Noise reduction method and system for high-performance TWS Bluetooth audio chip

Abstract

The invention discloses a noise reduction method and system for a high-performance TWS Bluetooth audio chip, and relates to the technical field of acoustoelectric integration, and the method comprises the steps: collecting an environment noise signal, synchronously obtaining a Bluetooth radio frequency interference signal, converting the Bluetooth radio frequency interference signal into a digital signal through a 24-bit analog-to-digital converter, and outputting a multi-modal noise feature matrix; performing noise separation and reverse sound wave generation on the multi-modal noise characteristic matrix by using a pre-trained lightweight CycleGAN model, and outputting a reverse sound wave signal; monitoring auditory meatus tightness parameters in real time, dynamically adjusting phase offset and frequency response balance parameters of the mixed noise reduction signals, and outputting optimized noise reduction signals adaptive to the current wearing state; and performing parameter calibration on the optimized noise reduction signal through anechoic chamber scene simulation and laser interferometer phase detection to generate a noise reduction control instruction set. According to the invention, the noise interference problem of the TWS earphone in a complex environment is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of acoustic and electrical integration technology, and in particular to a noise reduction method and system for a high-performance TWS Bluetooth audio chip. Background Art

[0002] Active noise cancellation (ANC) technology for true wireless stereo (TWS) Bluetooth earbuds has made significant progress in recent years. Traditional ANC systems rely primarily on feedforward and feedback microphones to capture ambient noise, using a digital signal processor (DSP) to generate inverse sound waves to achieve noise cancellation. Existing technologies use multi-band decomposition and a least mean square (LMS) algorithm to dynamically adjust noise cancellation parameters, effectively suppressing mid- and low-frequency ambient noise (such as traffic and air conditioning). This technology has been widely adopted in mainstream TWS earbuds, but its signal processing typically targets only acoustic noise and fails to fully consider the coupling effects of Bluetooth RF interference and power supply noise on the noise cancellation effect.

[0003] The performance of existing ANC systems in complex electromagnetic environments still needs improvement. Specifically, the Bluetooth chip's RF signals (such as 2.4GHz carrier harmonics) can couple into the audio signal chain through power and ground lines, introducing high-frequency background noise into the noise reduction signal. Furthermore, traditional noise separation algorithms (such as independent component analysis (ICA)) have limited modal decoupling capabilities for mixed noise, making it difficult to achieve high-precision sound field reconstruction while maintaining low power consumption. These issues can significantly reduce the depth and bandwidth of noise reduction in real-world scenarios, especially in conditions of dense RF interference or unstable wear. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a noise reduction method for a high-performance TWS Bluetooth audio chip to solve the problems of insufficient mixed noise decoupling capability and high-frequency background noise introduced by radio frequency.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, the present invention provides a noise reduction method for a high-performance TWS Bluetooth audio chip, which includes collecting an environmental noise signal and synchronously obtaining a Bluetooth radio frequency interference signal, converting it into a digital signal using a 24-bit analog-to-digital converter, and outputting a multimodal noise feature matrix; Use the pre-trained lightweight CycleGAN model to perform noise separation and reverse sound wave generation on the multimodal noise feature matrix, and output the reverse sound wave signal; Injecting reverse current into the reverse acoustic wave signal to cancel the power domain noise and generate a hybrid noise reduction signal; Real-time monitoring of ear canal sealing parameters, dynamic adjustment of the phase offset and frequency response equalization parameters of the hybrid noise reduction signal, and output of an optimized noise reduction signal adapted to the current wearing state; The optimized noise reduction signal is calibrated through anechoic chamber scene simulation and laser interferometer phase detection to generate a noise reduction control instruction set.

[0007] As a preferred solution of the noise reduction method for a high-performance TWS Bluetooth audio chip described in the present invention, the environmental noise signal includes industrial noise, construction noise, transportation noise and social life noise.

[0008] As a preferred solution of the noise reduction method for high-performance TWS Bluetooth audio chip described in the present invention, wherein: the Bluetooth radio frequency interference signal is synchronously acquired, converted into a digital signal using a 24-bit analog-to-digital converter, and a multimodal noise feature matrix is ​​output. The specific steps are as follows: Bluetooth RF interference signals are synchronously collected through a shared clock source, and RF coupling and anti-aliasing filtering are performed, while frequency band limiting filtering is performed on ambient noise signals. The filtered ambient noise signal and Bluetooth RF interference signal are input into the independent differential channels of a 24-bit analog-to-digital converter for synchronous quantization and noise shaping to obtain a digitized signal with time-stamp alignment. Based on the digitized signal with time stamp alignment, the acoustic component features are extracted by joint time-frequency analysis of the ambient noise signal. The electromagnetic component features are extracted by spectrum pulse detection of the Bluetooth radio frequency interference signal. Cross-modal analysis is then performed using a time series correlation algorithm to generate cross-modal features. The acoustic component features, electromagnetic component features and cross-modal features are fused to generate a multimodal noise feature matrix.

[0009] As a preferred solution of the noise reduction method for high-performance TWS Bluetooth audio chip described in the present invention, wherein: the pre-trained lightweight CycleGAN model is used to perform noise separation and reverse sound wave generation on the multimodal noise feature matrix, and output the reverse sound wave signal. The specific steps are as follows: Domain adversarial training is performed on the multimodal noise feature matrix. The distribution differences of the acoustic and electromagnetic component features are aligned through the gradient reversal layer to generate domain-invariant feature representations. The sliding time window is used to extract the temporal context dependencies and output the context-enhanced feature matrix. The context-enhanced feature matrix is ​​input into the pre-trained lightweight CycleGAN model, which separates the acoustic noise and electromagnetic noise components through cross-modal adversarial training and generates the initial reverse acoustic wave signal based on a multi-scale discriminator. The initial reverse acoustic wave signal is constrained for complex spectrum consistency, and the phase offset is corrected using a dynamic time warping algorithm to output a preliminary reverse acoustic wave signal aligned in the time domain. The time-varying noise reduction coefficient and phase compensation parameters are extracted from the time-domain aligned preliminary reverse acoustic wave signal, and dynamically adjusted through reinforcement learning strategy to generate the reverse acoustic wave signal.

[0010] As a preferred solution of the noise reduction method for high-performance TWS Bluetooth audio chip described in the present invention, wherein: the reverse current is injected into the reverse acoustic wave signal to offset the power domain noise and generate a mixed noise reduction signal. The specific steps are as follows: Real-time acquisition of power circuit ripple signals, extraction of switching frequency harmonic components and broadband noise floor characteristics through time-frequency analysis, and the acquisition of dynamic noise spectrum characteristics; The time-varying noise reduction coefficient and phase compensation parameters are combined with the dynamic noise spectrum characteristics to generate the reverse current injection amount and phase synchronization parameters. The time delay error of the reverse acoustic wave signal transmission path is calculated through the cross-correlation function, and a pre-distortion algorithm is used to generate the compensated reverse current signal. The reverse acoustic wave signal and the compensated reverse current signal are jointly aligned in time and frequency domains to generate a preliminary mixed noise reduction signal; The preliminary mixed denoised signal is monitored in real time and its time-frequency characteristics are extracted. The spectral consistency error with the target noise-free signal is calculated through a multi-scale discriminator. The reverse current injection amount and reverse acoustic wave phase compensation parameters are dynamically adjusted according to the spectrum consistency error to generate a hybrid noise reduction signal.

[0011] As a preferred solution of the noise reduction method for high-performance TWS Bluetooth audio chip described in the present invention, wherein: the real-time monitoring of ear canal sealing parameters, dynamic adjustment of the phase offset and frequency response equalization parameters of the mixed noise reduction signal, and output of the optimized noise reduction signal adapted to the current wearing state, the specific steps are as follows, Real-time collection of air pressure fluctuation data inside the ear canal to generate ear canal sealing parameters, and the phase offset of the hybrid noise reduction signal is calculated through a nonlinear dynamic model; According to the ear canal sealing parameters and phase offset, the frequency response equalization parameters of the hybrid noise reduction signal are calculated using the dynamic coupled resonance transfer function, and weighted correction is performed to generate the optimized frequency response equalization parameters; Based on the optimized frequency response equalization parameters, the phase offset is iteratively corrected through a multi-objective loss function to generate an optimized phase offset; The optimized phase offset and the optimized frequency response equalization parameters are superimposed on the mixed noise reduction signal to generate an optimized noise reduction signal adapted to the current ear canal wearing state.

[0012] As a preferred solution of the noise reduction method for a high-performance TWS Bluetooth audio chip described in the present invention, wherein: the optimized noise reduction signal is calibrated by anechoic room scene simulation and laser interferometer phase detection to generate a noise reduction control instruction set, the specific steps are as follows, A three-dimensional laser interferometer array was constructed in the anechoic chamber to measure the vibration phase data of the diaphragm at the entrance of the ear canal and reconstruct the sound pressure field distribution of the optimized noise reduction signal in the ear canal. Decompose the sound pressure field distribution into amplitude spectrum and phase spectrum, compare and analyze with the preset target spectrum of the anechoic room background noise, and generate the amplitude error distribution and phase error distribution of each frequency band; Construct a nonlinear transfer function mapping from amplitude error distribution and phase error distribution to frequency response equalization parameters and phase offset, and use quantum optimization algorithm to solve the optimal parameter correction that meets the error constraints; The optimal correction parameters are superimposed on the optimized noise reduction signal to generate a calibration signal. The vibration phase data of the calibration signal is measured again using a laser interferometer. If the error exceeds the limit, it is iteratively corrected until convergence. The converged frequency response equalization parameters and phase offset are encoded into a binary instruction sequence with priority marks, dynamic compensation instructions are inserted into key frequency bands, and a noise reduction control instruction set is generated.

[0013] In a second aspect, the present invention provides a noise reduction system for a high-performance TWS Bluetooth audio chip, comprising a noise acquisition module for collecting ambient noise signals and synchronously acquiring Bluetooth radio frequency interference signals, converting them into digital signals using a 24-bit analog-to-digital converter, and outputting a multimodal noise feature matrix; The noise separation module is used to perform noise separation and reverse sound wave generation on the multimodal noise feature matrix using the pre-trained lightweight CycleGAN model, and output the reverse sound wave signal; A noise reduction optimization module is used to offset the power domain noise by injecting reverse current into the reverse acoustic wave signal to generate a hybrid noise reduction signal; The calibration control module is used to monitor the ear canal sealing parameters in real time, dynamically adjust the phase offset and frequency response equalization parameters of the hybrid noise reduction signal, and output an optimized noise reduction signal adapted to the current wearing state; The power supply noise reduction module is used to calibrate the parameters of the optimized noise reduction signal through anechoic chamber scene simulation and laser interferometer phase detection, and generate a noise reduction control instruction set.

[0014] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the noise reduction method for a high-performance TWS Bluetooth audio chip as described in the first aspect of the present invention is implemented.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the noise reduction method for a high-performance TWS Bluetooth audio chip as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are: high-precision separation and reverse sound wave generation of multimodal noise are achieved through a pre-trained lightweight CycleGAN model, solving the problem of insufficient noise decoupling capability of traditional methods; at the same time, combined with real-time ear canal seal monitoring and dynamic parameter adjustment, the noise reduction signal is ensured to accurately match the wearing state, overcoming the adaptability defects of fixed parameter noise reduction. Through the coordinated cooperation of multimodal noise acquisition, power supply noise cancellation, and laser interferometer calibration technologies, the entire solution achieves professional-grade noise reduction performance under low power conditions, effectively solving the noise interference problem of TWS headphones in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 The figure is a flowchart of a noise reduction method for a high-performance TWS Bluetooth audio chip.

[0019] Figure 2 Schematic diagram of the noise reduction system for a high-performance TWS Bluetooth audio chip.

[0020] Figure 3 Flowchart of the method for generating multimodal noise feature matrix.

[0021] Figure 4 Flowchart of the adaptive optimization method for ear canal wearing status. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0025] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a noise reduction method for a high-performance TWS Bluetooth audio chip, comprising the following steps: S1. Collect environmental noise signals.

[0026] S1.1 Environmental noise signals include industrial noise, construction noise, transportation noise and social life noise.

[0027] It should be noted that environmental noise signals primarily fall into four categories: industrial noise refers to the noise generated by the operation of mechanical equipment during factory production, such as textile machine noise, with a noise level of 85 decibels; construction noise refers to the noise generated by the operation of construction equipment such as pile drivers and concrete mixers, with a noise level of 100 decibels for impact pile drivers; transportation noise refers to the noise generated by the operation of vehicles such as motor vehicles, aircraft, and trains, with a noise level of 90 decibels for heavy trucks; and social noise refers to the noise generated by commercial operations, entertainment venues, and public sound amplification equipment, with a noise level of 70 decibels for shopping mall broadcasts. These noises have different spectral characteristics and temporal distributions. Industrial noise is primarily low- to medium-frequency and continuous, construction noise is impulsive and has large fluctuations in sound level, transportation noise exhibits intermittent peaks that vary with vehicle volume, and social noise exhibits a distinct temporal pattern.

[0028] S2. Synchronously acquire the Bluetooth radio frequency interference signal, convert it into a digital signal using a 24-bit analog-to-digital converter, and output a multimodal noise feature matrix.

[0029] S2.1. Synchronously collect Bluetooth RF interference signals through a shared clock source, perform RF coupling and anti-aliasing filtering, and perform frequency band limiting filtering on ambient noise signals.

[0030] It should be explained that a shared clock source ensures time synchronization between the Bluetooth RF interference signal acquisition device and the environmental noise signal acquisition device. After receiving the 2.4GHz band signal, the Bluetooth RF interference signal acquisition device first performs RF coupling processing to match impedance and improve signal transmission efficiency. The signal then passes through an anti-aliasing filter to suppress high-frequency noise and prevent sampling aliasing. During the synchronous acquisition process, the environmental noise signal acquisition device uses a band-limiting filter to band-limit industrial noise, construction noise, transportation noise, and social life noise, retaining the valid frequency band signal and filtering out out-of-band interference. RF coupling processing uses a directional coupler to extract the signal. The anti-aliasing filter cutoff frequency is set slightly higher than the highest frequency of the Bluetooth signal. The band-limiting filter has different passband ranges based on the characteristics of the environmental noise. For example, the industrial noise passband is set to 50Hz-2kHz.

[0031] S2.2. Input the filtered ambient noise signal and Bluetooth radio frequency interference signal into independent differential channels of a 24-bit analog-to-digital converter, perform synchronous quantization and noise shaping, and obtain a digitized signal with time stamp alignment.

[0032] It should be noted that the filtered ambient noise signal and Bluetooth RF interference signal are connected to two independent differential channels of a 24-bit analog-to-digital converter, respectively. The differential channels use a symmetrical input structure to suppress common-mode interference. The 24-bit analog-to-digital converter synchronously samples the two signals under the control of a shared clock source. During the sampling process, Σ-Δ modulation technology is used to achieve noise shaping, pushing the quantization noise to the high frequency range. The quantized ambient noise signal and Bluetooth RF interference signal are both accurately timestamped. The timestamps are generated by the frequency-divided signal of the shared clock source to ensure the time alignment accuracy of the two digitized signals. The output data format of the 24-bit analog-to-digital converter is in two's complement format. The number of quantization levels is 2^24, and the sampling rate is set according to the highest frequency of the signal. For example, the sampling rate of the ambient noise signal is 48kHz.

[0033] S2.3. Based on the timestamp-aligned digitized signal, perform a joint time-frequency domain analysis on the ambient noise signal to extract acoustic component features. Perform spectrum pulse detection on the Bluetooth RF interference signal to extract electromagnetic component features. Then, perform cross-modal analysis using a time series correlation algorithm to generate cross-modal features. It should be explained that, based on the timestamp-aligned digitized signal, a short-time Fourier transform is used to perform time-frequency analysis on the ambient noise signal, extracting acoustic component features such as sound pressure level, spectral centroid, and zero-crossing rate. The analysis window length is exemplified as 1024 points. Power spectral density estimation is performed on the Bluetooth RF interference signal, detecting pulse interference features within the 2.4 GHz frequency band, and extracting electromagnetic component features such as peak amplitude, pulse width, and repetition period. The frequency resolution is exemplified as 1 MHz. A dynamic time warping algorithm is used to establish a temporal correspondence between the acoustic component features of the ambient noise signal and the electromagnetic component features of the Bluetooth RF interference signal. The cross-correlation function is calculated to evaluate the cross-modal coupling strength, ultimately generating a cross-modal feature vector containing acoustic-electromagnetic correlation characteristics. The feature dimension is exemplified as 128 dimensions.

[0034] S2.4. Fuse the acoustic component features, electromagnetic component features, and cross-modal features to generate a multimodal noise feature matrix.

[0035] It should be noted that after aligning the acoustic, electromagnetic, and cross-modal features by timestamp, a feature-level fusion approach is used to construct a multimodal noise feature matrix. Acoustic component features include parameters such as time-domain sound pressure level and frequency-domain spectral centroid. Electromagnetic component features include parameters such as pulse amplitude and spectral power. Cross-modal features include parameters such as acoustic-electromagnetic cross-correlation strength. The rows of the multimodal noise feature matrix correspond to sampling time points, while the columns arrange the acoustic, electromagnetic, and cross-modal features, respectively. The eigenvalues ​​of the acoustic, electromagnetic, and cross-modal features are Z-score normalized. The matrix is ​​stored in column-major format, with a data type of 32-bit floating point. The time dimension is exemplified as 1000 frames, and the feature dimension is exemplified as 64 dimensions. The multimodal noise feature matrix retains both the original features and the cross-modal correlation features, forming a complete noise representation.

[0036] S3. Use the pre-trained lightweight CycleGAN model to perform noise separation and reverse sound wave generation on the multimodal noise feature matrix, and output the reverse sound wave signal.

[0037] S3.1. Perform domain adversarial training on the multimodal noise feature matrix, and align the distribution differences of the acoustic component features and the electromagnetic component features through the gradient reversal layer to generate domain-invariant feature representation. Use the sliding time window to extract the temporal context dependency and output the context-enhanced feature matrix.

[0038] It should be explained that when domain adversarial training is implemented on the multimodal noise feature matrix, a gradient reversal layer is added during the feature extraction process. The gradient reversal layer multiplies the discriminator gradient by a negative coefficient during back propagation, so that the distribution difference between the acoustic component features and the electromagnetic component features in the latent space is reduced. During the training process, the strategy of minimizing the domain classification loss and maximizing the feature confusion is adopted so that the generated domain-invariant feature representation retains the original feature discriminability and eliminates modal differences. In the temporal processing stage, a sliding time window is used to truncate the domain-invariant feature representation. The window length example takes 10 frames, and the step size example takes 5 frames. The temporal context dependency of the features in the window is extracted through a bidirectional long short-term memory network. The final output context-enhanced feature matrix contains a time dimension and a feature dimension. The time dimension is determined by the number of sliding windows, and the feature dimension example takes 128 dimensions. It should also be noted that when training the bidirectional long short-term memory network, the input sequence is a domain-invariant feature representation captured by a sliding time window. The sequence length is determined by the window parameter, assuming 10 frames in this example. The network structure consists of two long short-term memory layers, one forward and one backward, with 64 hidden units in this example. The output layer concatenates the forward and backward features into a 128-dimensional feature vector through a fully connected network. The training process uses the mean squared error loss function and the Adam algorithm as the optimizer. The initial learning rate is 0.001 in this example, and the batch size is 32 in this example. To prevent overfitting, a dropout layer is added between the input and hidden layers, with a dropout rate of 0.2 in this example, and gradient clipping is applied to constrain the gradient norm. Training is terminated when the validation set loss decreases by no more than 0.1% in this example for five consecutive epochs, and the final model's temporal feature reconstruction error on the test set is within 0.05 in this example.

[0039] S3.2. Input the context-enhanced feature matrix into the pre-trained lightweight CycleGAN model, separate the acoustic noise and electromagnetic noise components through cross-modal adversarial training, and generate the initial reverse acoustic wave signal based on the multi-scale discriminator.

[0040] It should be noted that after the context-enhanced feature matrix is ​​input into the pre-trained lightweight CycleGAN model, the generator network first performs modal separation on the acoustic noise component and the electromagnetic noise component, and uses a residual network structure to extract deep features. The lightweight CycleGAN model contains two generator networks and two multi-scale discriminator networks. The generator network learns the mutual conversion relationship between acoustic and electromagnetic noise through cross-modal adversarial training. The multi-scale discriminator network evaluates the generation quality under different receptive fields. The receptive field size examples are 16×16, 32×32, and 64×64. The cycle consistency loss is maintained during the training process to ensure that the acoustic noise component maintains feature consistency after bidirectional conversion. Finally, the generator network outputs the initial inverse sound wave signal; It should also be noted that the pre-trained lightweight CycleGAN model was trained using a paired acoustic-electromagnetic noise dataset, with the input and output dimensions matching the 128-dimensional features of the context-enhanced feature matrix. The generator network uses a nine-residual block architecture, with a convolution kernel size of 3×3, a number of channels of 64, and two downsampling cycles. The discriminator network constructs three parallel multi-scale subnetworks, processing feature maps with receptive fields of 16×16, 32×32, and 64×64, respectively. Training utilizes an alternating optimization strategy, first fixing the generator to update the discriminator, then fixing the discriminator to update the generator. The adversarial loss weight for each iteration is 1.0, and the cycle consistency loss weight is 10.0. The Adam algorithm is used as the optimizer, with an initial learning rate of 0.0002, a momentum parameter β1 of 0.5, and a batch size of 16. The training termination condition is that the ratio of the generator loss to the discriminator loss stabilizes in the range of 0.8-1.2 for more than 10,000 iterations, and the modal conversion accuracy of the final model on the test set reaches more than 95% of the example value.

[0041] S3.3. Perform complex spectrum consistency constraints on the initial reverse acoustic wave signal, use the dynamic time warping algorithm to correct the phase offset, and output a preliminary reverse acoustic wave signal that is aligned in the time domain.

[0042] It should be noted that the initial reverse acoustic wave signal is processed with a complex spectrum consistency constraint. The signal is converted to the frequency domain via a short-time Fourier transform. The analysis window length is 1024 points, and the overlap ratio is 75%. This optimizes the phase information while maintaining the spectral amplitude consistent with the original ambient noise signal. A dynamic time warping algorithm is then applied to calculate the time offset between the initial reverse acoustic wave signal and the ambient noise signal. The waveform is stretched or compressed in the time domain to eliminate phase deviation. The warping step size constraint is exemplified by ±5ms. After complex spectrum reconstruction, the output is a time-domain aligned preliminary reverse acoustic wave signal.

[0043] S3.4. Extract the time-varying noise reduction coefficient and phase compensation parameters from the time-domain aligned preliminary reverse acoustic wave signal, dynamically adjust them through reinforcement learning strategy, and generate the reverse acoustic wave signal.

[0044] It should be noted that a sliding time window is used to extract the time-varying noise reduction coefficient and phase compensation parameters from the time-domain aligned preliminary reverse acoustic wave signal. The window length is 20ms, and the step size is 10ms. The time-varying noise reduction coefficient is obtained by calculating the instantaneous energy ratio between the preliminary reverse acoustic wave signal and the ambient noise signal, and the phase compensation parameter is determined by extracting the instantaneous phase difference through Hilbert transform. A reinforcement learning strategy based on a deep Q-network dynamically adjusts these parameters. The state space contains the noise reduction coefficient and phase compensation parameters for the current frame, as well as the state features of the past five frames. The action space is the noise reduction coefficient adjustment amount (example value ±3dB) and the phase compensation amount (example value ±0.1π). During training, signal-to-noise ratio improvement is used as the reward function. After policy optimization, the smoothness of the time-varying noise reduction coefficient of the generated reverse acoustic wave signal is improved by 30%, and the phase compensation accuracy reaches 0.05π, which is an example value. The final reverse acoustic wave signal is output.

[0045] S4. Injecting a reverse current into the reverse acoustic wave signal to offset the power domain noise, thereby generating a hybrid noise reduction signal.

[0046] S4.1. Real-time acquisition of power supply circuit ripple signals. Extraction of switching frequency harmonic components and broadband noise floor characteristics through time-frequency analysis to obtain dynamic noise spectrum characteristics.

[0047] It should be noted that when acquiring power supply circuit ripple signals in real time, an isolated differential probe is used to obtain the AC component on the DC power line. The sampling rate is set according to the operating frequency of the switching power supply, for example, 1 MHz. A short-time Fourier transform (SFT) time-frequency analysis is performed on the acquired ripple signal. The Hanning window function is used, with an example window length of 4096 points and an example overlap ratio of 50%. The fundamental and harmonic component characteristics of the switching frequency are extracted from the time-frequency spectrum, including the center frequency, amplitude, and phase of each harmonic. The harmonic order analysis range is 1-20. The broadband noise floor characteristics of each frequency band are also calculated, using 1 / 3 octave analysis to obtain the noise power in different frequency bands. The frequency band division is 20 Hz-100 kHz. The dynamic noise spectrum characteristics include time-varying harmonic component parameters and broadband noise floor parameters. The time resolution is 10 ms and the frequency resolution is 100 Hz.

[0048] S4.2. The time-varying noise reduction coefficient and phase compensation parameters are combined with the dynamic noise spectrum characteristics to generate the reverse current injection amount and phase synchronization parameters. The time delay error of the reverse acoustic wave signal transmission path is calculated through the cross-correlation function, and the pre-distortion algorithm is used to generate the compensated reverse current signal.

[0049] It should be explained that the time-varying noise reduction coefficient and phase compensation parameters are integrated with the dynamic noise spectrum characteristics. The time-varying noise reduction coefficient is weighted by 1 / 3 octave frequency bands, with the weighting coefficient determined based on the noise power contribution of each frequency band, for example, taking a value of 0.1-0.8. The phase compensation parameters are adjusted synchronously with the phase characteristics of the harmonic components, with an adjustment step size constraint of ±0.05π as an example. The delay error of the reverse acoustic signal in the transmission path is calculated using a cross-correlation function, with a calculation window length of 50ms and a delay resolution of 10μs as an example. A predistortion algorithm is used to compensate the reverse current signal, with a predistortion filter order of 16 as an example. The compensation range covers both delay error and amplitude distortion, with an amplitude compensation accuracy of ±0.5dB as an example. The resulting compensated reverse current signal maintains a sampling rate of 1MHz and is time-synchronized with the power supply circuit ripple signal. The synchronization error is controlled within a value of 20μs, for example, and the current amplitude dynamic range matches the original noise characteristics.

[0050] S4.3. Jointly align the reverse acoustic wave signal and the compensated reverse current signal in the time and frequency domains to generate a preliminary mixed noise reduction signal.

[0051] It should be noted that the reverse acoustic signal and the compensated reverse current signal are time-aligned, and a dynamic time warping algorithm is used to eliminate timing deviations caused by sampling rate differences. The warping window width is, for example, ±5ms. The aligned signals are converted to the frequency domain using a short-time Fourier transform. The analysis window length is, for example, 1024 points, and the overlap ratio is, for example, 75%. A weighted fusion of the amplitude and phase spectra of the two signals is performed in the frequency domain. The reverse acoustic signal is the primary weight in the acoustic frequency band (for example, 20Hz-20kHz), with a weight coefficient of, for example, 0.7. The compensated reverse current signal is the primary weight in the electromagnetic frequency band (for example, 20kHz-100kHz), with a weight coefficient of, for example, 0.6. Phase continuity is maintained across all frequency bands during the fusion process, and the phase transition band is smoothed using cosine smoothing. The transition bandwidth is, for example, 1kHz. This ultimately generates a preliminary mixed noise-reduced signal.

[0052] S4.4. Monitor the preliminary mixed denoised signal in real time and extract its time-frequency characteristics, and calculate the spectral consistency error with the target noise-free signal through a multi-scale discriminator.

[0053] It should be noted that when monitoring the preliminary hybrid denoised signal in real time, a sliding time window is used to extract time-frequency characteristics. The window length is 50ms, and the step size is 25ms. Multiresolution analysis is performed on the signal within each window, using a three-layer wavelet packet decomposition to obtain the energy distribution of different frequency bands. The wavelet basis function used is db4, and the number of decomposition layers corresponds to the frequency resolution, for example, 125Hz, 250Hz, and 500Hz. The multi-scale discriminator consists of three subnetworks operating in parallel, processing characteristics of different frequency bands: the low-frequency subnetwork (for example, 20-500Hz) uses a fully connected structure, the mid-frequency subnetwork (for example, 500-5kHz) uses a one-dimensional convolutional structure, and the high-frequency subnetwork (for example, 5kHz-20kHz) uses an attention mechanism. Spectral consistency error is calculated using a weighted mean squared error criterion, with weights of 0.5 for the low-frequency band, 0.3 for the mid-frequency band, and 0.2 for the high-frequency band. Error values ​​are back-propagated to update the hybrid denoising parameters in real time. The error convergence threshold is set to an example value of 0.05. When the error of each frequency band is lower than the error convergence threshold, the spectrum consistency is determined to meet the standard. It should also be noted that the error convergence threshold is set based on benchmark test data of a target noise-free signal in an anechoic chamber environment. First, 100 sample sets of noise-free signal samples were collected. Using the same multi-resolution analysis method, the inherent fluctuation range of each frequency band (low frequency 20-500Hz, mid-frequency 500-5kHz, and high frequency 5kHz-20kHz) was calculated. An example value of ±0.8dB for the inherent fluctuation range in the low frequency band corresponds to an error threshold of 0.04, a value of ±1.2dB for the mid-frequency band corresponds to 0.06, and a value of ±1.5dB for the high frequency band corresponds to 0.08. The final error convergence threshold of 0.05 is the weighted average of the thresholds for the three frequency bands, with the weighting factors consistent with the spectral consistency error calculation (0.5 for low frequency, 0.3 for mid-frequency, and 0.2 for high frequency). The error convergence threshold ensures that the spectral consistency meets the standard when the spectral fluctuation of the hybrid denoised signal falls within the 90% confidence interval of the example values ​​of the inherent fluctuation range of the noise-free signal. The threshold validation adopts the cross-validation method, and the misjudgment rate on the independent test set is controlled within 5% of the sample value.

[0054] S4.5. Dynamically adjust the reverse current injection amount and the reverse acoustic wave phase compensation parameters according to the spectrum consistency error to generate a hybrid noise reduction signal.

[0055] It should be noted that when dynamically adjusting the reverse current injection amount and reverse acoustic wave phase compensation parameters based on spectral consistency error, a proportional-integral control algorithm is used to update the parameters. The reverse current injection amount is adjusted proportionally to the low-frequency error. The proportional coefficient is taken as an example value of 0.8, the integral time constant is taken as an example value of 100ms, and the current adjustment step size is constrained to within the example value of ±5mA. The reverse acoustic wave phase compensation parameters are adjusted based on the weighted sum of the mid- and high-frequency errors. The mid-frequency weight is taken as an example value of 0.6, and the high-frequency weight is taken as an example value of 0.4. The phase compensation amount is calculated using a sliding average filter with an example window length of 10 frames. The parameter update cycle is synchronized with the spectrum analysis window, taking an example value of 25ms. The spectral consistency error of the hybrid noise reduction signal is recalculated after each adjustment. When the error value falls below the error convergence threshold (taken as an example value of 0.05), the current parameter combination is locked. The total harmonic distortion of the generated hybrid noise reduction signal is kept below the example value of 1%, and the signal-to-noise ratio is improved by 20dB (taken as an example value). The sampling rate of the hybrid noise reduction signal is maintained at 48kHz, and the correlation coefficient between the time domain waveform and the target noise-free signal reaches an example value of 0.95.

[0056] S5: Real-time monitoring of ear canal sealing parameters, dynamic adjustment of the phase offset and frequency response equalization parameters of the mixed noise reduction signal, and output of an optimized noise reduction signal adapted to the current wearing state.

[0057] S5.1. Real-time collection of air pressure fluctuation data inside the ear canal to generate ear canal sealing parameters, and calculation of the phase offset of the hybrid noise reduction signal using a nonlinear dynamic model.

[0058] It should be noted that when collecting the air pressure fluctuation data inside the ear canal in real time, a miniature air pressure sensor is used to obtain the pressure change in the sealed cavity at an example sampling rate of 1kHz, and the example value of the sensor sensitivity is 10mV / Pa. The ear canal sealing parameter is obtained by calculating the ratio of the air pressure fluctuation amplitude to the external sound pressure. The ratio range is 0.1-0.9, corresponding to different sealing levels. The nonlinear dynamic model uses the improved Van der Pol equation to describe the acoustic characteristics of the ear canal. The equation parameters are dynamically adjusted according to the sealing parameters. The stiffness coefficient is 100-500N / m, and the damping coefficient is 0.1-0.5Ns / m. The fourth-order Runge-Kutta method is used for the solution, and the step size is 0.1ms. The phase offset accuracy of the calculated output hybrid noise reduction signal reaches the example value of 0.01π, and the phase compensation response time is controlled within the example value of 5ms. The phase offset calculation result is updated synchronously with the real-time spectrum analysis data, and the update cycle is consistent with the hybrid noise reduction signal processing window; It should also be noted that the training of the nonlinear dynamic model (modified Van der Pol equation) uses 100 sets of measured ear canal pressure-phase data at different seal levels (0.1-0.9) as the training set. The nonlinear dynamic model parameter training process is divided into two steps: first, the damping coefficient is fixed to the example value of 0.3 Ns / m, and the stiffness coefficient is optimized using the least squares method to fit the pressure-phase relationship within the range of 100-500 N / m, with an iteration step size of 10 N / m. Then, the optimized stiffness coefficient is fixed, and the damping coefficient is adjusted using the gradient descent method (range 0.1-0.5 Ns / m), with a learning rate of 0.01 and a batch size of 10. The training objective is to minimize the root mean square value of the phase prediction error, and the termination criterion is that the error decreases by no more than 0.001π after 20 consecutive iterations. The final nonlinear dynamics model achieved a phase offset prediction error on the test set within the sampled value of 0.01π, and a response delay below the sampled value of 5ms, meeting real-time processing requirements. This was confirmed through cross-validation during the validation phase.

[0059] S5.2. Based on the ear canal sealing parameters and the phase offset, the frequency response equalization parameters of the hybrid noise reduction signal are calculated using the dynamic coupled resonant transfer function, and weighted correction is performed to generate optimized frequency response equalization parameters.

[0060] It should be noted that a second-order resonant model is used to construct the dynamic coupled resonant transfer function based on the ear canal seal parameter and phase offset. The resonant frequency varies with the ear canal seal parameter, with an example range of 100Hz-5kHz, and the quality factor (Q) value is 2-10. Frequency response equalization parameters are calculated by solving the amplitude response of the transfer function at the center frequency of a 1 / 3 octave band, with an example number of calculated frequency points of 24. A dual-parameter fusion strategy is used for weighting correction, with an example weight of 0.6 for the ear canal seal parameter and 0.4 for the phase offset parameter. The weighting coefficients are dynamically adjusted using a sigmoid function. The optimized frequency response equalization parameters are expressed in decibels, with an adjustment range within ±12dB of the example values. Parameter smoothing uses a three-point moving average, with the update cycle synchronized with the phase offset calculation. The resulting optimized frequency response equalization parameters have a frequency resolution of 1 / 3 octave, and a time domain response latency within 10ms.

[0061] S5.3. Based on the optimized frequency response equalization parameters, the phase offset is iteratively corrected through a multi-objective loss function to generate an optimized phase offset.

[0062] It should be noted that based on the optimized frequency response equalization parameters, the multi-objective loss function includes three items: phase consistency loss, frequency response flatness loss, and transient response loss, with weights assigned as examples: 0.5, 0.3, and 0.2, respectively. The conjugate gradient method is used for iterative correction of the phase offset, with an example iteration step size of 0.01π and a maximum number of iterations of 100. During each iteration, the phase consistency loss calculates the mean square error between the current phase offset and the target phase, the frequency response flatness loss evaluates the fluctuation within the 1 / 3 octave band, and the transient response loss is calculated by extracting the envelope signal using the Hilbert transform. The final convergence error of the optimized phase offset is controlled within the example value of 0.05π, the frequency response fluctuation range is reduced to ±3dB, and the transient response overshoot is reduced to less than 10% of the example value. The optimized phase offset update period is synchronized with the frequency response equalization parameters, and the timestamp alignment accuracy is maintained at the example value of 1ms.

[0063] S5.4. Superimpose the optimized phase offset and the optimized frequency response equalization parameters on the mixed noise reduction signal to generate an optimized noise reduction signal adapted to the current ear canal wearing state.

[0064] It should be noted that when the optimized phase offset and optimized frequency response equalization parameters are superimposed on the mixed noise reduction signal, frequency domain processing is used to achieve parameter fusion. First, a short-time Fourier transform (SFT) is performed on the mixed noise reduction signal, with an example window length of 1024 points and an example overlap ratio of 75%. The optimized phase offset is applied to the mixed noise reduction signal in the frequency domain, with the phase adjustment step size constrained to ±0.05π / frame (example value). The optimized frequency response equalization parameters are simultaneously applied for amplitude correction, with the correction amount smoothly transitioning over a 1 / 3 octave band, with an example transition bandwidth of 1 / 8 octave. The processed spectrum is reconstructed into the time domain signal using an inverse short-time Fourier transform (ISFT). Phase continuity is maintained during the reconstruction process, and group delay fluctuation is controlled within the example value of 0.1ms. The resulting optimized noise reduction signal, adapted to the current ear canal fit, maintains a sampling rate of 48kHz. The correlation coefficient between the time domain waveform and the target noise reduction characteristics reaches the example value of 0.98, and the frequency response flatness is improved to within ±2dB of the example value. The delay matching accuracy between the optimized noise reduction signal and the original noise signal reaches an example value of 0.05ms, achieving an effective active noise reduction effect.

[0065] S6. Parameter calibration of the optimized noise reduction signal is performed through anechoic chamber scene simulation and laser interferometer phase detection to generate a noise reduction control instruction set.

[0066] S6.1. Construct a three-dimensional laser interferometer array in an anechoic chamber to measure the vibration phase data of the diaphragm at the entrance of the ear canal and reconstruct and optimize the sound pressure field distribution of the noise reduction signal in the ear canal.

[0067] It should be explained that a three-dimensional laser interference array is arranged in the anechoic chamber, and 8 laser Doppler vibrometers are used to form a spatial measurement network. The example value of the spacing between the vibrometers is 50mm, and the example value of the sampling rate is 100kHz. When measuring the vibration phase data of the diaphragm at the entrance of the ear canal, the laser beam is focused on the measurement point on the surface of the diaphragm, the example value of the spot diameter is 0.1mm, and the phase resolution reaches an example value of 0.01π. The vibration data of each measuring point are fused through the three-dimensional point cloud reconstruction algorithm, and the reconstructed frequency range covers an example value of 20Hz-20kHz, and the spatial interpolation accuracy is an example value of 0.1mm. The sound pressure field distribution of the reconstructed optimized noise reduction signal in the ear canal is calculated using the boundary element method, with the diaphragm vibration phase data as the boundary condition, the calculation grid size is 1mm³, and the sound pressure field reconstruction frequency resolution is 1Hz. The final sound pressure field distribution obtained contains three-dimensional spatial sound pressure amplitude and phase information. The amplitude dynamic range is 40-120dB, and the phase consistency error is controlled within the example value of 0.05π, which effectively characterizes the propagation characteristics of the optimized noise reduction signal in the ear canal.

[0068] S6.2. Decompose the sound pressure field distribution into an amplitude spectrum and a phase spectrum, compare and analyze them with the preset target spectrum of the anechoic chamber background noise, and generate the amplitude error distribution and phase error distribution for each frequency band.

[0069] It should be explained that the sound pressure field distribution is decomposed into amplitude and phase spectra using a three-dimensional Fourier transform. The transform window length is 256×256×256 voxels, and the frequency resolution is 10 Hz. The amplitude spectrum is compared with the preset target spectrum of the anechoic chamber background noise floor in frequency bands. The frequency bands are divided into 1 / 3 octaves, and the center frequency range is 20 Hz to 20 kHz. The amplitude error distribution is calculated using the logarithmic scale difference method. The error value is expressed in decibels, and the calculation accuracy is 0.1 dB. The phase error distribution is calculated using the cyclic angle difference method. The error range is controlled within ±π of the example value, and the accuracy reaches 0.01π of the example value. The error analysis results are statistically analyzed by frequency band. The error weight coefficient for the low-frequency band (for example, 20-200 Hz) is 0.6, the weight coefficient for the mid-frequency band (for example, 200-5 kHz) is 0.3, and the weight coefficient for the high-frequency band (for example, 5 kHz to 20 kHz) is 0.1. The spatial resolution of the generated amplitude error distribution and phase error distribution remains consistent with the original sound pressure field distribution, with an example value of 1 mm³. The time domain stability error is controlled within 5% of the example value.

[0070] S6.3. Construct a nonlinear transfer function mapping from amplitude error distribution and phase error distribution to frequency response equalization parameters and phase offset, and use a quantum optimization algorithm to solve the optimal parameter correction that satisfies the error constraints.

[0071] It should be noted that when constructing the nonlinear transfer function mapping from the amplitude and phase error distributions to the frequency response equalization parameters and phase offset, a radial basis function network was used to establish the input-output relationship. The network contained 50 hidden nodes (as an example), and the Gaussian kernel width was 0.5. The quantum optimization algorithm employed quantum annealing to solve for parameter corrections. A Hamiltonian model containing 100 qubits (as an example) was constructed. The amplitude error constraint was converted into a potential energy term with a weight of 0.7 (as an example), and the phase error constraint weight was 0.3 (as an example). During the optimization process, the quantum tunneling effect intensity was 0.1 (as an example), the annealing time was 100 μs (as an example), and an exponential decay curve was used for temperature scheduling. The optimal parameter corrections obtained were: the frequency response equalization parameter correction range was controlled within the example ±6 dB, with a step size accuracy of 0.1 dB; the phase offset correction range was ±0.2π (as an example), with a step size accuracy of 0.01π (as an example). The optimization results meet the constraints that the RMS value of the amplitude error distribution is less than 1 dB from the example value, and the RMS value of the phase error distribution is less than 0.05π from the example value. The parameter update cycle is synchronized with the error analysis.

[0072] S6.4. Superimpose the optimal correction parameters on the optimized noise reduction signal to generate a calibration signal. Use a laser interferometer to measure the vibration phase data of the calibration signal for a second time. If the error exceeds the limit, repeat the iterative correction until convergence.

[0073] It should be noted that when the optimal correction parameters are superimposed on the optimized noise reduction signal to generate the calibration signal, the frequency response equalization parameters are implemented using a digital filter bank. The filter bank contains multiple 1 / 3 octave bandpass filters, each with a center frequency distributed according to a standard 1 / 3 octave distribution, and the passband fluctuation is controlled within a small range. The phase offset is implemented using an all-pass filter, and the group delay fluctuation is constrained within a strict range. The calibration signal is measured using a three-dimensional laser interferometer array to measure the vibration phase data of the diaphragm at the entrance of the ear canal. The laser beam focus spot diameter is kept small, achieving a high level of phase measurement accuracy. The amplitude error root mean square value threshold and the phase error root mean square value threshold are each set to specific values. When the error exceeds the limit, the amplitude error distribution and phase error distribution are recalculated, and the constraints of the quantum optimization algorithm are updated before the parameter correction is solved again. During the iterative process, the correction step size is attenuated according to a specific sequence. The initial step sizes of the amplitude correction and phase correction are each set to fixed values, and the attenuation ratio is kept constant. The convergence condition is that the amplitude error root mean square value and the phase error root mean square value of multiple consecutive iterations are lower than the amplitude error root mean square value threshold and the phase error root mean square value threshold, and the correlation coefficient between the calibration signal and the target spectrum reaches a high level. It should also be noted that the setting process of the amplitude error RMS threshold is based on the background noise characteristics of the anechoic chamber. By measuring the minimum discernible difference in each 1 / 3 octave band and combining it with the psychoacoustic equal loudness curve, a weighted calculation is performed to ensure that the human ear cannot perceive the amplitude distortion. The process of setting the phase error RMS threshold is as follows: Based on the phase response limit of the ear canal diaphragm under transient sound stimulation and the phase integration characteristics of the critical hearing bandwidth, the phase tolerance value of each frequency band is calculated to ensure that no audible phase distortion effect is generated.

[0074] S6.5. Encode the converged frequency response equalization parameters and phase offset into a binary instruction sequence with a priority mark, insert dynamic compensation instructions for the key frequency band, and generate a noise reduction control instruction set.

[0075] It should be noted that when encoding the converged frequency response equalization parameters and phase offset into a binary instruction sequence with priority markings, the frequency response equalization parameters are represented using 16-bit fixed-point numbers, with an example quantization step size of 0.01dB. The priority markings are divided into four levels (example values ​​00-11 correspond to priorities 1-4). The phase offset is encoded in 12-bit two's complement format, with an example quantization step size of 0.001π. The insertion position of the key frequency band dynamic compensation instructions is determined based on the error analysis results. The key frequency band determination criteria are: the frequency band where the historical maximum amplitude error exceeds the example value of 3dB or the phase error exceeds the example value of 0.1π. The dynamic compensation instructions use a 32-bit format, with the first 16 bits encoding the band center frequency (example value: 20Hz-20kHz logarithmically distributed), and the last 16 bits containing the compensation amount and refresh period parameters. The total length of the noise reduction control instruction set is 1024 bytes, the instruction execution interval is 10ms, and the refresh period of the key frequency band instructions is shortened to 2ms. The instruction set uses a cyclic redundancy check (CRC) code, with the checksum polynomial taking the example value of 0x1021, to ensure transmission reliability. The resulting noise reduction control instruction set contains 256 executable instructions with example values, supporting real-time dynamic updates of key frequency band parameters.

[0076] This embodiment also provides a noise reduction system for a high-performance TWS Bluetooth audio chip, including: a noise acquisition module for collecting environmental noise signals and simultaneously obtaining Bluetooth radio frequency interference signals, converting them into digital signals using a 24-bit analog-to-digital converter, and outputting a multimodal noise feature matrix; The noise separation module is used to perform noise separation and reverse sound wave generation on the multimodal noise feature matrix using the pre-trained lightweight CycleGAN model, and output the reverse sound wave signal; A noise reduction optimization module is used to offset the power domain noise by injecting a reverse current into the reverse acoustic wave signal to generate a hybrid noise reduction signal; The calibration control module is used to monitor the ear canal sealing parameters in real time, dynamically adjust the phase offset and frequency response equalization parameters of the hybrid noise reduction signal, and output an optimized noise reduction signal adapted to the current wearing state; The power supply noise reduction module is used to calibrate the parameters of the optimized noise reduction signal through anechoic chamber scene simulation and laser interferometer phase detection, and generate a noise reduction control instruction set.

[0077] This embodiment also provides a computer device, which is suitable for the noise reduction method for a high-performance TWS Bluetooth audio chip, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the noise reduction method for a high-performance TWS Bluetooth audio chip proposed in the above embodiment.

[0078] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0079] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the noise reduction method for a high-performance TWS Bluetooth audio chip proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0080] In summary, the present invention achieves high-precision separation and reverse sound wave generation of multimodal noise through a pre-trained lightweight CycleGAN model, solving the problem of insufficient noise decoupling capability of traditional methods; at the same time, it combines real-time ear canal seal monitoring and dynamic parameter adjustment to ensure that the noise reduction signal accurately matches the wearing state, overcoming the adaptability defects of fixed parameter noise reduction. The entire solution achieves professional-grade noise reduction performance under low power consumption conditions through the coordinated cooperation of technologies such as multimodal noise acquisition, power supply noise cancellation, and laser interferometer calibration, effectively solving the noise interference problem of TWS headphones in complex environments.

[0081] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A noise reduction method for a high-performance TWS Bluetooth audio chip, characterized by: include, Collect environmental noise signals and simultaneously obtain Bluetooth radio frequency interference signals, convert them into digital signals using a 24-bit analog-to-digital converter, and output a multimodal noise feature matrix; Use the pre-trained lightweight CycleGAN model to perform noise separation and reverse sound wave generation on the multimodal noise feature matrix, and output the reverse sound wave signal; Injecting reverse current into the reverse acoustic wave signal to cancel the power domain noise and generate a hybrid noise reduction signal; Real-time monitoring of ear canal sealing parameters, dynamic adjustment of the phase offset and frequency response equalization parameters of the hybrid noise reduction signal, and output of an optimized noise reduction signal adapted to the current wearing state; The optimized noise reduction signal is calibrated through anechoic chamber scene simulation and laser interferometer phase detection to generate a noise reduction control instruction set.

2. The noise reduction method for a high-performance TWS Bluetooth audio chip according to claim 1, characterized in that: The environmental noise signals include industrial noise, construction noise, transportation noise and social life noise.

3. The noise reduction method for a high-performance TWS Bluetooth audio chip according to claim 1, characterized in that: The output multimodal noise feature matrix is ​​specifically constructed as follows: Bluetooth RF interference signals are synchronously collected through a shared clock source, and RF coupling and anti-aliasing filtering are performed, while frequency band limiting filtering is performed on ambient noise signals. The filtered ambient noise signal and Bluetooth RF interference signal are input into the independent differential channels of a 24-bit analog-to-digital converter for synchronous quantization and noise shaping to obtain a digitized signal with time-stamp alignment. Based on the digitized signal with time stamp alignment, the acoustic component features are extracted by joint time-frequency analysis of the ambient noise signal. The electromagnetic component features are extracted by spectrum pulse detection of the Bluetooth radio frequency interference signal. Cross-modal analysis is then performed using a time series correlation algorithm to generate cross-modal features. The acoustic component features, electromagnetic component features and cross-modal features are fused to generate a multimodal noise feature matrix.

4. The noise reduction method for a high-performance TWS Bluetooth audio chip according to claim 1, wherein: The specific steps of outputting the reverse sound wave signal are as follows: Domain adversarial training is performed on the multimodal noise feature matrix. The distribution differences of the acoustic and electromagnetic component features are aligned through the gradient reversal layer to generate domain-invariant feature representations. The temporal context dependencies are extracted using a sliding time window, and the context-enhanced feature matrix is ​​output. The context-enhanced feature matrix is ​​input into the pre-trained lightweight CycleGAN model, which separates the acoustic noise and electromagnetic noise components through cross-modal adversarial training and generates the initial reverse acoustic wave signal based on a multi-scale discriminator. The initial reverse acoustic wave signal is constrained for complex spectrum consistency, and the phase offset is corrected using a dynamic time warping algorithm to output a preliminary reverse acoustic wave signal aligned in the time domain. The time-varying noise reduction coefficient and phase compensation parameters are extracted from the time-domain aligned preliminary reverse acoustic wave signal, and dynamically adjusted through reinforcement learning strategy to generate the reverse acoustic wave signal.

5. The noise reduction method for a high-performance TWS Bluetooth audio chip according to claim 1, characterized in that: The specific steps of generating the mixed noise reduction signal are as follows: Real-time acquisition of power circuit ripple signals, extraction of switching frequency harmonic components and broadband noise floor characteristics through time-frequency analysis, and the acquisition of dynamic noise spectrum characteristics; The time-varying noise reduction coefficient and phase compensation parameters are combined with the dynamic noise spectrum characteristics to generate the reverse current injection amount and phase synchronization parameters. The time delay error of the reverse acoustic wave signal transmission path is calculated through the cross-correlation function, and a pre-distortion algorithm is used to generate the compensated reverse current signal. The reverse acoustic wave signal and the compensated reverse current signal are jointly aligned in time and frequency domains to generate a preliminary mixed noise reduction signal; The preliminary mixed denoised signal is monitored in real time and its time-frequency characteristics are extracted. The spectral consistency error with the target noise-free signal is calculated through a multi-scale discriminator. The reverse current injection amount and reverse acoustic wave phase compensation parameters are dynamically adjusted according to the spectrum consistency error to generate a hybrid noise reduction signal.

6. The noise reduction method for a high-performance TWS Bluetooth audio chip according to claim 1, wherein: The output is adapted to the current wearing state of the optimized noise reduction signal, the specific steps are as follows: Real-time collection of air pressure fluctuation data inside the ear canal to generate ear canal sealing parameters, and the phase offset of the hybrid noise reduction signal is calculated through a nonlinear dynamic model; According to the ear canal sealing parameters and phase offset, the frequency response equalization parameters of the hybrid noise reduction signal are calculated using the dynamic coupled resonance transfer function, and weighted correction is performed to generate the optimized frequency response equalization parameters; Based on the optimized frequency response equalization parameters, the phase offset is iteratively corrected through a multi-objective loss function to generate an optimized phase offset; The optimized phase offset and the optimized frequency response equalization parameters are superimposed on the mixed noise reduction signal to generate an optimized noise reduction signal adapted to the current ear canal wearing state.

7. The noise reduction method for a high-performance TWS Bluetooth audio chip according to claim 1, characterized in that: The specific steps of generating the noise reduction control instruction set are as follows: A three-dimensional laser interferometer array was constructed in the anechoic chamber to measure the vibration phase data of the diaphragm at the entrance of the ear canal and reconstruct the sound pressure field distribution of the optimized noise reduction signal in the ear canal. Decompose the sound pressure field distribution into amplitude spectrum and phase spectrum, compare and analyze with the preset target spectrum of the anechoic chamber background noise, and generate the amplitude error distribution and phase error distribution of each frequency band; Construct a nonlinear transfer function mapping from amplitude error distribution and phase error distribution to frequency response equalization parameters and phase offset, and use quantum optimization algorithm to solve the optimal parameter correction that meets the error constraints; The optimal correction parameters are superimposed on the optimized noise reduction signal to generate a calibration signal. The vibration phase data of the calibration signal is measured again using a laser interferometer. If the error exceeds the limit, it is iteratively corrected until convergence. The converged frequency response equalization parameters and phase offset are encoded into a binary instruction sequence with priority marks, dynamic compensation instructions are inserted into key frequency bands, and a noise reduction control instruction set is generated.

8. A noise reduction system for a high-performance TWS Bluetooth audio chip, based on the noise reduction method for a high-performance TWS Bluetooth audio chip according to any one of claims 1 to 7, characterized in that: include, The noise acquisition module is used to collect environmental noise signals and simultaneously obtain Bluetooth radio frequency interference signals, convert them into digital signals using a 24-bit analog-to-digital converter, and output a multimodal noise feature matrix; The noise separation module is used to perform noise separation and reverse sound wave generation on the multimodal noise feature matrix using the pre-trained lightweight CycleGAN model, and output the reverse sound wave signal; A noise reduction optimization module is used to offset the power domain noise by injecting a reverse current into the reverse acoustic wave signal to generate a hybrid noise reduction signal; The calibration control module is used to monitor the ear canal sealing parameters in real time, dynamically adjust the phase offset and frequency response equalization parameters of the hybrid noise reduction signal, and output an optimized noise reduction signal adapted to the current wearing state; The power supply noise reduction module is used to calibrate the parameters of the optimized noise reduction signal through anechoic chamber scene simulation and laser interferometer phase detection, and generate a noise reduction control instruction set.

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 steps of the noise reduction method for a high-performance TWS Bluetooth audio chip according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of the noise reduction method for a high-performance TWS Bluetooth audio chip according to any one of claims 1 to 7 are implemented.

Images