously update the signal to extract all IMF components; By decomposing and filtering the characteristics represented by the modal classification at different frequencies, the filtering mechanism can be obtained as follows: In formula (9): μ sd It represents the residual ratio of the screening mechanism in any discrete sequence, which is used to determine whether the extraction of IMF components is complete; h k-1 (t) and h k (t) represent the screening components of two adjacent regions, i.e., the intermediate results in the IMF extraction process; Under the screening mechanism, the output value of the engine noise signal can be obtained. At this time, the engine noise signal can be expressed as follows under different decomposition units: In formula (10), x(t) represents the original noise signal; IMF i represents the i-th IMF component, which represents the modal characteristics of the signal at different frequencies; r n Represents the final residual component, that is, the remaining signal part after all IMF components are extracted.

6. The noise control method of a micro engine according to claim 1, characterized in that: The principle of optimizing the number of neurons and learning rate of the LSTM hidden layer through EGMGWO in step 3 is to optimize the number of neurons and learning rate of the LSTM hidden layer through the GWO algorithm, and to improve the ability of the GWO algorithm to obtain the optimal parameters by introducing the dynamic adjustment of the convergence factor and the Gaussian variation local search mechanism. The process is as follows: Step 31. Algorithm initialization, including: 1) Population initialization, setting the population size N, search space dimension d, and maximum number of iterations T max ; Randomly generate the initial population position X i =(x i1 ,x i2 ,…,x id ), where i = 1, 2, ..., N; 2) Parameter initialization, initialization convergence factor a init , decay constant T delay , Dynamically adjust the parameter λ min , Dynamically adjust the parameter λ max , dynamically adjust the parameter β; initialize the Gaussian mutation parameters, mean μ = 0, standard deviation σ; initialize the positions and fitness values ​​of Alpha, Beta, and Delta of the three gray wolf individuals; set the acceptance probability p of the Gaussian mutation local search accept ; Step 32. Fitness evaluation, for each individual X i , use the LSTM hyperparameter configuration represented by it to train the LSTM neural network model, and combine it with the engine noise data processed by empirical mode decomposition (EMD) for prediction; calculate the mean square error (MSE) between the prediction result and the actual noise data as the fitness value f(X i ): Among them, y k is the real noise data, is the prediction result of the LSTM neural network model, and n is the number of data points; Step 33. Location update, including: 1) Calculate the coefficient vector and For each individual X i , generate random numbers r1,r2∈[0,1]; calculate the convergence factor of nonlinear attenuation: Compute the coefficient vector: A i =2a(t)·r i -a(t), C i =2·sin(2πr2) (13); 2) Calculate the dynamic adjustment factor based on the population fitness and the number of iterations: 3) The individual position of the gray wolf is updated. For each individual X i , according to its current position X i (t), Alpha's position X α 、Beta's position X β 、Delta's position X δ , and the coefficient vector A i , C i And dynamically adjust the factor λ to update the individual position: 4) Gaussian mutation local search, for the updated individual X i Each dimension x of (t+1) ij (t+1), apply Gaussian mutation: Among them, N(0,1) represents the standard normal distribution; For the newly generated candidate solution Perform fitness evaluation and obtain According to the set acceptance probability p accept Update individual location: Where r is a randomly generated number, r∈[0,1]; Step 34. Boundary processing, assuming that the value range of the hyperparameters of the LSTM neural network model is [L j ,U j ], where L j and U j are the lower and upper bounds of the jth hyperparameter respectively; after each position update, the updated individual position X i (t+1)=(x i1 (t+1),x i2 (t+1),…,x id (t+1)) performs boundary checking and processing; For each dimension j=1,2,…,d, perform the following operations: The individual position X after boundary processing can be realized i (t+1) satisfies the value range constraints of all hyperparameters; Step 35. Iteration termination condition, repeat fitness evaluation, position update and boundary processing until the maximum number of iterations T is reached max ; Step 36. Output the optimal solution. According to the fitness value, output the optimal individual position X best and its fitness value f(X best ), use X best The hyperparameter configuration of the representative LSTM neural network model trains the final EMD-EGMGWO-LSTM model.

7. A noise control method for a micro engine according to claim 6, characterized in that: The LSTM neural network model is constructed and optimized with the gray wolf as follows: Build a micro-engine noise suppression model and set the neural network to have three layers: one input layer, one hidden layer, and one output layer; The input layer has 12 nodes and a linear activation function is selected; The neural network structure used in the hidden layer is a standard LSTM layer, without adding an additional activation function, and the number of hidden layer nodes is optimized and determined according to the GWO algorithm; The output layer has 1 node, Epoch is set to 500, and the gradient descent optimization algorithm is Adam; The initial parameters of the GWO algorithm are: the number of wolf populations is 50, the maximum number of iterations is 500, the dimension is 3, the learning rate search range is 0.001 to 0.1, the hidden layer node search range is 10 to 100, and the L2 regularization coefficient search range is 0.001 to 0.

1.

8. A noise control system for a micro engine, characterized in that: include: The acquisition module is used for data acquisition and preprocessing. It uses noise sensors to collect the intake / exhaust noise, mechanical noise and combustion noise data of the micro-engine in real time during operation, and performs empirical mode decomposition (EMD) to decompose the noise signal into several IMF components of intrinsic mode functions to obtain the different frequency band components of the micro-engine noise. A model building module is used to build an LSTM neural network model according to the characteristics of micro-engine noise. The LSTM neural network model includes an input layer, a hidden layer and an output layer. The input layer is responsible for receiving the IMF component of the intrinsic mode function or the time series fragment of the original noise data after empirical mode decomposition (EMD) preprocessing, the hidden layer extracts the noise characteristics through nonlinear transformation, and the output layer outputs the noise prediction value or control signal. The training module is used for training and optimization. The LSTM neural network model is trained based on the micro-engine noise data under multiple working conditions. The number of neurons in the LSTM hidden layer and the learning rate are optimized through EGMGWO to obtain the EMD-EGMGWO-LSTM model capable of predicting micro-engine noise. At the same time, L2 regularization is used to prevent the EMD-EGMGWO-LSTM model from overfitting and achieve effective generalization of the EMD-EGMGWO-LSTM model. The control module is used to realize real-time noise control. During the operation of the micro-engine, the noise data is collected in real time and input into the trained EMD-EGMGWO-LSTM model to obtain the predicted value of the noise data. Based on the predicted value, an anti-phase noise control signal with opposite phase and matching amplitude is generated instantly, and a reverse sound wave or vibration is generated through a vibration suppression device to offset the original noise waveform, thereby achieving a highly efficient active noise reduction effect.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the noise control method for a micro engine described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the noise control method for a micro engine described in any one of claims 1 to 7 is implemented.

Images