A Smart Noise Control Method for Protective Earmuffs

Abstract

This invention relates to the field of noise reduction technology, and more particularly to an intelligent noise reduction control method for protective earmuffs. The method includes the following steps: acquiring multi-source acoustic data through a multi-array sensor network; extracting key noise features and identifying noise types using a spectral embedded density peak clustering algorithm, generating noise analysis results; constructing a noise reduction decision model based on a near-end strategy optimization algorithm, and generating an optimal strategy matrix; independently adjusting multiple active noise reduction channels using an adaptive PID controller to achieve precise multi-channel noise suppression; and dynamically optimizing and updating control parameters based on historical noise reduction effects and user comfort feedback using a Bayesian optimization algorithm. This invention not only achieves more accurate, efficient, and comfortable intelligent noise reduction in varied and complex noise environments, but also possesses continuous adaptive and self-optimizing capabilities, improving the intelligence level of protective earmuffs and the user experience.

Description

Technical Field

[0001] This invention relates to the field of noise reduction technology, and more particularly to an intelligent noise reduction control method for protective earmuffs. Background Technology

[0002] With the continuous increase in noise from industrial production, transportation, and urban environments, the harm caused by noise pollution to human health, especially the hearing system, has received increasing attention. Protective earmuffs, as a commonly used hearing protection device, are widely used in high-noise environments such as factory workshops, airports, and construction sites to isolate external noise and improve work comfort. However, traditional protective earmuffs mostly use passive noise reduction materials, relying solely on sound-absorbing pads to absorb and block noise. Their effectiveness in suppressing low-frequency noise, impact noise, and non-stationary noise is limited, and they cannot achieve precise and dynamic noise control. Currently, the following problems still exist: Existing technologies mostly use general noise feature extraction and recognition algorithms, which are difficult to efficiently distinguish between various noise types such as transient, impulsive, and periodic noise. This results in noise reduction measures not being optimized for specific noise characteristics, affecting the noise reduction effect; Traditional protective earmuffs mostly use static or empirical parameters for noise reduction control, which cannot dynamically adjust noise reduction parameters according to real-time environmental changes and individual user differences, easily leading to insufficient noise reduction performance or user discomfort; Most existing active noise cancellation methods only target a single noise cancellation unit or channel, making it difficult to achieve intelligent coordination and adaptive adjustment between multi-channel noise cancellation units, affecting the noise reduction capability of earmuffs in complex multi-source noise environments. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides an intelligent noise reduction control method for protective earmuffs. This method solves the problem of how existing protective earmuffs struggle to efficiently identify various noise types, such as transient, impact, and periodic noise, in complex multi-source noise environments in industrial and urban settings. It also enables intelligent adaptive multi-channel noise reduction control based on real-time environmental and user differences, thereby improving overall noise reduction performance and wearing comfort.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A smart noise control method for protective earmuffs includes the following steps:

[0006] S1: Acquire multi-source acoustic data of the protective earmuffs through a multi-array sensor network;

[0007] S2: Based on the multi-source acoustic data, the key noise feature factors are extracted and the noise type is identified using the spectral embedded density peak clustering algorithm to generate noise analysis results;

[0008] S3: Based on the noise analysis results, construct a noise reduction decision model based on the near-end strategy optimization algorithm, dynamically evaluate the noise reduction effect and comfort index through the reward function, and generate the optimal noise reduction strategy matrix for transient noise, impact noise and periodic noise in real time.

[0009] S4: Based on the optimal noise reduction strategy matrix, an adaptive PID controller is used to assign control commands to multiple active noise reduction units inside the earmuff, thereby realizing the adjustment of the multi-channel noise reduction execution module and adaptive noise suppression.

[0010] S5: Based on historical noise reduction effects and user comfort feedback, dynamically adjust noise reduction parameters and control strategies using a Bayesian optimization algorithm, and feed back the optimization results to update the noise reduction decision model and control commands.

[0011] Furthermore, the multi-source acoustic data includes external environmental noise signals, user voice signals, earmuff structure vibration sound signals, bone conduction sound signals, and sound pressure signals inside and outside the earmuff cavity.

[0012] Furthermore, step S2 includes the following steps:

[0013] The multi-source acoustic data is normalized, and a unified acoustic time-spectrum diagram is constructed using short-time Fourier transform to extract the energy distribution characteristics and amplitude spectrum characteristics within the corresponding time period.

[0014] The acoustic time-spectrum graph is input into a spectrum-embedded density peak clustering algorithm to identify potential noise-dense regions by utilizing the relationship between local density and distance, and to extract key noise feature factors that include noise intensity, frequency band characteristics, and instantaneous amplitude.

[0015] Based on feature factors, a multi-class discrimination model based on graph regularization constraints is constructed to classify and identify external background noise, structural vibration noise, bone conduction sound and user speech.

[0016] The classification results are assigned weight labels based on recognition confidence, and information on noise category, sound source spatial label, and feature intensity is integrated to generate noise analysis results.

[0017] Furthermore, the construction process of the multi-class discrimination model includes the following steps:

[0018] A set of feature vectors is constructed based on the key noise feature factors, and the similarity matrix between feature samples is calculated;

[0019] Based on the set of feature vectors, a feature graph structure is constructed using a graph regularization constraint strategy. The similarity matrix is ​​then subjected to spectral embedding mapping using the Laplacian operator to generate a low-dimensional feature subspace.

[0020] Labeled training samples are introduced into the low-dimensional feature subspace to construct a multi-class discrimination model. The objective is jointly optimized by the minimum graph regularization loss and the maximum inter-class discrimination divergence. The multi-class discrimination mapping function is solved by iterative training to achieve the classification of background noise, structural noise, bone conduction sound and speech signals.

[0021] The graph structure parameters and regularization coefficients are fine-tuned through cross-validation to complete the training of the multi-class discrimination model, and finally output noise analysis data containing classification labels, confidence coefficients and noise type mapping results.

[0022] Furthermore, the formula for the multi-class discrimination model is as follows:

[0023]

[0024] in, This represents the final output of the multi-class discrimination, i.e., the noise category label to which the final judgment belongs; k represents the index of the k-th noise type; K represents the total number of noise types; N represents the number of key noise feature factors extracted in the current frame; This represents the response intensity of the i-th key noise feature factor on the k-th noise model; This represents the response strength of the i-th key noise feature factor on the l-th noise model; This represents the average acoustic energy value of the i-th feature factor in the k-th class of training samples; This represents the overall average acoustic energy of the i-th feature factor across all training categories; This represents the energy standard deviation of the i-th feature factor across all categories; This represents a small positive constant introduced to prevent the denominator from being zero; This represents the confidence label value of the i-th feature factor in the k-th class; This represents the embedding distance between the i-th feature and the center of the k-th noise subspace; The influence of the response normalization ratio, the spectral energy difference term, and the spatial distance confidence term are controlled separately.

[0025] Furthermore, step S3 includes the following steps:

[0026] The noise analysis results are processed to extract the dominant noise type, characteristic frequency band, energy weight, spatial label and its identification confidence in the current period, and a dynamic environmental state vector is constructed by combining it with historical identification data.

[0027] The environmental state vector is input into the noise reduction decision model constructed based on the near-end policy optimization algorithm, with the joint optimization objectives of minimizing noise residue, improving response speed and maintaining user comfort, to generate a set of policy candidates;

[0028] During the strategy generation process, a hierarchical control framework for multiple types of noise is constructed, a rapid response priority is set for transient noise, a short-term burst suppression mechanism is enabled for impact noise, and a predictive intervention strategy is introduced for periodic noise.

[0029] By combining the candidate strategy set with user-personalized parameters, the strategy results are differentiated and weighted to obtain the optimal noise reduction strategy matrix for the current noise scenario. The optimal noise reduction strategy matrix is ​​then used as an output signal and transmitted to the control execution layer. The user-personalized parameters include earmuff tightness, ear canal reflection characteristics, adaptive historical parameters, and subjective comfort feedback level.

[0030] Furthermore, the formula for the noise reduction decision model is as follows:

[0031]

[0032] in, This represents the policy score corresponding to the j-th type of noise; This represents the characteristic frequency band energy value of the j-th type of noise in the current time period; This represents the amplitude spectrum activation coefficient corresponding to the j-th type of noise; This indicates the degree of spatial disturbance of the noise within the current period; This represents the average noise reduction gain of the periodic strategy response to the j-th type of noise. The time-delay sensitivity parameter represents the strategy corresponding to the j-th type of noise; This represents the system's response delay to type j noise when the current strategy is executed; This represents the control path gradient under the j-th type of noise; This represents the intensity of false triggering or vibration interference under the influence of type j noise interference. This represents a stability constant set to prevent the denominator from becoming too small; These represent the weights of the acoustic driving factor, historical response factor, and control stability factor on the scoring function, respectively.

[0033] Furthermore, the multi-channel noise reduction execution module integrates multiple digital signal processing units, each corresponding to a different acoustic region, and can perform independent signal modulation and amplitude and phase adjustment according to the instructions output by the adaptive PID controller.

[0034] Furthermore, the Bayesian optimization algorithm takes maximizing the user's long-term comfort as the objective function, dynamically selects different families of parameter adjustment strategies based on the noise identification results, and outputs an updated set of candidate control parameters.

[0035] The beneficial effects of this invention are as follows:

[0036] This invention utilizes a multi-array sensor network to collect multi-source acoustic data in real time, enabling multi-dimensional monitoring and dynamic response of the sound field inside and outside the earcups, significantly improving the comprehensiveness and accuracy of noise perception. Employing a spectral embedded density peak clustering algorithm, it effectively extracts and distinguishes key noise feature factors. Combined with intelligent identification of different types of noise (such as transient, impulsive, and periodic noise), it ensures a more targeted and adaptable noise reduction solution. Through dynamic evaluation based on a near-end strategy optimization algorithm and reward function, it comprehensively considers noise reduction effect and user comfort under different noise environments, achieving dynamic trade-offs and optimal strategy generation, effectively avoiding side effects such as auditory fatigue that are easily caused by traditional noise reduction methods. An adaptive PID controller is used to assign control commands to multiple active noise reduction units and achieve multi-channel coordinated adjustment, enabling the earcups to maintain stable and efficient noise reduction performance even when facing complex and varied noise sources. By combining historical noise reduction effects and user subjective comfort feedback, a Bayesian optimization algorithm is introduced to continuously learn and dynamically adjust noise reduction parameters and control strategies, achieving continuous self-optimization of earcup performance, improving long-term wearing experience and adaptability to individual user needs. This invention updates the noise reduction decision model and control commands in real time by optimizing the results, forming an intelligent closed loop of data-decision-feedback, which significantly improves the autonomous evolution capability of noise reduction control and the overall intelligence level of the system. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating an intelligent noise reduction control method for protective earmuffs according to the present invention.

[0038] Figure 2 This is a flowchart illustrating step S2 provided in an embodiment of the present invention.

[0039] Figure 3 This is a flowchart illustrating step S3 provided in an embodiment of the present invention. Detailed Implementation

[0040] Please see Figure 1-3 As shown, the present invention relates to an intelligent noise reduction control method for protective earmuffs.

[0041] Example

[0042] A smart noise control method for protective earmuffs includes the following steps:

[0043] S1: Collect multi-source acoustic data of the protective earmuff through a multi-array sensor network; the multi-source acoustic data includes external environmental noise signals, user voice signals, earmuff structure vibration sound signals, bone conduction sound signals, and sound pressure signals inside and outside the earmuff cavity.

[0044] In one embodiment, the intelligent acoustic acquisition of the protective earmuff consists of multiple integrated micro-sensor units, forming a multi-array sensor network with spatial distribution sensing capability, for high-precision, full-coverage acquisition of multi-source acoustic data.

[0045] The specific composition and layout of the sensor array are as follows:

[0046] 1. Ambient sound microphone array (external noise acquisition): 3-5 MEMS condenser microphones are installed on each side of the earcup shell in a ring or fan-shaped array to support stereo acquisition and sound source localization. This array is used to capture broadband noise in the external environment (such as industrial noise, traffic noise, crowd noise, etc.) and has good directionality and wind noise resistance.

[0047] 2. Bone conduction sensor (user voice extraction): A piezoelectric ceramic or bone conduction resistor sensor is embedded in the inner side of the ear cup, in the area that contacts the user's cheekbone or auricle. It is used to pick up low-frequency vibration signals transmitted by the user's skull, separate the user's own voice signal for subsequent recognition and processing, and avoid the user's voice being mistakenly identified as external noise.

[0048] 3. Structural vibration sensor (shell-borne noise detection): A miniature MEMS accelerometer and a piezoelectric diaphragm are attached between the inner shell of the earcup and the speaker structure to capture structural noise caused by shell resonance, mechanical impact, etc. These signals are usually difficult to detect through air conduction.

[0049] 4. Sound pressure sensor group (internal and external sound pressure comparison monitoring): One sound pressure sensor is installed inside the ear cup cavity near the ear canal opening and another is installed on the outer edge. The sound pressure changes inside and outside the ear cup can be measured in real time to help identify problems that affect noise reduction performance, such as decreased airtightness and sound leakage paths.

[0050] All sensors are synchronously sampled by a multi-channel A / D sampling chip driven by a single master clock, ensuring acoustic data timing alignment between different channels and avoiding recognition deviations caused by delays. The sampling frequency can be set from 16kHz to 48kHz, and the sampling resolution is 16-bit or 24-bit, dynamically switching according to different noise frequency band requirements to ensure speech clarity and low-frequency noise resolution. Basic processing such as initial noise reduction, gain control, and DC offset cancellation is completed locally in the DSP chip built into the earcups, reducing the burden on wireless transmission.

[0051] All acquired data is aggregated via multiple MUX channels and then transmitted to the central processing chip (MCU / DSP) via I2S or SPI bus. Some earcups may also utilize a USB bridge chip for high-throughput debugging and uploading. The earcups integrate a low-power Bluetooth module (Bluetooth 5.0 or higher) for uploading real-time acoustic data, noise type information, and user feedback to smart terminals (such as apps), supporting remote parameter distribution and model updates. For critical nodes (such as bone conduction channels), the system supports hot backup channels, automatically switching to backup sensors in case of primary channel failure to ensure uninterrupted signal acquisition.

[0052] S2: Based on the multi-source acoustic data, the key noise feature factors are extracted and the noise type is identified using the spectral embedded density peak clustering algorithm to generate noise analysis results;

[0053] Step S2 includes the following steps:

[0054] The multi-source acoustic data is normalized, and a unified acoustic time-spectrum diagram is constructed using short-time Fourier transform to extract the energy distribution characteristics and amplitude spectrum characteristics within the corresponding time period.

[0055] Specifically, the multi-source acoustic data (including ambient sound, bone conduction sound, and structural vibration sound) collected by the sensor array is normalized. The normalization process includes amplitude normalization and sampling rate unification for different source data to ensure consistency in subsequent feature processing. Short-Time Fourier Transform (STFT) is used to convert the normalized data into acoustic time-spectrum graphs. The system frames and windows the acoustic signal, extracting the frequency energy distribution and amplitude spectrum variation features within each frame, forming a series of two-dimensional spectrogram sets covering different time periods, which serve as the input basis for subsequent clustering.

[0056] The acoustic time-spectrum graph is input into a spectrum-embedded density peak clustering algorithm to identify potential noise-dense regions by utilizing the relationship between local density and distance, and to extract key noise feature factors that include noise intensity, frequency band characteristics, and instantaneous amplitude.

[0057] Specifically, after constructing the acoustic time-spectrum graph, the graph is used as input to be processed by the spectral embedded density peak clustering module. The specific operation includes the following steps:

[0058] (1) Construction of spectral feature space:

[0059] The system first extracts the frequency-amplitude data from each frame of the acoustic spectrogram into vector form, and then extracts high-dimensional feature sets from multiple time windows to form a frame-level feature library covering the dynamic spectral distribution. To enhance the expressive power of local patterns, the following auxiliary information is also introduced during feature processing: the energy centroid of each frame (reflecting the concentrated energy frequency band); the amplitude rise rate (used to identify burst signals); the narrowband bandwidth distribution (used to distinguish between periodic and white noise sound sources); and the spectral entropy (characterizing the complexity of the frequency domain distribution). These features together form a multi-dimensional feature vector set, serving as the basic data for density peak analysis.

[0060] (2) Calculation of density and distance indices:

[0061] By calculating the local density of each feature point in the feature space (such as the number of its neighboring points) and its relative distance to other high-density points, it can be determined whether it is a potential cluster center.

[0062] For example:

[0063] Density peaks often appear in high-frequency, high-energy regions, corresponding to mechanical impact sound or structural resonance.

[0064] The low-frequency broadband region has low density but high energy, which may be background wind noise or low-frequency traffic noise.

[0065] This approach avoids relying on manually setting the number of clusters and can more adaptively discover the distribution patterns of real noise types in different environments.

[0066] (3) Spectral embedding fusion:

[0067] To more effectively distinguish acoustic features with nonlinear distributions, the system introduces a spectral embedding mechanism. By constructing a similarity matrix of the spectrogram and its graphical Laplace representation, the high-dimensional complex spectral structure is mapped to a lower-dimensional continuous space, enhancing the separability of cluster boundaries. The embedding results not only preserve the structural connectivity between features but also enhance the ability to separate signals in overlapping frequency bands, making it particularly suitable for scenarios with mixed sound sources (such as speech superimposed with mechanical noise).

[0068] (4) Feature factor output:

[0069] Ultimately, each noise cluster in the clustering results will generate corresponding key feature factors. These factors include: the main frequency band range of noise energy; the time and intensity of peak amplitude occurrence; the energy rise / fall trend; and the spatial channel source identifier corresponding to the time-frequency window (used for subsequent spatial localization). These factors will serve as inputs for subsequent discrimination and classification, and will also be used for semantic-level noise interpretation and analysis.

[0070] Based on feature factors, a multi-class discrimination model based on graph regularization constraints is constructed to classify and identify external background noise, structural vibration noise, bone conduction sound and user speech.

[0071] The classification results are assigned weight labels based on recognition confidence, and information on noise category, sound source spatial label, and feature intensity is integrated to generate noise analysis results.

[0072] The construction process of the multi-class discriminant model includes the following steps:

[0073] A set of feature vectors is constructed based on the key noise feature factors, and the similarity matrix between feature samples is calculated;

[0074] Specifically, several key noise feature factors are obtained from the spectrum embedded density peak clustering module, including the following sub-dimensions: energy center frequency; amplitude peak time position; spectrum shape features (such as slope and bandwidth); feature intensity and signal abrupt change; source channel number and time window label.

[0075] The above content is uniformly encoded as fixed-length vectors, with each vector representing a segment of noise event to be classified. All feature vectors constitute the feature set, denoted as the training dataset.

[0076] To more effectively preserve the structural relationships between different feature samples, the system performs pairwise similarity calculations on all feature vector pairs, forming a symmetric similarity matrix. Specifically, this includes: optimizing weighted cosine similarity, spectral correlation coefficient, or inverse Euclidean distance, with weights determined by the importance of different feature dimensions; to improve computational efficiency, the system retains the similarity relationship between each feature point and its K nearest neighbors, setting other positions to zero, forming a sparse connectivity graph; each feature point is treated as a node in the graph, and similarity constitutes the weight of the edges in the graph, achieving spatial-spectral domain fusion structural modeling.

[0077] Based on the set of feature vectors, a feature graph structure is constructed using a graph regularization constraint strategy. The similarity matrix is ​​then subjected to spectral embedding mapping using the Laplacian operator to generate a low-dimensional feature subspace.

[0078] Specifically, a graph regularization mechanism is introduced based on the feature map structure to ensure that feature nodes with high similarity tend to obtain consistent class predictions in subsequent discrimination, while enhancing stability at class boundaries: the system encodes the connectivity and weight relationships between all feature points in the form of a graph; a graph smoothing constraint is introduced to make the outputs of similar nodes as similar as possible during model training; the sparsity and edge weight distribution of the graph structure are calibrated from prior acoustic data, which can adapt to non-uniform feature spaces. This graph regularization strategy plays a crucial role in processing high-dimensional complex structures after spectral embedding and can effectively maintain the topological stability of features.

[0079] To reduce the computational complexity of classification and improve class separability, the system performs the following operations: calculates the Laplacian matrix of the graph structure and performs eigenvalue decomposition to extract several principal feature dimensions; maps the original high-dimensional features to a low-dimensional feature subspace through spectral embedding, so that samples with similar structures are clustered and samples between categories are separated; the feature representation after low-dimensional mapping not only preserves the connectivity of the original graph, but also improves the discrimination efficiency of linear or nonlinear classification models.

[0080] Labeled training samples are introduced into the low-dimensional feature subspace to construct a multi-class discrimination model. The objective is jointly optimized by the minimum graph regularization loss and the maximum inter-class discrimination divergence. The multi-class discrimination mapping function is solved by iterative training to achieve the classification of background noise, structural noise, bone conduction sound and speech signals.

[0081] Specifically, in the reduced-dimensional feature subspace, the system loads training samples with manually labeled information as supervision information, representing the following noise types: background broadband noise (such as traffic noise, wind noise); structural vibration noise (such as earmuff material resonance, impact feedback); bone conduction noise (user speech is transmitted through the skull); and user direct speech (microphone pickup signal), etc.

[0082] The classifier is constructed as follows: Support Vector Machine (SVM), multi-class Softmax classifier or lightweight convolutional neural network can be used as the discriminant model; during training, the graph regularization loss term and the classification error loss term are minimized at the same time; the system adopts an objective function with graph structure constraints, so that the model can maximize the inter-class discriminability while maintaining intra-class consistency.

[0083] To ensure strong generalization ability and high classification stability, the system performs the following training and optimization process: multiple rounds of batch training are used to update the discriminant parameters until convergence; the training set is divided into multiple folds to verify the classification performance of the model under different data distributions; hyperparameter tuning, including regularization strength coefficient, graph connectivity sparsity, embedding dimension, etc., is optimized through grid search or Bayesian parameter tuning methods; early stopping mechanism, Dropout, or data augmentation techniques are introduced to improve the robustness of the model.

[0084] The graph structure parameters and regularization coefficients are fine-tuned through cross-validation to complete the training of the multi-class discrimination model, and finally output noise analysis data containing classification labels, confidence coefficients and noise type mapping results.

[0085] Furthermore, the formula for the multi-class discrimination model is as follows:

[0086]

[0087] in, This represents the final output of the multi-class discrimination, i.e., the noise category label to which the final judgment belongs; k represents the index of the k-th noise type; K represents the total number of noise types; N represents the number of key noise feature factors extracted in the current frame; This represents the response intensity of the i-th key noise feature factor on the k-th noise model; This represents the response intensity of the i-th key noise feature factor on the l-th noise model; This represents the average acoustic energy value of the i-th feature factor in the k-th class of training samples; This represents the overall average acoustic energy of the i-th feature factor across all training categories; This represents the energy standard deviation of the i-th feature factor across all categories; This represents a small positive constant introduced to prevent the denominator from being zero; This represents the confidence label value of the i-th feature factor in the k-th class; Let represent the embedding distance between the i-th feature and the center of the k-th noise subspace, and calculate the geometric position of the feature in the Laplacian spectrum mapping space; The influence of the response normalization ratio, the spectral energy difference term, and the spatial distance confidence term are controlled separately.

[0088] The calculation formula is as follows:

[0089]

[0090] in, This represents the response intensity of the i-th key noise feature factor on the k-th noise model; This represents the set of principal frequencies corresponding to the i-th feature factor; This represents the power spectral density of the characteristic factor at frequency f, i.e., its energy at that frequency; This represents the spectral weight of the k-th type of noise at frequency f.

[0091] The calculation formula is as follows:

[0092]

[0093] in, This represents the average acoustic energy value of the i-th feature factor in the k-th class of training samples; This represents the number of training samples in the k-th class; This represents the j-th training sample of the k-th class; Indicates the i-th feature in the sample The normalized energy value in.

[0094] S3: Based on the noise analysis results, construct a noise reduction decision model based on the near-end strategy optimization algorithm, dynamically evaluate the noise reduction effect and comfort index through the reward function, and generate the optimal noise reduction strategy matrix for transient noise, impact noise and periodic noise in real time.

[0095] Step S3 includes the following steps:

[0096] The noise analysis results are processed to extract the dominant noise type, characteristic frequency band, energy weight, spatial label and its identification confidence in the current period, and a dynamic environmental state vector is constructed by combining it with historical identification data.

[0097] Specifically, the noise analysis results are subjected to secondary processing to extract the following:

[0098] Dominant noise types: such as sudden mechanical shocks, periodic equipment operation noises, voice interference, or background environmental noise;

[0099] Featured frequency bands and energy weight distribution: Identify the concentrated regions of current noise in the spectrum and their relative energy proportions;

[0100] Spatial labeling: using array sound source localization methods to pinpoint the main direction of noise or the location of sound sources;

[0101] Identification confidence: Outputs an identification reliability assessment value based on cluster stability and historical consistency.

[0102] By combining the identification records over a period of time, the system constructs a dynamic environmental state vector that can be updated over time to depict the overall picture and evolution trend of the current acoustic environment.

[0103] The environmental state vector is input into the noise reduction decision model constructed based on the near-end policy optimization algorithm, with the joint optimization objectives of minimizing noise residue, improving response speed and maintaining user comfort, to generate a set of policy candidates;

[0104] Specifically, after constructing the environment state vector, the system inputs it into a noise reduction decision model based on the Proximal Policy Optimization (PPO) algorithm. The model learns the policy effects from historical interactions through a reinforcement learning mechanism and generates a set of noise reduction policy candidates suitable for the current environment. Specific implementation details are as follows:

[0105] Each candidate strategy includes a combination of control parameters across multiple dimensions, such as: output power adjustment factor for each channel; phase adjustment delay time; response type of the active noise cancellation filter; switching rate setting and gain response curve; suppression intensity level and duration window.

[0106] Generation mechanism: The decision model automatically evaluates the adaptability of different strategies to the current noisy scene based on the environmental state vector; it also considers the short-term noise reduction effect and the medium- and long-term user experience, thus forming multiple strategy combinations with different performance focuses; these strategy combinations are temporarily stored in the candidate strategy cache area, with a strategy prediction score attached, for subsequent weighted screening.

[0107] During the generation strategy process, the model dynamically adjusts the weights of the three objectives: "maximizing noise reduction," "minimizing response latency," and "ensuring user comfort." The strategy bias can be adjusted based on real-time feedback, such as favoring a high-intensity suppression strategy in noisy environments and prioritizing comfort in quiet environments.

[0108] To ensure response speed and computational efficiency, the system limits the number of candidate strategies to no more than a preset threshold (e.g., 5-10 groups); redundant solutions are automatically eliminated according to the current scoring priority, retaining the subset of strategies with the greatest implementation potential. Through the above process, the system can quickly generate a set of noise reduction strategy candidates that are highly adaptable and stable in different acoustic environments, providing a structured basis for subsequent fine-tuning.

[0109] During the strategy generation process, a hierarchical control framework for multiple types of noise is constructed, a rapid response priority is set for transient noise, a short-term burst suppression mechanism is enabled for impact noise, and a predictive intervention strategy is introduced for periodic noise.

[0110] Specifically, to improve the accuracy of processing different types of noise, a hierarchical control mechanism is introduced during the noise reduction strategy generation process. Different control logic paths are triggered based on the noise category to achieve differentiated protection, including:

[0111] a. Transient noise processing mechanism

[0112] Applicable scenarios: such as sudden, high-intensity, short-term noise such as elevator doors closing, express delivery, or objects falling.

[0113] Processing logic: After identifying transient noise signals, the system switches the noise reduction strategy to high response priority mode; activates the fast dynamic gain adjustment in the noise reduction module to increase the output rate of the channel signal; and disables the buffering mechanism of some filters so that the controller can complete the loading of the corresponding parameters with minimal delay.

[0114] b. Impact noise suppression mechanism

[0115] Applicable scenarios: such as noise with impact amplitude and typical waveform characteristics, such as equipment knocking and metal impact.

[0116] Processing logic: Initiate a suppression strategy within a short time window, such as a dynamic limiting mechanism to intercept peak signals; activate the local frequency band suppression module to suppress energy in the main noise frequency band; the noise reduction unit maintains high-intensity output for a short time, and then quickly returns to the basic state to avoid excessive suppression causing adverse effects.

[0117] c. Periodic noise prediction intervention strategy

[0118] Applicable scenarios: such as continuous and regular noise from fans, motors, air conditioner compressors, etc.

[0119] Processing logic: Based on the identification results of the previous rounds, the system infers the repetition period and amplitude characteristics of the noise waveform; constructs a period prediction module to preload the phase adjustment and filtering settings within the target period; and implements "feedforward" noise reduction intervention, that is, pre-setting a reverse signal before the noise reaches its peak to enhance the efficiency of interference intervention.

[0120] d. Cross-type fusion scenario processing

[0121] In mixed noise scenarios, the system will weight and combine the response logic according to the energy ratio and confidence level of various noises; periodic prediction and transient fast response can be executed in parallel to ensure that the system has composite noise reduction capability in complex acoustic fields.

[0122] Through this multi-level noise processing framework, the noise reduction system can not only "identify different noises" but also "respond to different noises in different ways," significantly improving the overall noise processing efficiency, strategy matching, and wearing experience.

[0123] By combining the candidate strategy set with user-personalized parameters, the strategy results are differentiated and weighted to obtain the optimal noise reduction strategy matrix for the current noise scenario. The optimal noise reduction strategy matrix is ​​then used as an output signal and transmitted to the control execution layer. The user-personalized parameters include earmuff tightness, ear canal reflection characteristics, adaptive historical parameters, and subjective comfort feedback level.

[0124] In one embodiment, to achieve individualized adaptation of the noise reduction strategy, user-specific parameters are introduced to weight and correct the strategy candidate set, ensuring that the output not only matches the noise characteristics but also the user's own characteristics. Implementation details are as follows:

[0125] a. Source of personalized parameters

[0126] Earmuff fit tightness: By integrating a flexible pressure sensor into the inner wall of the earmuff, the contact pressure between the earmuff and the user's auricle is sensed in real time; the degree of sealing is inferred based on the pressure value, thereby judging the actual performance of the earmuff's sound insulation; when the sealing is too low, the system will automatically increase the noise reduction intensity to compensate for insufficient physical attenuation.

[0127] Ear canal reflection characteristics: During the first use or intermittent use of the device, the earcups can emit a weak amplitude test signal and receive the reflection; the acoustic response characteristics of the user's ear canal can be determined based on the echo time and frequency response curves; for users with special ear shapes (such as deep ear canals or high reflection), an adaptive filter curve can be automatically matched.

[0128] Adaptive historical parameters: The system records the user's acceptance history of various strategies in similar noise situations; including whether the user manually adjusts the volume, whether "discomfort feedback" is triggered, and other operational behaviors; and constructs a user response preference curve based on historical behavior as the basis for current weighting.

[0129] Subjective comfort feedback level: Users can provide feedback on their current comfort level through buttons, the APP rating system, or voice commands; the system converts the comfort level into a strategy adjustment signal, affecting the fine-tuning of the output strategy in terms of intensity and rhythm.

[0130] b. Weighted Correction Process

[0131] Personalized parameters are mapped to weighting factors to prioritize the control parameters in the strategy candidate set. If a strategy has the best technical indicators but a poor user experience record, the system will automatically lower its priority. If multiple strategies have similar scores, the one with the highest historical preference consistency will be selected first. The weighting process can be implemented through a preset rule base or machine learning model to ensure that it follows physical laws while respecting subjective experience.

[0132] c. Output of the optimal policy matrix

[0133] After weighting, the system selects the strategy combination with the highest comprehensive score to form the final optimal noise reduction strategy matrix. This matrix encapsulates information such as the target output gain, delay control, and modulation type of each channel. The matrix is ​​synchronously transmitted to the control execution layer in step S4, where it is parsed and executed by the adaptive PID module. During the control process, this matrix also serves as a feedback reference baseline, used to compare with the subsequent actual noise reduction effect and correct the closed-loop model.

[0134] Furthermore, the formula for the noise reduction decision model is as follows:

[0135]

[0136] in, This represents the policy score corresponding to the j-th type of noise; This represents the characteristic frequency band energy value of the j-th type of noise in the current time period; This represents the amplitude spectrum activation coefficient corresponding to the j-th type of noise; This indicates the degree of spatial disturbance of the noise within the current period; This represents the average noise reduction gain of the periodic strategy response to the j-th type of noise. The time-delay sensitivity parameter represents the strategy corresponding to the j-th type of noise; This represents the system's response delay to type j noise when the current strategy is executed; This represents the control path gradient under the j-th type of noise; This represents the intensity of false triggering or vibration interference under the influence of type j noise interference. This represents a stability constant set to prevent the denominator from becoming too small; These represent the weights of the acoustic driving factor, historical response factor, and control stability factor on the scoring function, respectively.

[0137] The calculation formula is as follows:

[0138]

[0139] in, Let f represent the complex amplitude spectrum of the j-th type of noise at frequency f and time t; This represents the frequency weighting function, used to represent the frequency bands that the human ear is sensitive to; represents the set of characteristic frequency bands identified by the spectrum embedded density peak clustering algorithm for the j-th type of noise; T represents the total number of frames in the current noise reduction analysis period.

[0140] The calculation formula is as follows:

[0141]

[0142] in, This represents the average sound pressure level (dB) inside the earcup under the influence of Class j noise before the noise reduction strategy is implemented. This indicates the residual sound pressure level after applying noise reduction strategies; This represents the user's comfort rating at time t for the current noise cancellation state; This represents a comfort reference value for the same user in either silent or baseline conditions. This represents the system's preset weighting coefficient for comfort and sound pressure control; P indicates that it is consistent with the average sampling time window of sound pressure and feedback.

[0143] The calculation formula is as follows:

[0144]

[0145] in, This represents the control command signal value that the controller outputs to the active noise reduction unit at time q when dealing with type j noise. It can be a voltage, current, or frequency adjustment parameter. Q represents the rate of change of the control output over time. A larger value indicates that the control signal is unstable and may cause sudden changes in the user's hearing. Q represents the time step of the current analysis cycle. The larger the value, the stronger the control path fluctuation of the strategy, and the greater the impact on user comfort.

[0146] S4: Based on the optimal noise reduction strategy matrix, an adaptive PID controller is used to assign control commands to multiple active noise reduction units inside the earcups, thereby adjusting the multi-channel noise reduction execution module and adaptive noise suppression; the multi-channel noise reduction execution module integrates multiple sets of digital signal processing units, each set of units corresponding to different acoustic regions, and can perform independent signal modulation and amplitude and phase adjustment according to the commands output by the adaptive PID controller.

[0147] Specifically, the earcups integrate multiple active noise cancellation units (ANC units), each covering different acoustic areas: the forward sound field channel (close to the user's mouth); the rear sound field channel (the back of the earcup); the vertical sound field channel (the upper and lower sides); and the ear canal proximal channel (close to the eardrum).

[0148] Each unit includes: a high-sensitivity microphone for real-time monitoring of residual noise in the target area; a miniature low-distortion speaker for emitting anti-phase sound waves; a built-in digital signal processor (DSP) for signal modulation and phase control; and an adaptive PID control module that accepts policy matrix instructions for local parameter tuning.

[0149] Specifically, based on the optimal noise reduction strategy matrix, control commands are issued channel by channel. Each channel corresponds to a local active noise reduction region (such as the front, lower, or rear side) and is bound to a corresponding adaptive PID controller. The controller's task is to continuously adjust the output acoustic characteristics of the current channel to precisely match the target strategy requirements.

[0150] The specific process is as follows:

[0151] The controller first receives the current sensor data for that channel, including local sound pressure changes, sound source direction estimation results, vibration response amplitude, and other information, to determine the noise intensity and characteristic type faced by that channel.

[0152] Based on the labels such as "Target Control Level", "Response Speed ​​Level", and "Priority Strategy Type" corresponding to the channel in the strategy matrix, the preset PID parameter template group (such as fast response mode, smooth response mode, or low interference mode) is called.

[0153] During controller execution, if a sudden environmental change (such as a sudden impact sound) is detected, the system will trigger the dynamic gain adjustment module to temporarily increase the weight of the proportional channel and shorten the adjustment lag time; conversely, when the noise enters a stable state, the response sensitivity will be reduced again to reduce comfort interference.

[0154] The control system monitors the status of each channel via a bus. If the response of a certain area is insufficient (such as output lag or phase cancellation failure), it can schedule neighboring channels to share some strategy coefficients to achieve neighbor redundancy compensation.

[0155] Through the above scheduling strategy, adaptive PID control not only has targeted local control capabilities, but also maintains global noise reduction coordination in complex sound field environments, taking into account both user experience and response accuracy.

[0156] Specifically, the DSP digital signal processing unit in each noise reduction channel is responsible for receiving control commands and converting them into specific sound wave synthesis control signals to drive the speaker units in the corresponding area to emit canceling sound waves. Its core task is to perform amplitude modulation and phase adjustment on the output signal to achieve spatial interference with the noise signal.

[0157] The specific execution process is as follows:

[0158] The DSP digital signal processing unit first analyzes the original noise waveform of the current channel and extracts its characteristic data such as main frequency, waveform shape, and rate of change.

[0159] Based on the noise reduction target level set by the control strategy, the energy of the output sound wave is modulated point by point. For high-energy impact noise, the system prioritizes increasing the output amplitude to ensure effective cancellation of interference waves; while for low-frequency background noise, the output amplitude is controlled to decrease gradually to avoid generating new interference waves.

[0160] The DSP digital signal processing unit controls the phase advance or delay of the output signal based on the required "interference point" and "current sound wave propagation path" to ensure that the sound wave forms anti-phase interference with the noise at the user's eardrum. If reflected or diffracted noise occurs in the environment, the system will also correct the phase response curve to enable the active sound wave to be predicted in advance.

[0161] Due to the time difference and path difference between multiple channels, the DSP digital signal processing unit performs a synchronization calibration process every certain time period to match and correct the output waveform with the input noise waveform in real time, so as to avoid superposition interference between multiple channels.

[0162] Through the above process, the system not only ensures that the energy and phase of the output sound wave are precisely matched with the external noise, but also maintains the stability of the interference effect in complex multi-sound field environments, thereby achieving a higher level of auditory noise reduction experience.

[0163] During system operation, the adaptive PID controller not only issues control commands but also possesses continuous learning and optimization capabilities. The system establishes multiple feedback channels to monitor and correct the current noise reduction effect and user experience in real time, forming a dynamic closed loop.

[0164] The feedback mechanism consists of the following three parts:

[0165] (1) Real-time monitoring and feedback: Each channel is embedded with a sensor module to monitor the local sound pressure change trend after noise reduction; the system continuously evaluates core indicators such as "noise residual intensity", "suppression speed" and "sound balance". Once it is found that a certain channel does not reach the set response effect, it will automatically trigger gain fine-tuning, modulation parameter compensation or strategy switching.

[0166] (2) User interaction feedback: The system supports users to quickly provide feedback on the current noise cancellation experience through the APP or the earcup button, such as "poor noise cancellation effect", "uncomfortable ear pressure", "voice masking", etc. Each feedback will be packaged into an event sample and recorded together with the current control parameters, noise status and environmental characteristics for subsequent model optimization reference.

[0167] (3) Intelligent memory and adjustment: The system has a built-in short-term memory cache module to compare the feedback results with the control response each time; if the same problem occurs repeatedly in a certain type of noise situation, the system will automatically call the optimized parameter combination in the next similar environment; the controller parameter adjustment process adopts a progressive interpolation strategy to avoid user discomfort or system oscillation caused by sudden changes.

[0168] The above triple feedback system constructs a stable, closed-loop, and evolvable noise reduction control structure, which not only improves response flexibility but also enhances the system's adaptability to complex environments and individual differences.

[0169] S5: Based on historical noise reduction effects and user comfort feedback, the noise reduction parameters and control strategies are dynamically adjusted using a Bayesian optimization algorithm, and the optimization results are fed back to update the noise reduction decision model and control commands; the Bayesian optimization algorithm takes maximizing long-term user comfort as the objective function, combines noise identification results to dynamically select different parameter adjustment strategy families, and outputs an updated set of control parameter candidates.

[0170] Specifically, during the continuous operation of the smart protective earmuffs, the system will collect multi-dimensional feedback information from users in real time or periodically, using this information as input for optimization and updates, including but not limited to the following data:

[0171] Noise reduction performance metrics include: the difference in sound pressure level before and after the change, the change in spectral energy, and the noise residual rate.

[0172] User comfort data is obtained through physical feedback from the user (such as temperature sensors detecting temperature rise inside the earcups) and user interface ratings (such as discomfort buttons and comfort scores).

[0173] Power consumption recording: The system records the drive current and battery consumption rate of each noise reduction unit;

[0174] Noise scene label: The noise type identification result from step S2, used as an optimization background constraint.

[0175] After collecting sufficient data samples, the system enters the optimization phase:

[0176] Parameter space construction: The system constructs a multi-dimensional parameter space from all adjustable parameters, such as the gain coefficient of each channel, filter order, response delay threshold, adaptive control range, etc.

[0177] Strategy family division: Based on historical records, the noise reduction strategy is divided into several sub-strategy families, such as "high response and low power consumption", "low interference and high comfort" and "periodic noise priority processing". The optimization objective can be dynamically switched between different strategy families.

[0178] The Bayesian optimization process specifically includes: the system automatically selects a set of the most representative control parameter combinations within the policy family based on the current environment identification results and feedback trends; when the control model is unclear or the effect of policy changes is difficult to predict, the system adopts an exploration mechanism based on the acquisition function to prioritize parameter combinations with higher expected returns; the optimization mechanism can continuously iterate and update the parameter selection strategy without explicit gradient information.

[0179] The optimization results are written to the local parameter cache and the current policy model is structurally updated to enhance the model's adaptability to similar environments. At the same time, when the model fails to effectively deal with noise features, the system automatically triggers "exploration mode" to try previously unsampled policy spaces, thereby improving the overall robustness of the system.

[0180] For users who wear the device for extended periods, the system automatically creates an individual user profile, recording their comfort thresholds and response preferences in different sound environments. When users switch or the usage environment changes, the system can call upon existing optimization results as a basis for migration, quickly adapting to strategic decisions in new situations and reducing the initial period of discomfort.

[0181] In summary, this invention constructs a multi-array sensor network, integrating environmental microphones, bone conduction sensors, structural vibration sensors, and internal and external sound pressure sensors to achieve high-precision, full-frequency coverage acquisition of multi-source acoustic data. This multi-channel synchronous sampling method effectively improves the ability to identify the direction, frequency band, and type of sound sources, especially demonstrating significant advantages in separating multi-source signals in mixed noise environments. The invention also introduces a spectral embedded density peak clustering method, constructing a spectrogram feature space to extract high-dimensional feature factors such as energy centroid and spectral entropy, and combining this with graph regularization constraints to achieve accurate identification of different noise types (such as speech, structured sound, and background noise). This adaptive clustering and classification mechanism overcomes the limitations of traditional algorithms that rely on a preset number of categories, significantly improving the model's recognition accuracy and adaptability.

[0182] This invention, based on a near-end policy optimization algorithm, constructs an environmental state vector by combining real-time noise type, energy distribution, and spatial labels. Multiple policy candidate sets are generated through reinforcement learning, and a hierarchical control mechanism is introduced to achieve differentiated suppression of transient, impulsive, and periodic noise, improving response speed and noise reduction accuracy. Furthermore, user-personalized parameters (such as earmuff sealing, ear canal reflection characteristics, and historical feedback) are integrated to weight and correct the policy candidate sets, ensuring that the noise reduction strategy achieves physical suppression while also considering user comfort, effectively reducing ear pressure discomfort and sudden changes in hearing. An adaptive PID controller independently adjusts the multi-channel noise reduction units, achieving precise amplitude and phase matching and coordinated control of the multi-region acoustic field. The system also incorporates a Bayesian optimization algorithm, using long-term user comfort as the objective function to dynamically optimize the control strategy, improving the overall system's robustness and adaptability.

[0183] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligent noise control of protective earmuffs, characterized in that, Includes the following steps: S1: Acquire multi-source acoustic data of the protective earmuffs through a multi-array sensor network; S2: Based on the multi-source acoustic data, the key noise feature factors are extracted and the noise type is identified using the spectral embedded density peak clustering algorithm to generate noise analysis results; S3: Based on the noise analysis results, construct a noise reduction decision model based on the near-end strategy optimization algorithm, dynamically evaluate the noise reduction effect and comfort index through the reward function, and generate the optimal noise reduction strategy matrix for transient noise, impact noise and periodic noise in real time. S4: Based on the optimal noise reduction strategy matrix, an adaptive PID controller is used to assign control commands to multiple active noise reduction units inside the earmuff, thereby realizing the adjustment of the multi-channel noise reduction execution module and adaptive noise suppression. S5: Based on historical noise reduction effects and user comfort feedback, dynamically adjust noise reduction parameters and control strategies using a Bayesian optimization algorithm, and feed back the optimization results to update the noise reduction decision model and control commands.

2. The intelligent noise control method for protective earmuffs according to claim 1, characterized in that, The multi-source acoustic data includes external environmental noise signals, user voice signals, earmuff structure vibration sound signals, bone conduction sound signals, and sound pressure signals inside and outside the earmuff cavity.

3. The intelligent noise control method for protective earmuffs according to claim 1, characterized in that, Step S2 includes the following steps: The multi-source acoustic data is normalized, and a unified acoustic time-spectrum diagram is constructed using short-time Fourier transform to extract the energy distribution characteristics and amplitude spectrum characteristics within the corresponding time period. The acoustic time-spectrum graph is input into a spectrum-embedded density peak clustering algorithm to identify potential noise-dense regions by utilizing the relationship between local density and distance, and to extract key noise feature factors that include noise intensity, frequency band characteristics, and instantaneous amplitude. Based on feature factors, a multi-class discrimination model based on graph regularization constraints is constructed to classify and identify external background noise, structural vibration noise, bone conduction sound and user speech. The classification results are assigned weight labels based on recognition confidence, and information on noise category, sound source spatial label, and feature intensity is integrated to generate noise analysis results.

4. The intelligent noise control method for protective earmuffs according to claim 3, characterized in that, The construction process of the multi-class discriminant model includes the following steps: A set of feature vectors is constructed based on the key noise feature factors, and the similarity matrix between feature samples is calculated; Based on the set of feature vectors, a feature graph structure is constructed using a graph regularization constraint strategy. The similarity matrix is ​​then subjected to spectral embedding mapping using the Laplacian operator to generate a low-dimensional feature subspace. Labeled training samples are introduced into the low-dimensional feature subspace to construct a multi-class discrimination model. The objective is jointly optimized by the minimum graph regularization loss and the maximum inter-class discrimination divergence. The multi-class discrimination mapping function is solved by iterative training to achieve the classification of background noise, structural noise, bone conduction sound and speech signals. The graph structure parameters and regularization coefficients are fine-tuned through cross-validation to complete the training of the multi-class discrimination model, and finally output noise analysis data containing classification labels, confidence coefficients and noise type mapping results.

5. The intelligent noise control method for protective earmuffs according to claim 4, characterized in that, The formula for the multi-class discriminant model is as follows: in, This represents the final output of the multi-class discrimination, i.e., the noise category label to which the final judgment belongs; k represents the index of the k-th noise type; K represents the total number of noise types; N represents the number of key noise feature factors extracted in the current frame; This represents the response intensity of the i-th key noise feature factor on the k-th noise model; This represents the response intensity of the i-th key noise feature factor on the l-th noise model; This represents the average acoustic energy value of the i-th feature factor in the k-th class of training samples; This represents the overall average acoustic energy of the i-th feature factor across all training categories; This represents the energy standard deviation of the i-th feature factor across all categories; This represents a small positive constant introduced to prevent the denominator from being zero; This represents the confidence label value of the i-th feature factor in the k-th class; This represents the embedding distance between the i-th feature and the center of the k-th noise subspace; The influence of the response normalization ratio, the spectral energy difference term, and the spatial distance confidence term are controlled separately.

6. The intelligent noise reduction control method for protective earmuffs according to claim 1, characterized in that, Step S3 includes the following steps: The noise analysis results are processed to extract the dominant noise type, characteristic frequency band, energy weight, spatial label and its identification confidence in the current period, and a dynamic environmental state vector is constructed by combining it with historical identification data. The environmental state vector is input into the noise reduction decision model constructed based on the near-end policy optimization algorithm, with the joint optimization objectives of minimizing noise residue, improving response speed and maintaining user comfort, to generate a set of policy candidates; During the strategy generation process, a hierarchical control framework for multiple types of noise is constructed, a rapid response priority is set for transient noise, a short-term burst suppression mechanism is enabled for impact noise, and a predictive intervention strategy is introduced for periodic noise. By combining the candidate strategy set with user-personalized parameters, the strategy results are differentiated and weighted to obtain the optimal noise reduction strategy matrix for the current noise scenario. The optimal noise reduction strategy matrix is ​​then used as an output signal and transmitted to the control execution layer. The user-personalized parameters include earmuff tightness, ear canal reflection characteristics, adaptive historical parameters, and subjective comfort feedback level.

7. The intelligent noise control method for protective earmuffs according to claim 6, characterized in that, The formula for the noise reduction decision model is as follows: in, This represents the policy score corresponding to the j-th type of noise; This represents the characteristic frequency band energy value of the j-th type of noise in the current time period; This represents the amplitude spectrum activation coefficient corresponding to the j-th type of noise; This indicates the degree of spatial disturbance of the noise within the current period; This represents the average noise reduction gain of the periodic strategy response to the j-th type of noise. The time-delay sensitivity parameter represents the strategy corresponding to the j-th type of noise; This represents the system's response delay to type j noise when the current strategy is executed; This represents the control path gradient under the j-th type of noise; This represents the intensity of false triggering or vibration interference under the influence of type j noise interference. This represents a stability constant set to prevent the denominator from becoming too small; These represent the weights of the acoustic driving factor, historical response factor, and control stability factor on the scoring function, respectively.

8. The intelligent noise reduction control method for protective earmuffs according to claim 1, characterized in that, The multi-channel noise reduction execution module integrates multiple digital signal processing units, each corresponding to a different acoustic region, and can perform independent signal modulation and amplitude and phase adjustment according to the instructions output by the adaptive PID controller.

9. The intelligent noise control method for protective earmuffs according to claim 1, characterized in that, The Bayesian optimization algorithm takes maximizing long-term user comfort as its objective function, dynamically selects different families of parameter adjustment strategies based on noise identification results, and outputs an updated set of candidate control parameters.

Images