OWS Bluetooth earphone noise reduction method and system based on artificial intelligence

Abstract

The invention belongs to the technical field of OWS Bluetooth headset noise reduction, and provides an OWS Bluetooth headset noise reduction method and system based on artificial intelligence, and the method comprises the steps: obtaining audio, visual and motion data through a microphone array, a camera and a six-axis IMU, carrying out the preprocessing, extracting a fusion feature through a 1D-CNN and an attention mechanism, and generating a multi-dimensional noise feature vector; noise prediction and adaptive noise reduction are carried out, historical features are analyzed based on a time sequence prediction network, future 100ms noise is predicted, an anti-phase filter is generated according to frequency bands, and parameters are adjusted in real time according to errors; according to the method, low-power-consumption optimization is achieved, SNN is used for replacing part of models, event-driven calculation is achieved in combination with a memristor, redundancy operation is reduced based on noise sparsity, dynamic noise quick response, multi-mode accurate sensing and low-power-consumption operation are achieved, and the noise reduction effect and endurance of the OWS earphone are improved.

Description

Technical Field

[0001] This invention belongs to the field of Bluetooth headset noise reduction technology, specifically the OWS Bluetooth headset noise reduction method and system based on artificial intelligence. Background Technology

[0002] OWS Bluetooth earphones, as an emerging audio device, have gained widespread attention in scenarios such as sports, commuting, and office work due to their features such as not blocking the ear canal, comfortable wearing, and natural perception of ambient sound.

[0003] However, its open design also leads to natural noise reduction challenges: lacking an ear canal closure structure, it cannot use passive sound insulation to reduce ambient noise and mainly relies on active noise cancellation (ANC) technology to suppress noise.

[0004] The existing OWS Bluetooth headphone noise cancellation technology mainly has the following problems: Response delay to dynamic noise: Existing algorithms typically employ a serial processing flow of detection-analysis-response. For sudden dynamic noises, such as car horns or braking sounds, the response time can be as long as 200ms or more, causing users to still perceive a significant noise impact the moment the noise appears.

[0005] Limitations of single-modal perception: It relies solely on microphones to collect audio signals, lacks multi-dimensional feature extraction of noise sources, and is difficult to accurately distinguish different types of noise and their changing trends.

[0006] The conflict between computing power and power consumption on the device side: OWS headphones usually use low-power chips, while complex noise reduction algorithms require high computing resources, making it difficult to balance noise reduction effect and battery life.

[0007] Therefore, how to achieve rapid response and efficient suppression of dynamic noise while taking into account low power consumption in the open-back design of OWS headphones is a technical problem that urgently needs to be solved.

[0008] Therefore, this invention provides an OWS Bluetooth headset noise reduction method and system based on artificial intelligence. Summary of the Invention

[0009] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0010] The technical solution adopted by this invention to solve its technical problem is: In a first aspect, the present invention provides an OWS Bluetooth headset noise reduction method based on artificial intelligence, comprising: S1: Construct a microphone array to collect audio signals, use a camera, gyroscope, and accelerometer to acquire environmental visual information and user motion state information, preprocess the collected multimodal data, and extract and fuse features from the preprocessed multimodal data based on a deep learning model to generate a multidimensional noise feature vector; S2: Based on the temporal prediction network, analyze the historical noise feature vectors to predict the noise prediction results within the next 100ms. Based on the noise prediction results, adjust the adaptive filter parameters in advance to generate a noise reduction signal. Monitor the deviation between the actual noise signal and the noise prediction results in real time and dynamically adjust the noise reduction parameters. S3: Replace the one-dimensional convolutional neural network and attention mechanism model in the feature fusion and extraction stage of S1 with a spiking neural network (SNN). Combine it with a memristor in-memory computing unit to achieve low-power computing. Based on the sparsity characteristics of the noise spectrum, redundant operations are reduced. A sound pressure change triggering mechanism is adopted to dynamically switch between energy-saving and enhancement modes according to the noise intensity.

[0011] As a further improvement of the present invention, the process of constructing a microphone array to collect audio signals is as follows: An 8-microphone array was constructed, with the following arrangement: 3 feedforward microphones, 3 feedback microphones, and 2 bone conduction microphones were placed on the audio channel; the spacing between the feedforward and feedback microphones was 15mm to form spatial filtering gain; the bone conduction microphones were placed close to the temporal bone to collect skull vibration signals; and the audio signals were obtained by collecting the feedforward, feedback, and bone conduction microphones.

[0012] As a further improvement of the present invention, the specific process of acquiring environmental visual information and user motion state information using a camera, gyroscope, and accelerometer is as follows: A miniature CMOS camera is deployed with a lens tilted at 30° to cover a 120° field of view in front, used to acquire environmental images; A six-axis IMU is integrated into the earphone hinge to acquire user motion data. The six-axis IMU integrates a gyroscope and an accelerometer, and its sampling frequency is 100Hz, which can capture key frequency changes in human motion.

[0013] As a further improvement of the present invention, the specific process of analyzing historical noise feature vectors based on a time-series prediction network and predicting noise prediction results within the next 100ms is as follows: Temporal feature sequence construction and preprocessing: sliding window acquisition: with a window length of 500ms and a step size of 10ms, multidimensional noise feature vectors are obtained to form a temporal sequence containing 50 frames of features, covering the duration of dynamic noise; Sequence enhancement calculates the first-order difference of the feature sequence to highlight the dynamic trend of noise changes. It normalizes the features of each dimension through Z-score and adopts a lightweight temporal prediction network strategy. It uses a one-dimensional convolutional neural network (1D-CNN) to extract local spectral features and uses the attention mechanism of Transformer to capture long-distance temporal dependencies. After analyzing 50 frames of feature sequences, it outputs the noise prediction results for the next 100ms.

[0014] As a further improvement of the present invention, the specific process of adjusting the adaptive filter parameters in advance based on the noise prediction results to generate the noise-reduced signal is as follows: A frequency band prediction strategy is adopted, with fine-grained prediction used for high-frequency noise to match sharp impact noise, and conventional prediction used for mid- and low-frequency noise. A 3-channel FIR filter bank is constructed, with the audio signal frequency of 20-500Hz as the low-frequency channel, the audio signal frequency of 500-2000Hz as the mid-frequency channel, and the audio signal frequency of 2000-20kHz as the high-frequency channel. Based on the noise spectrum predicted by the noise prediction results, the filter coefficients are adjusted in advance: Low-frequency channel: suppresses slowly varying noises such as engine roar; High-frequency channel: responds quickly to sudden noises such as horn blasts. If the energy envelope shows a rapid change in noise energy, the high-frequency channel filter will quickly adjust its coefficients. If the energy envelope shows a slow change in noise energy, the low-frequency channel filter will smoothly adjust its coefficients. Based on the spectrum prediction map of the noise prediction result, i.e. the energy distribution of noise in the 20Hz-20kHz frequency band, filter coefficients that can cancel the noise prediction result are generated in reverse phase, ensuring that the noise-reduced signal and the noise prediction result have opposite acoustic phase and similar amplitude. The DSP's SIMD instructions are used to process the 3-channel audio signal in parallel. The original audio is convolved with the pre-configured filter coefficients to generate an inverse noise-reduced signal. The processing delay of the low-frequency channel, mid-frequency channel, and high-frequency channel is controlled within 2ms. The weights of each channel are allocated according to the energy envelope, and the signals are superimposed to generate the final noise-reduced signal.

[0015] As a further improvement of the present invention, the specific process of real-time monitoring of the deviation between the actual noise signal and the noise prediction result, and dynamically adjusting the noise reduction parameters is as follows: The residual noise in the ear canal is collected in real time by a feedback microphone, the prediction error is calculated, the filter coefficients are updated by a variable step size NLMS algorithm, the step size is dynamically adjusted with the error, the attention weights of the prediction network are fine-tuned based on the error trend, and when the error rate PER>30%, the pure feedback noise reduction mode is switched to and the prediction function is disabled.

[0016] As a further improvement of the present invention, the specific process of achieving low-power computing is as follows: The one-dimensional convolutional neural network and attention mechanism model are transformed into an event-driven spiking neural network (SNN) to process multimodal data using neuronal spiking. The computation is activated only when the noise characteristics change, reducing power consumption in static scenes. The parameters connecting the neurons in the spiking neural network are stored in a memristor array, and vector multiplication is performed directly in memory. Each processing unit, including the audio processing unit, the visual processing unit, and the motion processing unit, operates independently and cooperates through pulse events to achieve low-power computation.

[0017] As a further improvement of the present invention, the specific process of reducing redundant calculations is as follows: The system extracts 50 frames of feature sequences using the attention mechanism of 1D-CNN and Transformer. The L1 norm of each frame feature is calculated. When the sparsity is >70%, i.e. a quiet scene, the sparse computation mode is activated. A sparse mask is generated by a spiking neural network to retain only the frequency points of energy abrupt changes. The sparse convolution is performed only on non-zero feature points to reduce redundant computation.

[0018] As a further improvement of the present invention, the specific process of dynamically switching between energy-saving and enhancement modes based on noise intensity is as follows: The mode is dynamically switched based on the noise intensity. Energy-saving mode: In the current static scene, the signal-to-noise ratio is less than 10dB, and only 10% of neurons are retained to achieve low-power operation; Enhanced Mode: For sudden strong noise, full impulse calculation is used to enhance the noise reduction effect.

[0019] Secondly, the present invention provides an AI-based OWS Bluetooth headset noise cancellation system, comprising: Multimodal data acquisition and feature fusion module: Constructs a microphone array to acquire audio signals, uses cameras, gyroscopes, and accelerometers to acquire environmental visual information and user motion state information, preprocesses the acquired multimodal data, and extracts and fuses features from the preprocessed multimodal data based on a deep learning model to generate a multidimensional noise feature vector; Noise prediction and adaptive noise reduction module: Based on the temporal prediction network, the module analyzes the historical noise feature vectors and predicts the noise prediction results within the next 100ms. Based on the noise prediction results, the module adjusts the adaptive filter parameters in advance, generates a noise reduction signal, and monitors the deviation between the actual noise signal and the noise prediction results in real time, dynamically adjusting the noise reduction parameters. Low-power computing and mode switching module: The one-dimensional convolutional neural network and attention mechanism model in the feature fusion and extraction stage of the multimodal data acquisition and feature fusion module are replaced by a spiking neural network (SNN). Combined with a memristor in-memory computing unit, low-power computing is achieved. Based on the sparsity characteristics of the noise spectrum, redundant operations are reduced. A sound pressure change triggering mechanism is adopted to dynamically switch between energy-saving and enhancement modes according to the noise intensity.

[0020] The beneficial effects of this invention are as follows: By predicting noise 100ms in advance using a time-series prediction network and generating an inverting filter by frequency band division, active pre-response noise reduction is achieved. This breaks through the delay bottleneck of the traditional detection-analysis-response serial process, significantly improving the response speed to sudden dynamic noises such as horns and brake sounds, and reducing the perceived impact of noise.

[0021] By integrating audio, visual, and motion multimodal data and dynamically allocating modal weights through an attention mechanism, combined with a bone conduction backup mechanism, the accuracy of noise type differentiation and feature extraction in complex environments is improved, avoiding the limitations of single audio perception.

[0022] By replacing the traditional model with a spiking neural network (SNN) and combining it with a memristor in-memory computing unit to achieve event-driven computation (activating only when noise changes), the sparsity of the noise spectrum reduces redundant computation, and dynamically switches between power-saving and enhancement modes. This reduces power consumption while maintaining noise reduction performance, making it compatible with low-computing-power chips in OWS headphones and extending battery life. In summary, this invention achieves rapid response to dynamic noise, accurate multimodal perception, and low-power operation, significantly improving the noise cancellation performance and user experience of OWS Bluetooth headsets. Attached Figure Description

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart of the steps of the OWS Bluetooth headset noise reduction method based on artificial intelligence in this invention; Figure 2 This is a system module diagram of the OWS Bluetooth headset noise cancellation system based on artificial intelligence, which is the subject of this invention. Detailed Implementation

[0025] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments. Example

[0026] like Figure 1 As shown in the embodiment of the present invention, the OWS Bluetooth headset noise reduction method based on artificial intelligence includes: S1: Construct a microphone array to collect audio signals, use a camera, gyroscope, and accelerometer to acquire environmental visual information and user motion state information, preprocess the collected multimodal data, and extract and fuse features from the preprocessed multimodal data based on a deep learning model to generate a multidimensional noise feature vector; The first specific step is to construct a microphone array to collect audio signals as follows: An 8-microphone array was constructed, with the following arrangement: 3 feedforward microphones, 3 feedback microphones, and 2 bone conduction microphones were placed on the audio channel; the spacing between the feedforward and feedback microphones was 15mm to form spatial filtering gain; the bone conduction microphones were placed close to the temporal bone to collect skull vibration signals; and the audio signals were obtained by collecting the feedforward, feedback, and bone conduction microphones. Secondly, the specific process of acquiring environmental visual information and user motion state information using cameras, gyroscopes, and accelerometers is as follows: A miniature CMOS camera is deployed with a lens tilted at 30° to cover a 120° field of view in front, used to acquire environmental images; A six-axis IMU is integrated into the earphone hinge to acquire user motion data. The six-axis IMU integrates a gyroscope and an accelerometer, and its sampling frequency is 100Hz, which can capture key frequency changes in human motion. Preprocessing of multimodal data collected by audio, vision, and motion sensors improves data quality and lays the foundation for subsequent processing; Thirdly, the specific process for preprocessing the collected multimodal data is as follows: The hardware trigger signal triggers multiple types of sensors to collect data synchronously with an error of <1ms. The timestamps are used to align the multimodal data, which includes audio signals (48kHz), environmental images (10fps), and motion data (100Hz), generating a spatiotemporally aligned data stream. It should be noted that the hardware trigger signal is a dedicated hardware trigger signal set inside the headset. When the hardware trigger signal is generated, the system simultaneously commands multiple sensors, such as the microphone, camera, and six-axis IMU sensor, to immediately start collecting data. The specific process of audio signal preprocessing is as follows: first, an 8th-order Butterworth filter is used to filter out audio signals in the range inaudible to the human ear, and then the transfer function is used... The audio signal within the range of human hearing is obtained, where, The cutoff angular frequency is the audio signal within the range of human hearing. For composite frequency variables, Let the order be the filter order. Let be the damping coefficient of the k-th second-order element, where k = 1, 2, 3, 4; The filter transfer function outputs an audio signal within the audible range of the human ear. The audible audio signal obtained by processing with an 8th-order Butterworth filter is then subjected to spectral subtraction to suppress steady-state noise and extract non-steady-state noise, facilitating subsequent analysis. The steady-state noise is suppressed using spectral subtraction, and the specific formula is as follows: ,in, For over-subtraction factor, The power spectral density of the noisy signal spectrum. The power spectral density of the noise spectrum. The power spectral density of the signal after noise reduction. This indicates taking the maximum value; the signal after steady-state noise suppression can be obtained using the above spectral subtraction formula. The signal after steady-state noise suppression includes the target signal and non-steady-state noise. For the original noisy signal and the signal after steady-state noise suppression Perform subtraction to obtain non-steady-state noise. ; Specifically, Combining the above formula The formula is obtained as follows: Approximately ; , If non-steady-state noise dominates, then non-steady-state noise should be extracted. Steady-state noise is characterized by a stable spectrum and slow changes over time. Spectral subtraction utilizes this characteristic to suppress steady-state noise by first estimating the noise spectrum and then subtracting it from the noisy signal. The specific process of environmental image preprocessing is as follows: CLAHE utilizes contrast-limited adaptive histogram equalization (CLAHE) to enhance image details and avoid noise amplification by adjusting the histogram of local regions. The specific process of preprocessing motion data involves fusing gyroscope and accelerometer data using the following formula: ,in, For attitude angle, Let be the angular velocity at time t. , For acceleration components, The fusion coefficient is... Time variable, For at any time attitude angle, For time intervals; Fourthly, specifically, feature extraction and fusion are performed on the preprocessed multimodal data based on deep learning models to generate multidimensional noise feature vectors; Mel-ray spectral features are extracted using a one-dimensional convolutional neural network (1D-CNN). The convolution operation formula is as follows: ,in, For the first Layer One convolutional kernel, This represents the convolution operation. For activation function, This is the bias term corresponding to the j-th output of the l-th layer. Let be the i-th input feature vector of the (l-1)-th layer, and be the number of feature vectors. Calculate modal weights using an attention mechanism: ,in, , For the first Modal features, It is a 2-layer perceptron. For layer normalization, For the modal weights, To Perform exponential function operations. For the i-th modal feature The result obtained after calculation Represents all modalities; The concatenated 448-dimensional features are compressed to 128 dimensions using a 1×1 convolution, generating a multidimensional noise feature vector. The formula is as follows: ,in, The weight matrix is ​​a 1×1 convolution. The spliced ​​448-dimensional features Indicates the bias term. This is the compressed 128-dimensional noise feature vector; The generated multidimensional noise feature vector is subjected to audio confidence calculation, and the reliability of the audio features is determined by calculating the short-time signal-to-noise ratio (SNR). In complex scenarios, OWS earphones often suffer from audio signal distortion due to strong noise interference. If the distorted audio features are used directly for predictive noise reduction, the noise reduction model will misjudge the noise type and increase the prediction error. By calculating the SNR to determine the audio confidence level, this unreliable audio feature can be identified, triggering bone conduction backup and ensuring the validity of the input audio feature. Audio confidence is used to calculate the short-time signal-to-noise ratio (SNR): ; This is the total energy obtained by summing the squares of the spectral amplitudes of the target signal at different frequencies f and time t. The total energy is the sum of the squares of the spectral amplitudes of the noise signal at different frequencies f and time t. when The time marker is set to low confidence, meaning the generated multidimensional noise feature vector is severely contaminated and constitutes unreliable audio features. At this point, adaptive modal backup is triggered, switching to the bone conduction voice enhancement signal. ,in, Based on bone conduction characteristics, the human voice frequency band is enhanced through bandpass filtering.

[0027] Mathematical modeling enables efficient processing and feature fusion of multimodal data, providing accurate input for subsequent predictive noise reduction.

[0028] S2: Based on the temporal prediction network, analyze the historical noise feature vectors to predict the noise prediction results within the next 100ms. Based on the noise prediction results, adjust the adaptive filter parameters in advance to generate a noise reduction signal. Monitor the deviation between the actual noise signal and the noise prediction results in real time and dynamically adjust the noise reduction parameters. Firstly, specifically, the process of analyzing historical noise feature vectors based on a time-series prediction network to predict noise results within the next 100ms is as follows: Temporal feature sequence construction and preprocessing: sliding window acquisition: with a window length of 500ms and a step size of 10ms, multidimensional noise feature vectors are obtained to form a temporal sequence containing 50 frames of features, covering the duration of dynamic noise; Sequence augmentation, calculating the first difference of the feature sequence , Let be the eigenvalues ​​of the characteristic sequence at time t. The feature values ​​of the feature sequence at time t-1 are used to highlight the dynamic change trend of noise and normalize the features of each dimension using Z-score. A lightweight temporal prediction network strategy is adopted, utilizing a one-dimensional convolutional neural network (1D-CNN) to extract local spectral features. The attention mechanism of the Transformer is used to capture long-range temporal dependencies. After analyzing 50 frames of feature sequences, the noise prediction results for the next 100ms are output. The noise prediction results include: Spectrum prediction diagram: i.e., the energy distribution in the 20Hz-20kHz frequency band; Energy envelope: Instantaneous energy change curve with a resolution of 10 ms; Based on the noise prediction results, the specific process of adjusting the adaptive filter parameters in advance to generate the noise-reduced signal is as follows: Secondly, a frequency band prediction strategy is adopted, with fine-grained prediction used for high-frequency noise (5-20kHz) to match sharp impact noise, and conventional prediction used for mid-to-low frequency noise (20Hz-5kHz) to improve computational efficiency. A 3-channel FIR filter bank is constructed: the low-frequency channel is for audio signals with frequencies between 20 and 500 Hz, the mid-frequency channel is for audio signals with frequencies between 500 and 2000 Hz, and the high-frequency channel is for audio signals with frequencies between 2000 and 20 kHz. The filter coefficients are adjusted in advance based on the noise spectrum of the noise prediction results: the low-frequency channel suppresses slowly varying noises such as engine roar; the high-frequency channel responds quickly to sudden noises such as horn blasts. If the energy envelope shows a rapid change in noise energy, such as the instantaneous burst of a horn, the high-frequency channel filter will more aggressively adjust the coefficients, for example, by shortening the response time. If the energy envelope shows that the noise energy changes slowly, such as the continuous and stable roar of an engine, the low-frequency channel filter will adjust the coefficient more smoothly to avoid over-cancellation. Based on the frequency spectrum prediction diagram of the noise prediction results, i.e., the energy distribution of noise in the 20Hz-20kHz frequency band, filter coefficients that can cancel out the noise prediction results are generated in reverse phase, ensuring that the noise-reduced signal has opposite acoustic phase and similar amplitude to the noise prediction results. The specific formula is as follows: ,in, For noise prediction results in the next 100ms, For Fourier transform, For inverse Fourier transform, These are the filter coefficients; The DSP's SIMD instructions are used to process the 3-channel audio signal in parallel. The original audio is convolved with the pre-configured filter coefficients to generate an inverse noise reduction signal. The processing delay of the low-frequency channel, mid-frequency channel and high-frequency channel is controlled within 2ms. The weights of each channel are assigned according to the energy envelope, and the final noise-reduced signal is generated by superimposing them and then mixed with the original audio before being output.

[0029] Traditional noise reduction involves first collecting noise and then canceling it, which is passive noise reduction. In contrast, this invention predicts noise and generates an inverse filter to prepare the noise reduction signal in advance, which is active pre-response noise reduction and optimizes the noise reduction delay.

[0030] The specific process of dynamically adjusting the noise reduction parameters by real-time monitoring of the deviation between the actual noise signal and the noise prediction result is as follows: Thirdly, specifically, residual noise in the ear canal is collected in real time using a feedback microphone, and the prediction error is calculated: ,in, Let be the actual residual noise collected by the feedback microphone at time t. Let be the noise prediction value at time t. This represents the prediction error.

[0031] The filter coefficients are updated using the variable step size NLMS algorithm, with the step size dynamically adjusted according to the error. The dynamic adjustment range is 0.1~0.01 to compensate for the amplitude deviation of burst noise. Fine-tuning the attention weights of the prediction network based on error trends optimizes long-term prediction accuracy. Prediction confidence control: When the error rate PER > 30%, temporarily switch to pure feedback noise reduction mode to disable the prediction function and avoid miscompensation.

[0032] S3: Replace the one-dimensional convolutional neural network and attention mechanism model in the feature fusion and extraction stage of S1 with a spiking neural network (SNN). Combine it with a memristor in-memory computing unit to achieve low-power computing. Based on the sparsity characteristics of the noise spectrum, redundant operations are reduced. A sound pressure change triggering mechanism is adopted to dynamically switch between energy-saving and enhancement modes according to the noise intensity.

[0033] The one-dimensional convolutional neural network and attention mechanism model in the feature fusion and extraction stage of S1 are replaced by a spiking neural network (SNN), and combined with a memristor in-memory computing unit to achieve low-power computing. Firstly, specifically, replacing traditional LSTM with spiking neural networks (SNNs). The one-dimensional convolutional neural network and attention mechanism model are transformed into an event-driven spiking neural network (SNN) to process multimodal data using neuronal spiking. Activating computation only when noise characteristics change reduces power consumption in static scenarios. Parameters connecting neurons in a spiking neural network are stored in a memristor array, allowing vector multiplication to be performed directly in memory, thus solving the bandwidth bottleneck caused by the separation of storage and computation in traditional DSP architectures.

[0034] Each processing unit, including the audio processing unit, the visual processing unit, and the motion processing unit, operates independently and collaborates through pulse events to achieve low-power computing. Secondly, specifically, based on the sparsity characteristics of the noise spectrum, the process of reducing redundant calculations is as follows: Based on the sparse distribution characteristics of dynamic noise spectrum in the frequency domain, redundant calculations are discarded in real time; The feature sequence of 50 frames is extracted by combining the attention mechanism of one-dimensional convolutional neural network (1D-CNN) and Transformer. The L1 norm of each frame feature in the feature sequence is calculated. When the sparsity is >70%, i.e. quiet scene, the sparse calculation mode is activated.

[0035] A sparse mask is generated by a spiking neural network, retaining only the frequency points with energy abrupt changes, such as the high-frequency band of a horn, for subsequent processing.

[0036] Sparse convolution performs convolution operations only on non-zero feature points, reducing redundant computations; Thirdly, specifically, the process of dynamically switching between energy-saving and enhancement modes based on noise intensity using a sound pressure surge triggering mechanism is as follows: Abandoning traditional timed sampling, an event-driven mechanism is adopted: Audio channel: When the microphone detects a sound pressure level change >15dB, it triggers the spiking neural network (SNN) calculation; The mode is dynamically switched based on the noise intensity. Energy-saving mode: In the current static scene, the signal-to-noise ratio is less than 10dB, and only 10% of neurons are retained to achieve low-power operation; Enhanced Mode: For sudden strong noise, full impulse calculation is used to enhance the noise reduction effect; The technical solution of this invention is as follows: A microphone array is constructed to collect audio signals, and environmental visual information and user motion state information are obtained using cameras, gyroscopes, and accelerometers. The collected multimodal data is preprocessed, and features are extracted and fused from the preprocessed multimodal data based on a deep learning model to generate multidimensional noise feature vectors. Historical noise feature vectors are analyzed based on a time-series prediction network to predict noise prediction results within the next 100ms. Based on the noise prediction results, adaptive filter parameters are adjusted in advance to generate a noise-reduced signal. The deviation between the actual noise signal and the noise prediction results is monitored in real time, and the noise reduction parameters are dynamically adjusted. A spiking neural network (SNN) replaces the one-dimensional convolutional neural network and attention mechanism model in the feature fusion and extraction stage of S1. Combined with a memristor in-memory computing unit, low-power computing is achieved. Based on the sparsity characteristics of the noise spectrum, redundant operations are reduced. A sound pressure surge triggering mechanism is adopted to dynamically switch between energy-saving and enhancement modes according to the noise intensity. Example

[0037] like Figure 2 As shown in Embodiment 1, the present invention provides an AI-based OWS Bluetooth headset noise cancellation system, comprising: Multimodal data acquisition and feature fusion module: Constructs a microphone array to acquire audio signals, uses cameras, gyroscopes, and accelerometers to acquire environmental visual information and user motion state information, preprocesses the acquired multimodal data, and extracts and fuses features from the preprocessed multimodal data based on a deep learning model to generate a multidimensional noise feature vector; Noise prediction and adaptive noise reduction module: Based on the temporal prediction network, the module analyzes the historical noise feature vectors and predicts the noise prediction results within the next 100ms. Based on the noise prediction results, the module adjusts the adaptive filter parameters in advance, generates a noise reduction signal, and monitors the deviation between the actual noise signal and the noise prediction results in real time, dynamically adjusting the noise reduction parameters. Low-power computing and mode switching module: The one-dimensional convolutional neural network and attention mechanism model in the feature fusion and extraction stage of the multimodal data acquisition and feature fusion module are replaced by a spiking neural network (SNN). Combined with a memristor in-memory computing unit, low-power computing is achieved. Based on the sparsity characteristics of the noise spectrum, redundant operations are reduced. A sound pressure change triggering mechanism is adopted to dynamically switch between energy-saving and enhancement modes according to the noise intensity.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An OWS Bluetooth headset noise cancellation method based on artificial intelligence, characterized in that: include: S1: Construct a microphone array to collect audio signals, use a camera, gyroscope, and accelerometer to acquire environmental visual information and user motion state information, preprocess the collected multimodal data, and extract and fuse features from the preprocessed multimodal data based on a deep learning model to generate a multidimensional noise feature vector; S2: Based on the temporal prediction network, analyze the historical noise feature vectors to predict the noise prediction results within the next 100ms. Based on the noise prediction results, adjust the adaptive filter parameters in advance to generate a noise reduction signal. Monitor the deviation between the actual noise signal and the noise prediction results in real time and dynamically adjust the noise reduction parameters. S3: Replace the one-dimensional convolutional neural network and attention mechanism model in the feature fusion and extraction stage of S1 with a spiking neural network (SNN). Combine it with a memristor in-memory computing unit to achieve low-power computing. Based on the sparsity characteristics of the noise spectrum, redundant operations are reduced. A sound pressure change triggering mechanism is adopted to dynamically switch between energy-saving and enhancement modes according to the noise intensity.

2. The OWS Bluetooth headset noise reduction method based on artificial intelligence according to claim 1, characterized in that: The process of constructing a microphone array to collect audio signals is as follows: An 8-microphone array was constructed, with the following arrangement: 3 feedforward microphones, 3 feedback microphones, and 2 bone conduction microphones were placed on the audio channel; the distance between the feedforward and feedback microphones was 15mm to form spatial filtering gain; the bone conduction microphones were placed close to the temporal bone to collect skull vibration signals; and the audio signals were obtained by collecting the feedforward, feedback, and bone conduction microphones.

3. The OWS Bluetooth headset noise reduction method based on artificial intelligence according to claim 1, characterized in that: The specific process of acquiring environmental visual information and user motion state information using cameras, gyroscopes, and accelerometers is as follows: A miniature CMOS camera is deployed with a lens tilted at 30° to cover a 120° field of view in front, used to acquire environmental images; A six-axis IMU is integrated into the earphone hinge to acquire user motion data. The six-axis IMU integrates a gyroscope and an accelerometer, and its sampling frequency is 100Hz, which can capture key frequency changes in human motion.

4. The OWS Bluetooth headset noise reduction method based on artificial intelligence according to claim 1, characterized in that: The specific process of analyzing historical noise feature vectors and predicting noise results within the next 100ms based on the time-series prediction network is as follows: Temporal feature sequence construction and preprocessing: sliding window acquisition: with a window length of 500ms and a step size of 10ms, multidimensional noise feature vectors are obtained to form a temporal sequence containing 50 frames of features, covering the duration of dynamic noise; Sequence enhancement calculates the first-order difference of the feature sequence to highlight the dynamic trend of noise changes. It normalizes the features of each dimension through Z-score and adopts a lightweight temporal prediction network strategy. It uses a one-dimensional convolutional neural network (1D-CNN) to extract local spectral features and uses the attention mechanism of Transformer to capture long-distance temporal dependencies. After analyzing 50 frames of feature sequences, it outputs the noise prediction results for the next 100ms.

5. The OWS Bluetooth headset noise reduction method based on artificial intelligence according to claim 1, characterized in that: The specific process of adjusting the adaptive filter parameters in advance based on the noise prediction results to generate the noise-reduced signal is as follows: A frequency band prediction strategy is adopted, with fine-grained prediction used for high-frequency noise to match sharp impact noise, and conventional prediction used for mid- and low-frequency noise. A 3-channel FIR filter bank is constructed, with the audio signal frequency of 20-500Hz as the low-frequency channel, the audio signal frequency of 500-2000Hz as the mid-frequency channel, and the audio signal frequency of 2000-20kHz as the high-frequency channel. Based on the noise spectrum predicted by the noise prediction results, the filter coefficients are adjusted in advance: Low-frequency channel: suppresses slowly varying noises such as engine roar; High-frequency channel: responds quickly to sudden noises such as horn blasts. If the energy envelope shows a rapid change in noise energy, the high-frequency channel filter will quickly adjust its coefficients. If the energy envelope shows a slow change in noise energy, the low-frequency channel filter will smoothly adjust its coefficients. Based on the spectrum prediction map of the noise prediction result, i.e. the energy distribution of noise in the 20Hz-20kHz frequency band, filter coefficients that can cancel the noise prediction result are generated in reverse phase, ensuring that the noise-reduced signal and the noise prediction result have opposite acoustic phase and similar amplitude. The DSP's SIMD instructions are used to process the 3-channel audio signal in parallel. The original audio is convolved with the pre-configured filter coefficients to generate an inverse noise-reduced signal. The processing delay of the low-frequency channel, mid-frequency channel, and high-frequency channel is controlled within 2ms. The weights of each channel are allocated according to the energy envelope, and the signals are superimposed to generate the final noise-reduced signal.

6. The OWS Bluetooth headset noise reduction method based on artificial intelligence according to claim 1, characterized in that: The specific process of dynamically adjusting the noise reduction parameters by real-time monitoring of the deviation between the actual noise signal and the noise prediction result is as follows: The residual noise in the ear canal is collected in real time by a feedback microphone, the prediction error is calculated, the filter coefficients are updated by a variable step size NLMS algorithm, the step size is dynamically adjusted with the error, the attention weights of the prediction network are fine-tuned based on the error trend, and when the error rate PER>30%, the pure feedback noise reduction mode is switched to and the prediction function is disabled.

7. The OWS Bluetooth headset noise reduction method based on artificial intelligence according to claim 1, characterized in that: The specific process for achieving low-power computing is as follows: The one-dimensional convolutional neural network and attention mechanism model are transformed into an event-driven spiking neural network (SNN) to process multimodal data using neuronal spiking. The computation is activated only when the noise characteristics change, reducing power consumption in static scenes. The parameters connecting the neurons in the spiking neural network are stored in a memristor array, and vector multiplication is performed directly in memory. Each processing unit, including the audio processing unit, the visual processing unit, and the motion processing unit, operates independently and cooperates through pulse events to achieve low-power computation.

8. The OWS Bluetooth headset noise reduction method based on artificial intelligence according to claim 1, characterized in that: The specific process for reducing redundant operations is as follows: The system extracts 50 frames of feature sequences using the attention mechanism of 1D-CNN and Transformer. The L1 norm of each frame feature is calculated. When the sparsity is >70%, i.e. a quiet scene, the sparse computation mode is activated. A sparse mask is generated by a spiking neural network to retain only the frequency points of energy abrupt changes. The sparse convolution is performed only on non-zero feature points to reduce redundant computation.

9. The OWS Bluetooth headset noise reduction method based on artificial intelligence according to claim 1, characterized in that: The specific process of dynamically switching between energy-saving and enhancement modes based on noise intensity is as follows: The mode is dynamically switched based on the noise intensity. Energy-saving mode: In the current static scene, the signal-to-noise ratio is less than 10dB, and only 10% of neurons are retained to achieve low-power operation; Enhanced Mode: For sudden strong noise, full impulse calculation is used to enhance the noise reduction effect.

10. An AI-based OWS Bluetooth headset noise cancellation system, characterized in that: include: Multimodal data acquisition and feature fusion module: Constructs a microphone array to acquire audio signals, uses cameras, gyroscopes, and accelerometers to acquire environmental visual information and user motion state information, preprocesses the acquired multimodal data, and extracts and fuses features from the preprocessed multimodal data based on a deep learning model to generate a multidimensional noise feature vector; Noise prediction and adaptive noise reduction module: Based on the temporal prediction network, the module analyzes the historical noise feature vectors and predicts the noise prediction results within the next 100ms. Based on the noise prediction results, the module adjusts the adaptive filter parameters in advance, generates a noise reduction signal, and monitors the deviation between the actual noise signal and the noise prediction results in real time, dynamically adjusting the noise reduction parameters. Low-power computing and mode switching module: The one-dimensional convolutional neural network and attention mechanism model in the feature fusion and extraction stage of the multimodal data acquisition and feature fusion module are replaced by a spiking neural network (SNN). Combined with a memristor in-memory computing unit, low-power computing is achieved. Based on the sparsity characteristics of the noise spectrum, redundant operations are reduced. A sound pressure change triggering mechanism is adopted to dynamically switch between energy-saving and enhancement modes according to the noise intensity.

Images