An automobile non-stationary aerodynamic noise active control system and method

Abstract

This invention discloses an active control system and method for automotive non-stationary aerodynamic noise. Addressing the problems of lack of coherent reference sources, poor convergence of linear algorithms, and insufficient real-time performance in broadband non-stationary aerodynamic noise under high-speed conditions, this system eliminates external reference sensors. Instead, it utilizes a microphone array to collect in-vehicle noise signals, extracts time-frequency features through wavelet packet decomposition, and constructs intrinsic modal components as a virtual reference source. This virtual reference source is then input into a deep neural network for nonlinear mapping to predict the control signals for the reverse sound waves in each frequency band. Finally, a bandpass filter with center frequency matching is used to combine these signals into a total reverse sound wave, driving the speaker to create a quiet zone at the occupant's ears. This invention overcomes hardware limitations by using a software-defined reference source, utilizes a deep neural network to fit turbulent nonlinear characteristics, and combines multi-layer sub-band parallel control to ensure real-time performance across the entire frequency band. It also possesses online adaptive fine-tuning capabilities, significantly improving the non-stationary wind noise suppression effect and system robustness.

Description

Technical Field

[0001] This invention belongs to the field of active control technology for automotive NVH (noise, vibration, and harshness), specifically relating to an active suppression system and method for broadband, non-stationary aerodynamic noise (wind noise) under high-speed driving conditions. This system utilizes wavelet packet decomposition to extract time-frequency features and employs deep neural networks for nonlinear prediction and speaker distortion compensation. Background Technology

[0002] With the popularization of new energy vehicles, the engine masking effect has disappeared, and aerodynamic noise at high speeds (>80km / h) has become the primary noise source inside the vehicle. Aerodynamic noise is characterized by a wide frequency band (100Hz-5kHz), strong non-stationarity (sudden changes with gusts and crosswinds), and lack of a coherent reference source (airborne propagation, which does not cause significant vibration of the vehicle body structure). Existing active noise cancellation technologies for automobiles have the following limitations, making them unable to effectively solve aerodynamic noise problems: (1) Inability to handle non-stationary signals: Existing road noise control schemes mostly use Fast Fourier Transform (FFT) for frequency domain processing. FFT is based on the assumption of stationary signals and cannot effectively capture the time-frequency local characteristics of transient changes (such as gusts) in aerodynamic noise, resulting in control lag or failure.

[0003] (2) Difficulty in obtaining reference signals: Existing technologies mostly rely on accelerometers (to collect road vibration) or engine speed signals as references, while some existing technologies rely on lidar to scan the road surface or accelerometers to collect vibrations. However, when dealing with aerodynamic noise, since airflow turbulence is transparent and random, lidar cannot detect its pulsations, and aerodynamic noise is not accompanied by significant chassis vibration. This causes such "physical feedforward" architectures that rely on external sensors to completely fail when facing wind noise. Aerodynamic noise is transmitted into the vehicle through direct air transmission or high-frequency fluid-structure interaction, and the accelerometer cannot collect a clean aerodynamic noise reference signal, causing traditional feedforward control systems to fail due to the lack of a coherent reference source.

[0004] (3) Limitations of linear control algorithms: The traditional adaptive filtering algorithm FxLMS is based on the assumptions of stationary signals and linear models. However, the carriage exhibits complex acoustic modes in the high-frequency band, and the generation mechanism of aerodynamic noise is nonlinear (Navier-Stokes equations), which has extremely strong nonlinear characteristics. When dealing with such sudden and wide-bandwidth signals, linear algorithms are prone to slow convergence speed or even system instability leading to "whistling" phenomena.

[0005] (4) The contradiction between algorithm complexity and real-time performance: In order to improve recognition accuracy, existing recognition technologies map one-dimensional sound signals to two-dimensional time-frequency diagrams for processing. This "dimensionality-upgrading" logic will generate huge computational overhead and signal oscillations caused by edge effects, making it difficult to meet the phase accuracy requirements for real-time cancellation of broadband noise when the car is traveling at high speed (>120km / h). Some existing technologies only generate reverse waves by looking up tables based on vehicle speed. This open-loop control cannot cope with noise abrupt changes caused by crosswinds, overtaking, etc., and the control accuracy is extremely low.

[0006] (5) Confusion between active control logic and signal denoising logic: Existing technologies mostly focus on signal enhancement and denoising for pipeline leaks or fan blades. Such methods are essentially "signal post-processing," and their goal is to obtain clean waveforms for fault diagnosis. Due to the high computational delay in their processing (such as VMD iteration or Wiener filtering), they cannot meet the stringent requirements of active noise control, that is, they cannot synthesize reverse sound waves in real time within milliseconds before the noise reaches the human ear. Summary of the Invention

[0007] The main objective of this invention is to provide an active control system and method for automotive non-stationary aerodynamic noise based on wavelet packet decomposition and deep neural networks, which solves the problems of poor noise reduction effect and system instability caused by the lack of effective reference signals, insufficient time-frequency resolution and poor fitting ability of linear algorithms when dealing with broadband random aerodynamic noise in existing technologies.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An active control method for non-stationary aerodynamic noise in automobiles based on wavelet packet decomposition and deep neural networks includes the following steps: S1: The raw aerodynamic noise signal is collected in real time using a microphone array located in the crew cabin, and then converted from analog to digital and subjected to anti-aliasing filtering by the signal acquisition and processing unit to obtain a digital noise signal; S2: Input the digital noise signal into the wavelet packet decomposition module to perform multi-layer wavelet packet decomposition to obtain residual components and multiple wavelet packet sub-band signals; perform discrete Fourier transform and power spectrum calculation on each sub-band signal respectively; determine the center frequency of each sub-band according to the distribution of power spectrum energy on the frequency axis; and construct a virtual reference signal by using sub-band signals with different center frequencies as intrinsic modal components. S3: Input the intrinsic modal components into the pre-trained deep neural network control module, and output the reverse acoustic wave control signal corresponding to each frequency band through nonlinear mapping; wherein, the deep neural network adopts a strategy of combining offline training and online fine-tuning, with the goal of minimizing the loss function between the network output and the desired reverse acoustic wave, and learns the nonlinear evolution law of aerodynamic noise caused by turbulent pressure fluctuations. S4: The reverse acoustic wave control signals of each frequency band are input to a bandpass filter bank whose center frequency matches the wavelet subband, and filtered and synthesized to obtain the total reverse acoustic wave signal, which drives the vehicle speaker to play, forming a quiet zone at the occupant's ear position to cancel the original aerodynamic noise.

[0009] The core idea of ​​this invention is to abandon the traditional external reference sensor, directly use the acoustic signal collected by the microphone, extract fine time-frequency features through wavelet packet decomposition to construct a virtual reference source, and use the nonlinear mapping capability of deep neural network to generate control signal.

[0010] In the above technical solution, the wavelet packet decomposition in step S2 adopts the Daubechies wavelet family, with a decomposition level J of 3 to 5, decomposing the broadband noise signal into 2^J wavelet packet subband signals.

[0011] In the above technical solution, the power spectrum is obtained by performing a discrete Fourier transform on each sub-band signal, taking the square of the modulus, and normalizing it; the center frequency is defined as the energy-weighted average frequency of the power spectrum.

[0012] In the above technical solution, the deep neural network in step S3 includes a multilayer perceptron structure with an input layer, at least two hidden layers and an output layer, or a convolutional neural network structure with a convolutional layer, a pooling layer, a fully connected layer and an output layer; the output layer has multiple output nodes that correspond one-to-one with each sub-band, and outputs the reverse acoustic wave control signal of each frequency band respectively.

[0013] In the above technical solution, the offline training in step S3 includes: collecting aerodynamic noise data under different vehicle speeds and wind conditions, extracting intrinsic modal components as training sample inputs through step S2, using the target reverse acoustic wave control signal obtained by simulation or actual measurement as training labels, and iteratively updating network parameters using the backpropagation algorithm; the online fine-tuning includes: dynamically correcting the bias term of the deep neural network output layer according to the residual noise signal collected by the error microphone, so as to adapt to the changes in the secondary path of the sound field inside the vehicle.

[0014] In the above technical solution, the bandpass filter bank in step S4 is composed of multiple parallel finite impulse response filters. The center frequency of each filter corresponds one-to-one with the center frequency of each wavelet subband obtained in step S2, and the order is set to 64 to 128. The total reverse acoustic wave signal is obtained by linear superposition of the reverse acoustic wave components of each frequency band.

[0015] In the above technical solution, the microphone array includes multiple capacitive microphones arranged in the roof lining, A-pillar and headrest, with a sampling rate of not less than 16kHz and a quantization bit depth of 16bit to 24bit.

[0016] In the above technical solution, the residual component in step S2 is obtained by reconstructing the low-frequency approximation coefficient and high-frequency detail coefficient through wavelet packet decomposition. It is used to characterize the narrowband noise component that is not completely covered by the wavelet packet subband and serves as the compensation channel for the input of the deep neural network.

[0017] In the above technical solution, the hidden layer of the deep neural network adopts the ReLU activation function, and the output layer adopts the linear activation function; the loss function is the mean squared error loss function or the Huber robust loss function.

[0018] In the above technical solution, the deep neural network described in step S3 introduces physical constraints on turbulent pressure fluctuations described by the incompressible Navier-Stokes equations during the training process, enabling the network to learn the nonlinear convection characteristics of aerodynamic noise in the time-frequency domain.

[0019] In the above technical solution, after step S4, the method further includes: using an error microphone to collect residual noise signals, and when the residual noise power exceeds a preset threshold, triggering the online fine-tuning mechanism of the deep neural network, with an update cycle of no more than 50 milliseconds.

[0020] An active control system for non-stationary aerodynamic noise in automobiles based on wavelet packet decomposition and deep neural networks includes: An acoustic acquisition array is placed in the roof lining, A-pillars and headrests to collect raw aerodynamic noise signals in the passenger compartment in real time. The signal acquisition and processing unit is connected to the acoustic acquisition array and is used to perform analog-to-digital conversion, anti-aliasing filtering and gain amplification on the original aerodynamic noise signal to obtain a digital noise signal. The algorithm control unit includes a wavelet packet decomposition module and a deep neural network control module. The wavelet packet decomposition module performs multi-level wavelet packet decomposition on the digital noise signal to obtain residual components and multiple wavelet packet sub-band signals. By performing discrete Fourier transform, power spectrum calculation, and center frequency extraction on each sub-band signal, the intrinsic modal components are determined as virtual reference signals. The deep neural network control module receives the intrinsic modal components and outputs inverse acoustic wave control signals corresponding to each frequency band through nonlinear mapping. The sound wave synthesis execution unit includes a multi-channel bandpass filter bank and a speaker system; the bandpass filter bank contains multiple parallel bandpass filters, the center frequency of each filter being matched with the center frequency of the wavelet subband, used to filter and synthesize the reverse sound wave control signals of each frequency band into a total reverse sound wave signal; the speaker system is used to play the total reverse sound wave signal, forming a quiet zone at the occupant's ear position.

[0021] In the above technical solution, the acoustic acquisition array includes at least four microphones, two of which are arranged in the front of the roof lining near the windshield, one is arranged inside the A-pillar, and one is arranged at the driver's headrest; the signal acquisition and processing unit integrates a synchronous sample-and-hold circuit and an analog-to-digital converter.

[0022] In the above technical solution, the wavelet packet decomposition module has a built-in wavelet basis function selection unit, which supports Daubechies, Symlets or Coiflets wavelet families, and the number of decomposition layers can be adjusted from 3 to 5 layers.

[0023] In the above technical solution, the deep neural network control module includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer; the convolutional layer is used to extract local features from the intrinsic modal components of the input; the pooling layer is used to reduce the feature dimension; the fully connected layer is used to flatten the features and construct a nonlinear mapping relationship; the output layer has K output nodes, each corresponding to a K inverse acoustic wave control signal, where K is equal to the number of subbands obtained by wavelet packet decomposition.

[0024] The above technical solution also includes a deep neural network training module, which includes: a training data acquisition unit for acquiring aerodynamic noise samples under different vehicle speeds and wind conditions; a feature construction unit for obtaining intrinsic modal component training features through wavelet packet decomposition and spectral analysis; a loss function calculation unit for comparing the network output with the target reverse acoustic wave control signal and calculating the loss value; and a parameter update unit for performing backpropagation and updating network parameters based on the loss value.

[0025] The above technical solution also includes an online fine-tuning module, which is connected to the deep neural network control module and is used to dynamically adjust the bias parameters of the deep neural network output layer according to the residual noise signal collected in real time by the error microphone, so as to adapt to window opening, changes in occupant position or changes in the secondary path of the sound field.

[0026] In the above technical solution, each bandpass filter in the bandpass filter bank is a finite impulse response filter, designed using the window function method or the equiripple method, with a passband ripple of no more than 0.5dB, a stopband attenuation of no less than 40dB, and an order of 64 to 128.

[0027] In the above technical solution, the speaker system includes multiple broadband speakers arranged in the door trim panel, ceiling and headrest, and each speaker channel corresponds one-to-one with the output channel of the bandpass filter group.

[0028] In the above technical solution, the algorithm control unit is integrated into an on-board digital signal processor or a field-programmable gate array, and the processing delay is no more than 2 milliseconds.

[0029] The above technical solution also includes a residual feedback channel, which feeds back the residual noise signal collected by the error microphone to the signal acquisition and processing unit for evaluating the active control effect and triggering online fine-tuning.

[0030] In summary, this invention proposes an active control architecture without a physical reference source. Utilizing the multi-resolution characteristics of wavelet packet decomposition (WPD), the intrinsic modal components (IMCs) in the time-frequency domain are directly extracted from the in-vehicle error microphone signal as a virtual reference signal, and then input into a deep neural network for nonlinear anti-phase acoustic wave prediction.

[0031] Compared with existing active noise reduction techniques based on FFT frequency domain processing or linear adaptive filtering, this invention has the following technical advantages and beneficial effects: (1) Overcoming hardware physical limitations to achieve a software-defined reference source. This invention eliminates expensive external reference sensors (such as lidar or chassis accelerometers) and directly utilizes acoustic signals collected by in-vehicle error microphones, extracting intrinsic modal components through wavelet packet decomposition as a virtual reference source. This architecture fundamentally solves the industry problem of being unable to obtain a pure feedforward reference signal due to the transparency of air propagation of aerodynamic noise, significantly reducing system hardware costs, wiring complexity, and integration difficulty, and is especially suitable for new energy vehicle platforms.

[0032] (2) Deeply coupled with fluid physics characteristics to enhance nonlinear noise reduction capability. Unlike traditional linear adaptive algorithms such as FxLMS, ​​this invention uses a deep neural network to directly fit the nonlinear evolution law of turbulent pressure fluctuations governed by the Navier-Stokes equations. The network can accurately predict sudden pulses caused by crosswinds and gusts, and compensate for the nonlinear distortion of the loudspeaker in the high-frequency broadband when generating reverse sound waves, effectively solving the problem of divergence or even howling in the high-frequency band due to the complexity of acoustic modes in traditional algorithms.

[0033] (3) Multi-layer sub-band parallel control ensures real-time performance and robustness across the entire frequency band. By using wavelet packet decomposition, the high-dimensional broadband control problem is reduced to multiple parallel narrow-band sub-band prediction problems. Combined with a lightweight deep neural network architecture, this avoids the huge computational overhead and edge effects caused by upgrading a one-dimensional signal to a two-dimensional time-frequency diagram. Each sub-band performs phase compensation independently, ensuring that the millisecond-level real-time requirements under high-speed driving conditions are met across the entire frequency band from 100Hz to 5kHz. Even in the face of noise transients, deterministic control response can be maintained.

[0034] (4) It has online adaptive fine-tuning capability to adapt to complex operating conditions. The system introduces an online bias correction mechanism based on residual error signals, which can dynamically track changes in the secondary path of the sound field inside the vehicle (such as window opening and closing, changes in the number and position of occupants). Compared with traditional open-loop lookup table methods or fixed parameter filters, this invention can maintain stable quietness under various extreme driving conditions, and the control robustness is significantly enhanced.

[0035] (5) Eliminating residual musical sounds and improving subjective listening quality. This invention avoids the nonlinear distortion, spectral smearing, and residual musical sound interference caused by traditional denoising algorithms when processing non-stationary signals by learning the physical laws of noise generation rather than simple threshold truncation or spectral envelope shaping. The reverse sound wave is highly matched with the original noise in the time and frequency domain, making the acoustic environment in the passenger cabin smoother and more natural, and significantly improving the driving comfort at high speeds. Attached Figure Description

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of the overall structure of the active aerodynamic noise control system for automobiles according to an embodiment of the present invention; Figure 2 This is a schematic flowchart of the active control method for automotive aerodynamic noise according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the deep neural network training process according to an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] Example 1: High-speed wind noise active control system based on WPD-CNN architecture like Figures 1-3 As shown, this embodiment provides a specific implementation of a non-stationary aerodynamic noise active control system and method based on wavelet packet decomposition and convolutional neural network applied to the passenger compartment of a pure electric car.

[0039] The system architecture of this invention comprises four main modules: (1) Acoustic acquisition array (input end): A microphone array (multiple microphone units 101) arranged in the roof lining, A-pillar and headrest, used to acquire wind noise signals. .

[0040] (2) Signal preprocessing unit ( Figure 1Signal acquisition and processing unit 102 (feature extraction): performs real-time multi-layer wavelet packet decomposition into residual components. and sub-band signal Extract time-frequency features.

[0041] (3) Deep neural control unit ( Figure 1 The algorithm control unit 103 in the model predicts the control parameters of the reverse acoustic wave based on the nonlinear control model of the MLP architecture.

[0042] (4) Sound field synthesis execution unit (output end, Figure 1 The acoustic synthesis unit 104 in the middle contains a multi-channel filter bank and a speaker system to emit anti-noise. .

[0043] The method of the present invention includes the following key steps: Step S1: Acquisition of non-stationary broadband acoustic signals: Using a microphone array at a sampling rate The raw noise signal inside the crew cabin was collected. The signal contains uncancelled aerodynamic noise, which satisfies:

[0044] Step S2: Time-frequency feature extraction based on wavelet packet decomposition (WPD) (1) Difference: Unlike the global transformation of FFT, this invention performs a transformation on digital noise signals. Layered wavelet packet decomposition, while preserving time information.

[0045] (2) Specific operation: Decompose the broadband noise signal into its components across the entire frequency band. Sub-band.

[0046] To each J-level wavelet packet decomposition is performed to decompose the signal into a residual component. and A small wavelet packet carries a signal ,Right now:

[0047] For each wavelet packet signal Perform a discrete Fourier transform to obtain the frequency domain representation. :

[0048] in, For discrete angular frequencies. Then calculate the... Power spectrum of the sub-band signal:

[0049] Based on the distribution of power spectral energy along the frequency axis, calculate the first... The center frequency of each sub-band It can be defined as:

[0050] For frequency point The corresponding physical frequency. Sub-band signals with different center frequencies. As an intrinsic modal component signal, it is denoted as:

[0051] (3) Technical effect: WPD can provide fine frequency resolution in both low and high frequency ranges, accurately capturing transient change components in aerodynamic noise.

[0052] (4) Feature construction: Calculate the energy spectrum and coefficient distribution of each wavelet subband and construct a set of "Intrinsic Mode Components (IMC)" as the virtual reference signal of the system.

[0053] Step S3: Nonlinear mapping of deep neural networks (DNN) (1) Difference: It abandons the traditional LMS linear adaptive filtering and adopts a deep neural network.

[0054] (2) Input: Used to receive the "intrinsic modal component" feature vector generated in step S2 .

[0055] (3) Model architecture: MLP (4) Output: Output the reverse acoustic wave control signal vector corresponding to each frequency band. ,satisfy:

[0056] in, The parameter is The mapping function of deep neural networks.

[0057] (5) Training method: An offline training strategy (collecting wind noise data under multiple operating conditions) + online fine-tuning is adopted to enable the network to learn the nonlinear evolution law of aerodynamic noise. The core of this step is to use the nonlinear fitting ability of deep neural networks to capture the characteristics of the incompressible Navier-Stokes (NS) equation that aerodynamic noise follows at the physical level, which is expressed as follows:

[0058] in, Representing the nonlinear convection term, it is the physical root cause of the randomness and strong nonstationarity of aerodynamic noise.

[0059] For each set of intrinsic modal component (IMC) feature vectors extracted in step S2 that contain the aforementioned physical features This invention minimizes Norm loss function To train network weights This allows it to approximate the complex pressure fluctuation evolution law described by the Navier-Stokes equations. Based on the desired noise attenuation effect in the control area, the target reverse acoustic wave control signal is designed or simulated. , as the network output label. The loss function is defined as:

[0060] in, N For the sample size, For network output, This is the expected output.

[0061] Step S4: Frequency band filtering synthesis and execution By using a set of bandpass filters whose center frequency matches the WPD subband, the control parameters output by the DNN are converted into actual analog drive signals.

[0062] (1) The filter combination consists of K parallel bandpass filters, the first of which is the first of the K parallel bandpass filters. The unit impulse response of a bandpass filter is Used to control the corresponding reverse acoustic wave signal Perform filtering to obtain the first... One reverse acoustic wave component :

[0063] in, For the first The order of a bandpass filter.

[0064] (2) The final reverse acoustic signal It is obtained by superimposing K frequency band components:

[0065] (3) Output: Drives the vehicle speaker to emit reverse sound waves This creates a quiet zone around the occupant's ears.

[0066] Example 2 Based on Example 1, this example provides the following specific implementation method: S1: Non-stationary broadband acoustic signal acquisition An acoustic acquisition array forming a noise acquisition module 201 is arranged within the passenger compartment. Specifically, it includes: two condenser microphone units 101 (model Knowles SPH0645LM4H) positioned at the front of the roof lining, near the junction of the windshield and the roof, spaced 15cm apart, to acquire wind noise from the windshield and A-pillar area; one microphone positioned inside the A-pillar trim panel on the driver's side to capture turbulence noise in the side window area; and one error microphone positioned to the left of the driver's headrest to acquire residual noise at the driver's ear position. All microphones are connected to the signal acquisition and processing unit via shielded twisted-pair cables.

[0067] The signal acquisition and processing unit 102 is preferably a noise signal digitization module 202, which uses a TI ADS1274 four-channel synchronous sampling analog-to-digital converter with a sampling rate set to 48kHz, a quantization bit depth of 24bit, and a built-in eighth-order Butterworth anti-aliasing filter with a cutoff frequency of 20kHz. The acquired raw aerodynamic noise signal is converted from analog to digital to obtain a digital noise signal x[n], which is then transmitted to the algorithm control unit via a CAN bus.

[0068] S2: Time-frequency feature extraction and virtual reference source construction based on wavelet packet decomposition The algorithm control unit 103 incorporates a wavelet packet decomposition module 203, which uses the Daubechies 4 (db4) wavelet as the basis function to perform J=3 layers of fully binary tree wavelet packet decomposition on the digital noise signal x[n], resulting in 2^3=8 wavelet packet subband signals b_k[n] (k=1,2,…,8) and a residual component r[n]. During the decomposition process, each layer simultaneously decomposes the low-frequency approximation coefficients and high-frequency detail coefficients to ensure balanced frequency resolution in the low-frequency and high-frequency bands within the range of 100Hz to 5kHz.

[0069] The algorithm control unit 103 has a built-in spectrum analysis module 204 and a center frequency calculation module 205.

[0070] The spectrum analysis module 204 is used to perform a windowed discrete Fourier transform on each wavelet packet sub-band signal b_k[n]. The window function is a Hanning window with 2048 FFT points and an overlap rate of 50%. The center frequency calculation module 205 is used to calculate the power spectrum P_k(ω) of each sub-band signal, and the center frequency f_{c,k} of each sub-band is determined based on the weighted average of the power spectrum energy. The eight sub-band signals with different center frequencies are used as intrinsic modal components u_k[n], which, together with the residual component r[n], form a 9-dimensional virtual reference signal vector, which is then input to the deep neural network control module.

[0071] S3: Generation of inverse acoustic control signals based on convolutional neural networks The core of the algorithm control unit 103 is the deep neural network control module 206. The deep neural network control module 206 adopts a one-dimensional convolutional neural network (1D-CNN) architecture, with the specific network structure as follows: Input layer: Receives 9-channel intrinsic modal component timing features with a length of 512 points; Convolutional layer 302: Contains two one-dimensional convolutional layers. The first layer has 32 convolutional kernels with a kernel size of 7, a stride of 2, and padding mode "SAME". The second layer has 64 convolutional kernels with a kernel size of 3 and a stride of 1. The activation function used for both layers is ReLU. Pooling layer 303: Max pooling layers are applied after two convolutional layers, with a pooling window size of 2, to reduce data dimensionality and extract translation-invariant features; Fully connected layer 304: The pooled feature map is flattened and then connected to two fully connected layers with 256 and 128 neurons respectively. The Dropout ratio is set to 0.3 to prevent overfitting. Output layer 305: contains K=8 linear output nodes, each corresponding to the inverse acoustic control signal y_k[n] of 8 wavelet subbands.

[0072] The network training employed a strategy combining offline training and online fine-tuning. In the offline training phase, aerodynamic noise data was collected for 120 hours under combined conditions of vehicle speeds of 80km / h, 100km / h, 120km / h, and 140km / h, and crosswind speeds of 0m / s, 5m / s, 10m / s, and 15m / s, in a wind tunnel and on actual roads. Step S2 was performed on the collected data to obtain the intrinsic modal components, which were then used as training sample inputs. Simultaneously, a swept-frequency signal was played through secondary sound source speakers placed at the headrests. The secondary sound field path from the main sound source to the error microphone was measured using the transfer function method, and a target back-propagation control signal was generated using an adaptive inverse modeling method as the training label. The Huber robust loss function was selected, and backpropagation training was performed using the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and 200 training iterations until the loss function converged.

[0073] In the online fine-tuning phase: After the system is deployed in a real vehicle, the error microphone continuously collects the residual noise signal e[n] at a sampling rate of 48kHz. The short-time energy of the residual noise is calculated every 40 milliseconds. If the energy exceeds a preset threshold of -40dB (relative to the original noise), online fine-tuning is triggered: the weights of the convolutional and fully connected layers are frozen, and only the bias term b_k of the 8 nodes in the output layer is updated using gradient descent. The update formula is b_k^{new}=b_k^{old}-η·e_k, where η=0.0005 is the fine-tuning learning rate, and e_k is the residual noise energy corresponding to the k-th sub-band. This mechanism enables the system to quickly adapt to changes in the secondary sound field path caused by window opening and closing and occupant position movement, without retraining the entire network.

[0074] S4: Frequency band filtering synthesis and sound field execution The acoustic synthesis execution unit 104 includes an inverse acoustic synthesis module 207 (composed of an 8-channel bandpass filter bank) and a speaker module 208 (containing 4 vehicle-mounted speakers). The bandpass filter bank uses a 128th-order finite impulse response (FIR) filter, designed with a Hamming window. The center frequency of each filter corresponds one-to-one with the center frequencies of the 8 wavelet packet sub-bands obtained in step S2: sub-band 1 covers 100-200Hz, sub-band 2 covers 200-400Hz, sub-band 3 covers 400-800Hz, sub-band 4 covers 800-1500Hz, sub-band 5 covers 1500-2500Hz, sub-band 6 covers 2500-3500Hz, sub-band 7 covers 3500-4500Hz, and sub-band 8 covers 4500-5000Hz. The passband ripple of each filter is controlled within 0.3dB, and the stopband attenuation is greater than 50dB.

[0075] The eight reverse acoustic wave control signals y_k[n] output from the deep neural network are input into corresponding bandpass filters. After filtering, the reverse acoustic wave components of each frequency band are obtained and linearly superimposed to form the total reverse acoustic wave signal y[n]. After digital-to-analog conversion and power amplification, the total signal drives four wideband speakers (frequency response 80Hz-6kHz) located on both sides of the driver's headrest, the center of the headliner, and the door trim panels. The speakers form a quiet zone with a diameter of about 15cm at the ear position, canceling out the original aerodynamic noise in real time.

[0076] Control effect verification Real-vehicle tests were conducted under conditions of constant speed of 120 km / h and crosswind of 10 m / s. The results show that the system in this embodiment achieves an average sound pressure level reduction of 15.2 dB (A-weighted) across the entire frequency band of 100 Hz-5 kHz. In the high-frequency band of 2 kHz-4 kHz, the traditional FxLMS algorithm, due to its inability to handle nonlinear distortion, achieves a noise reduction of only 3.1 dB and is accompanied by slight howling, while this system achieves a noise reduction of 12.8 dB, with stable operation and no divergence. Under simulated sudden gust wind conditions (crosswind increasing from 0 to 15 m / s within 0.5 seconds), the system response delay is 1.8 milliseconds, and the residual noise converges to a steady-state level within 80 milliseconds, significantly outperforming the control accuracy of the traditional frequency domain lookup table method.

[0077] Example 3: Lightweight Implementation Based on WPD-MLP Architecture As another specific implementation, the deep neural network control module 206 can be replaced with a multilayer perceptron (MLP). The input layer has 9×64 nodes (9-dimensional virtual reference signal, 64 temporal sampling points per dimension), the hidden layers are set to two layers with 512 and 256 neurons respectively, the activation function is LeakyReLU (negative slope 0.1), and the output layer has 8 linear nodes. The network parameters are approximately 1.2M, and a single-frame inference time of 0.6 milliseconds can be achieved on an automotive DSP chip (such as the TI TMS320C6678), meeting real-time requirements. The remaining steps are the same as in Example 1, suitable for cost-sensitive vehicle platforms.

[0078] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for active control of non-stationary aerodynamic noise in automobiles based on wavelet packet decomposition and deep neural networks, characterized in that, Includes the following steps: S1: The raw aerodynamic noise signal is collected in real time using a microphone array located in the crew cabin, and then converted from analog to digital and subjected to anti-aliasing filtering by the signal acquisition and processing unit to obtain a digital noise signal; S2: Input the digital noise signal into the wavelet packet decomposition module to perform multi-layer wavelet packet decomposition to obtain residual components and multiple wavelet packet sub-band signals; perform discrete Fourier transform and power spectrum calculation on each sub-band signal respectively; determine the center frequency of each sub-band according to the distribution of power spectrum energy on the frequency axis; and construct a virtual reference signal by using sub-band signals with different center frequencies as intrinsic modal components. S3: Input the intrinsic modal components into the pre-trained deep neural network control module, and output the reverse acoustic wave control signal corresponding to each frequency band through nonlinear mapping; wherein, the deep neural network adopts a strategy of combining offline training and online fine-tuning, with the goal of minimizing the loss function between the network output and the desired reverse acoustic wave, and learns the nonlinear evolution law of aerodynamic noise caused by turbulent pressure fluctuations. S4: The reverse acoustic wave control signals of each frequency band are input to a bandpass filter bank whose center frequency matches the wavelet subband, and filtered and synthesized to obtain the total reverse acoustic wave signal, which drives the vehicle speaker to play, forming a quiet zone at the occupant's ear position to cancel the original aerodynamic noise.

2. The method according to claim 1, characterized in that, The wavelet packet decomposition in step S2 adopts the Daubechies wavelet family, with a decomposition level J of 3 to 5, decomposing the broadband noise signal into 2^J wavelet packet sub-band signals; the power spectrum is obtained by performing discrete Fourier transform on each sub-band signal, taking the square of the modulus and normalizing it; the center frequency is defined as the power spectrum energy weighted average frequency.

3. The method according to claim 1, characterized in that, The deep neural network in step S3 includes a multilayer perceptron structure with an input layer, at least two hidden layers and an output layer, or a convolutional neural network structure with a convolutional layer, a pooling layer, a fully connected layer and an output layer; the output layer has multiple output nodes that correspond one-to-one with each sub-band, and outputs the reverse acoustic wave control signal of each frequency band respectively.

4. The method according to claim 1, characterized in that, The offline training in step S3 includes: collecting aerodynamic noise data under different vehicle speeds and wind conditions, extracting intrinsic modal components as training sample inputs through step S2, using the target reverse acoustic wave control signal obtained by simulation or actual measurement as training labels, and iteratively updating network parameters using the backpropagation algorithm; the online fine-tuning includes: dynamically correcting the bias term of the deep neural network output layer according to the residual noise signal collected by the error microphone to adapt to the changes in the secondary path of the sound field inside the vehicle.

5. The method according to claim 1, characterized in that, The bandpass filter bank in step S4 is composed of multiple parallel finite impulse response filters, and the center frequency of each filter corresponds one-to-one with the center frequency of each wavelet subband obtained in step S2; the total reverse acoustic signal is obtained by linear superposition of the reverse acoustic components of each frequency band.

6. The method according to claim 1, characterized in that, The residual component mentioned in step S2 is obtained by reconstructing the low-frequency approximation coefficient and high-frequency detail coefficient of wavelet packet decomposition. It is used to characterize the narrowband noise component that is not completely covered by the wavelet packet subband and serves as the compensation channel for the input of the deep neural network.

7. An active control system for non-stationary aerodynamic noise in automobiles based on wavelet packet decomposition and deep neural networks, characterized in that, include: An acoustic acquisition array is placed in the roof lining, A-pillars and headrests to collect raw aerodynamic noise signals in the passenger compartment in real time. The signal acquisition and processing unit is connected to the acoustic acquisition array and is used to perform analog-to-digital conversion, anti-aliasing filtering and gain amplification on the original aerodynamic noise signal to obtain a digital noise signal. The algorithm control unit includes a wavelet packet decomposition module and a deep neural network control module. The wavelet packet decomposition module performs multi-level wavelet packet decomposition on the digital noise signal to obtain residual components and multiple wavelet packet sub-band signals. By performing discrete Fourier transform, power spectrum calculation, and center frequency extraction on each sub-band signal, the intrinsic modal components are determined as virtual reference signals. The deep neural network control module receives the intrinsic modal components and outputs inverse acoustic wave control signals corresponding to each frequency band through nonlinear mapping. The sound wave synthesis execution unit includes a multi-channel bandpass filter bank and a speaker system; the bandpass filter bank contains multiple parallel bandpass filters, the center frequency of each filter being matched with the center frequency of the wavelet subband, used to filter and synthesize the reverse sound wave control signals of each frequency band into a total reverse sound wave signal; the speaker system is used to play the total reverse sound wave signal, forming a quiet zone at the occupant's ear position.

8. The active control system for non-stationary aerodynamic noise of automobiles based on wavelet packet decomposition and deep neural networks according to claim 7, characterized in that, The acoustic acquisition array includes at least four microphones, two of which are located at the front of the roof lining near the windshield, one on the inside of the A-pillar, and one at the driver's headrest; the signal acquisition and processing unit integrates a synchronous sample-and-hold circuit and an analog-to-digital converter.

9. The active control system for non-stationary aerodynamic noise of automobiles based on wavelet packet decomposition and deep neural networks according to claim 7, characterized in that, The wavelet packet decomposition module has a built-in wavelet basis function selection unit, which supports Daubechies, Symlets or Coiflets wavelet families, and the number of decomposition layers can be adjusted from 3 to 5 layers.

10. The active control system for non-stationary aerodynamic noise of automobiles based on wavelet packet decomposition and deep neural networks according to claim 7, characterized in that, It also includes a deep neural network training module, which includes: a training data acquisition unit for acquiring aerodynamic noise samples under different vehicle speeds and wind conditions; a feature construction unit for obtaining intrinsic modal component training features through wavelet packet decomposition and spectral analysis; a loss function calculation unit for comparing the network output with the target reverse acoustic wave control signal and calculating the loss value; and a parameter update unit for performing backpropagation and updating network parameters based on the loss value.

Images