Range hood semi-adaptive active noise reduction method with self-checking function

Abstract

The invention discloses a range hood semi-adaptive active noise reduction method with a self-checking function, which is applied to a range hood semi-adaptive active noise reduction system, and the range hood semi-adaptive active noise reduction system comprises an image acquisition device, an electroacoustic device and a controller. In the pre-training stage, a device position self-checking vector expression library, a device oil stain self-checking vector expression library, a range hood sound field database, a secondary channel estimation function library, a controller coefficient library, a range hood sound field feature library and a range hood sound field vector expression library are established respectively; in the actual noise reduction stage, firstly, self-inspection of the position and the oil stain state of the electroacoustic device is completed through the image acquisition equipment, and noise reduction is interrupted when abnormity occurs. Otherwise, the sound field environment type of the audio signal frame of the range hood is judged in real time, corresponding parameters are selected from the secondary channel estimation function library and the controller coefficient library for active noise reduction and parameter fine adjustment, the problems that a complete self-adaptive system is slow in convergence and prone to howling are solved, and the defect that a fixed parameter system is difficult to adapt to environment changes is overcome.

Description

Technical Field

[0001] This invention belongs to the field of noise reduction technology for range hoods, specifically relating to a semi-adaptive active noise reduction method for range hoods with self-testing function. Background Technology

[0002] Range hoods are indispensable appliances in modern family kitchens. Installed above the stove, they quickly remove cooking fumes and gas waste, preventing harm to the human body. However, the powerful suction fans that ensure effective smoke extraction often generate significant noise, and prolonged exposure to such noise can lead to hearing loss or even deafness. Therefore, quiet range hoods are undoubtedly a key focus of research and development for major range hood manufacturers. Given that most range hood noise is broadband below 1kHz, more and more manufacturers are exploring active noise reduction technology.

[0003] Active noise cancellation is now widely used in the headphone industry. However, in the range hood industry, it is severely affected by factors such as wind noise and grease, and there is still much room for improvement.

[0004] For example, in terms of system adaptation:

[0005] Existing patent CN108954442 A describes a range hood with a noise reduction device and a noise reduction method, comprising a range hood body and a three-dimensional spatial sound field noise reduction device for active noise reduction. The noise reduction device of this invention employs a fully adaptive active noise reduction algorithm, which iterates parameters autonomously in real time according to changes in the range hood's sound field. This algorithm involves a large computational load, extremely high hardware requirements, and high cost. This method suffers from drawbacks such as difficulty in convergence and a high tendency to diverge, potentially leading to feedback and negatively impacting user experience.

[0006] Existing patent CN116105191A describes a range hood and its noise reduction method, system, and storage medium. This method acquires a target sampling rate and target sound velocity, collects the actual sound velocity corresponding to the current speed setting, determines the current sampling rate corresponding to the current speed setting based on the target sound velocity, target sampling rate, and actual sound velocity, and uses preset corresponding noise reduction parameters to reduce noise in the range hood. It employs a fixed coefficient approach, which is simple to calculate, low-cost, and avoids system divergence issues. However, the actual sound field environment of a range hood is complex and variable. Determining the sound field type of the range hood solely based on non-audio acquisition equipment makes it difficult to achieve good noise reduction results.

[0007] For example, in terms of device self-testing:

[0008] Existing patent CN117831493A describes a scene-adaptive active noise reduction method and system. This method trains a scene classification model using reference microphone signals and multimodal sensor signals. If the scene classification model outputs a preset scene type and its influencing factors, it then searches an ANC parameter table based on these parameters to determine the ANC parameters for the ANC control module. While this method considers various range hood scenarios, it lacks a device self-test function. The active noise reduction system's components are placed inside the range hood cavity, essentially forming a black box. The system's effectiveness depends on the proper positioning of all components and the fact that the grease or grime on the components does not affect the noise reduction effect.

[0009] Existing patent CN108916944B describes a range hood with noise reduction and visual detection functions, and a noise reduction method thereof. The range hood includes a main body, a three-dimensional spatial sound field noise reduction device for active noise reduction, and a visual detection system. The visual detection system for monitoring smoke is mounted on the main body of the range hood. However, the visual detection system in this method is not located within the interior cavity of the range hood; it is only used to detect the smoke concentration from external kitchen appliances and cannot monitor the components of the active noise reduction system located within the interior cavity of the range hood.

[0010] Existing patent CN116132900B describes a self-testing method, device, and medium for an active noise cancellation system. It includes at least one microphone and at least one speaker. The speaker plays its corresponding preset test signal, and the microphone simultaneously acquires at least one target audio signal, extracts the target spectrum, and determines whether the device is functioning correctly. This self-testing method is low-cost, but it requires a quiet environment where the user is not using the range hood, placing high demands on the external environment. Furthermore, the testing process generates noise, causing additional disturbance to the user. Additionally, relying solely on audio signal detection only provides a simple assessment of the presence or absence of grease, making accurate diagnosis difficult. In practical applications, grease can be categorized into viscous / heavy grease and clear / light grease. The impact mechanisms and severity of these two types of grease on electroacoustic devices are drastically different. Existing audio self-testing methods cannot distinguish between these two different physical forms and impact mechanisms, thus failing to provide targeted maintenance guidance (such as determining whether cleaning or replacement of the grease filter is needed), and making it difficult to provide early warnings in the initial stages of performance degradation. Summary of the Invention

[0011] The purpose of this invention is to address the problems raised in the background art by proposing a semi-adaptive active noise reduction method and device for range hoods with self-testing function.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] This invention proposes a semi-adaptive active noise reduction method for range hoods with self-testing function, applicable to a semi-adaptive active noise reduction system for range hoods. The system includes an image acquisition device, electroacoustic components, and a controller. The electroacoustic components include a reference microphone, a secondary loudspeaker, and an error microphone. The self-testing method for semi-adaptive active noise reduction in range hoods includes:

[0014] Pre-training phase:

[0015] A device position self-check vector expression library, a device oil stain self-check vector expression library, a range hood sound field database, a secondary channel estimation function library, a controller coefficient library, a range hood sound field feature library, and a range hood sound field vector expression library are established respectively. The range hood sound field database contains M-type sound field environments where the electroacoustic devices are in normal positions. The secondary channel estimation function library contains M-type secondary channel estimation functions that correspond one-to-one with each of the M-type sound field environments. The controller coefficient library contains M-type controller coefficients that correspond one-to-one with each of the M-type sound field environments. The range hood sound field vector expression library contains M-type vector expressions that correspond one-to-one with each of the M-type sound field environments.

[0016] Actual noise reduction stage:

[0017] At the moment the range hood is turned on, the current image frame of all electroacoustic devices is acquired by the image acquisition device. The position type of each electroacoustic device in the current image frame is determined according to the device position self-check vector expression library. When the position of an electroacoustic device is abnormal, the operation of the range hood's semi-adaptive active noise reduction system is interrupted.

[0018] When the position of the electroacoustic device is normal, the oil stain status of the electroacoustic device in the current image frame is determined according to the device oil stain self-check vector expression library. When the oil stain status of the electroacoustic device is abnormal, the operation of the range hood semi-adaptive active noise reduction system is interrupted.

[0019] When the oil stains on the electroacoustic device are in normal condition, the current range hood audio signal frame collected in real time by the reference microphone is obtained. The features corresponding to the range hood audio signal frame are extracted according to the range hood sound field feature library, and the extracted features are normalized. Then, the normalized features are spliced ​​together to obtain the spliced ​​features.

[0020] The splicing features of the current range hood audio signal frame are input into the trained range hood noise classification model to obtain the vector representation of the current range hood audio signal frame. The similarity of the vector representation of the current range hood audio signal frame with the range hood sound field vector representation library is compared, and the sound field environment type of the current range hood audio signal frame is determined by combining the oil stain status of the electroacoustic devices in the current image frame.

[0021] Based on the sound field environment type of the current range hood audio signal frame, the corresponding secondary channel estimation function and controller coefficient are selected from the preset secondary channel estimation function library and controller coefficient library, respectively.

[0022] The controller uses an adaptive algorithm to calculate an anti-phase sound wave signal that is out of phase with the current range hood audio signal frame based on the selected secondary channel estimation function and controller coefficients, as well as the current range hood audio signal frame collected by the reference microphone and the residual noise collected by the error microphone. The calculated anti-phase sound wave signal is then transmitted to the secondary speaker for emission, thereby achieving noise reduction.

[0023] The noise reduction effect of a semi-adaptive active noise reduction system is monitored based on the residual noise collected by the error microphone.

[0024] Preferably, in the process of establishing the device position self-check vector expression library, multiple first image frames containing all electroacoustic devices are acquired using an image acquisition device. All first image frames are divided into two categories according to whether the electroacoustic device position is abnormal: normal electroacoustic device position and abnormal electroacoustic device position. Each category contains W first image frames, and each first image frame is labeled with the label and number of the electroacoustic device position category.

[0025] For each first image frame in each class, a vector representation is generated. Then, the average of the vector representations of W first image frames in each class is calculated to obtain the vector representations of the first image frames in the two classes. The vector representations of the first image frames in the two classes constitute the device position self-test vector representation library.

[0026] Preferably, the oil contamination status of the electroacoustic device includes the type of oil contamination and the adhesion status;

[0027] In the process of establishing the device oil contamination self-inspection vector expression library, multiple second image frames containing all electroacoustic devices were acquired using image acquisition equipment. All second image frames were divided into H categories according to oil contamination type and adhesion state, with each category containing K second image frames. Second image frames of the same category had the same oil contamination type and adhesion state. Each second image frame was labeled with the label and number of the electroacoustic device oil contamination state.

[0028] For each second image frame in each class, a vector representation is generated. Then, the mean of the vector representations of K second image frames in each class is calculated to obtain the vector representations of the second image frames of class H. The vector representations of the second image frames of class H constitute the device oil stain self-inspection vector representation library.

[0029] Preferably, each type of sound field environment in the range hood sound field database contains J range hood data. A reference microphone is used to collect I consecutive frames of range hood audio signal as one range hood data. Each range hood data is then calibrated to obtain calibration information, which includes the sound field environment type, the oil stain status of the electroacoustic devices, the fan speed, and the serial number.

[0030] Preferably, the establishment of the secondary channel estimation function library and the controller coefficient library includes:

[0031] For the sound field environment corresponding to each range hood data, a secondary channel identification algorithm is used to identify the secondary channel and obtain the secondary channel estimation function. The secondary channel estimation function is substituted into the semi-adaptive active noise reduction system, and an adaptive algorithm is used for active noise reduction. When the semi-adaptive active noise reduction system reaches the best noise reduction effect, the secondary channel estimation function and controller coefficient corresponding to the current range hood data are recorded.

[0032] For the M-class sound field environment, the secondary channel estimation functions of the J range hood data in each class are summed and averaged to obtain the secondary channel estimation functions of each type of sound field environment, and the secondary channel estimation functions of the M-class range hood data are used to form a secondary channel estimation function library;

[0033] For the M-type sound field environment, the controller coefficients of the J range hood data in each type are summed and averaged to obtain the controller coefficients of each type of sound field environment, and the controller coefficients of the M-type range hood data are used to form a controller coefficient library.

[0034] Preferably, establishing the sound field feature library of the range hood includes:

[0035] For each range hood data point in the range hood sound field database, the following 15 features are calculated for each range hood audio signal frame: zero crossing rate, short-time amplitude, energy, absolute mean, root mean square value, variance, standard deviation, kurtosis, skewness, peak index, peak factor, margin coefficient, spectral centroid, mean square frequency, and frequency variance. Thus, 15 features with a length of I frames are obtained for each feature.

[0036] Calculate the variance of each of the fifteen features of length I-frame. Select the ten features with the highest variance from the fifteen features, and then use the Pearson correlation coefficient to calculate the correlation between any two features among these ten features:

[0037] When the correlation between two features is less than a preset value, it means that the correlation between the two features is low, and both features are retained.

[0038] When the correlation between two features is greater than or equal to a preset value, it indicates that the correlation between the two features is high, and the feature with the larger variance is selected from the two features for retention.

[0039] Ultimately, only five features are retained, referred to as the five main features: short-term amplitude, kurtosis, peak index, margin coefficient, and spectral centroid. For each I-frame of range hood audio signal in the range hood sound field database, the five main features with a length of I frames are retained.

[0040] Normalize each frame of the five main features to obtain the five main features with a normalized length of I frames;

[0041] The five main features of the range hood sound field are normalized to a length of I frames and constitute the range hood sound field feature library.

[0042] Preferably, in the pre-training stage, the range hood noise classification model is trained using all range hood sound field features in the range hood sound field feature library to obtain a trained range hood noise classification model, and the range hood noise classification model is a deep neural network.

[0043] During the pre-training phase, for each range hood data point in each sound field environment, five normalized main features of length I frames corresponding to the current range hood data point in the current sound field environment of the current category are extracted from the range hood sound field feature library. The five normalized main features are then concatenated and input into the trained range hood noise classification model. In the feature layer of the trained range hood noise classification model, the vector representation of the current range hood data point in the current sound field environment of the current category is output. The mean of the J vector representations of the current sound field environment of the current category is calculated to obtain the vector representation of the current sound field environment of the current category. The vector representations of the M sound field environments constitute the range hood sound field vector representation library.

[0044] Preferably, the monitoring of the noise reduction effect of the semi-adaptive active noise reduction system based on the residual noise collected by the error microphone includes:

[0045] The residual noise e(n) and its mean square value are collected in real time using an error microphone. The calculation formula is as follows:

[0046] ;

[0047] Among them, when the range hood does not produce a whistling sound, the mean square value of the residual noise is... If the preset number of seconds remains unchanged and is less than the first threshold, the semi-adaptive active noise cancellation system will achieve the best noise reduction effect. This represents the nth number in the residual noise e(n);

[0048] When the amplitude of the residual noise exceeds the second threshold, it is determined that the range hood is whistling and the noise reduction is immediately interrupted.

[0049] Preferably, the process of determining the position type of each electroacoustic device in the current image frame based on the device position self-test vector expression library is as follows: normalize the current image frame, express the position of the electroacoustic device in the normalized current image frame as a vector, compare the similarity of the current image frame's electroacoustic device position vector expression with the vector expression in the device position self-test vector expression library, and take the device position type corresponding to the vector expression with the highest similarity in the device position self-test vector expression library as the electroacoustic device position type of the current image frame;

[0050] The process of determining the oil contamination status of the electroacoustic device in the current image frame based on the device oil contamination self-inspection vector expression library is as follows: the normalized current image frame is expressed as a vector of the device oil contamination status; the similarity of the vector of the current image frame's electroacoustic device oil contamination status with the vector expression in the device oil contamination self-inspection vector expression library is compared; and the oil contamination status corresponding to the vector expression with the highest similarity in the device oil contamination self-inspection vector expression library is taken as the current image frame's electroacoustic device oil contamination status.

[0051] The process of comparing the similarity of the vector representation of the current range hood audio signal frame with the range hood sound field vector representation library, and determining the sound field environment type of the current range hood audio signal frame in conjunction with the oil stain status of the electroacoustic devices in the current image frame, includes: first, reducing the range hood sound field vector representation library based on the oil stain status of the electroacoustic devices in the current image frame; then, comparing the similarity of the vector representation of the current range hood audio signal frame with the reduced range hood sound field vector representation library, obtaining the sound field environment corresponding to the vector representation with the highest similarity in the reduced range hood sound field vector representation library, and taking this sound field environment as the sound field environment type of the current range hood audio signal frame.

[0052] Preferably, the semi-adaptive active noise reduction method for range hoods with self-testing function further includes: during the actual noise reduction stage, if the sound field environment type of the current range hood audio signal frame is the same as that of the previous range hood audio signal frame, then there is no need to reselect the corresponding secondary channel estimation function and controller coefficients. The controller coefficients are fine-tuned using an adaptive algorithm based on the noise reduction result, and the calculation of the inverted sound wave signal continues; if, starting from the current range hood audio signal frame, the sound field environment type of the range hood audio signal frames greater than or equal to Q consecutive frames is different from that of the previous range hood audio signal frame, then the corresponding secondary channel estimation function and controller coefficients need to be reselected.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. This method constructs a secondary channel estimation function library and a controller coefficient library, so that the actual noise reduction process selects the corresponding secondary channel estimation function and controller coefficient from the preset secondary channel estimation function library and controller coefficient library according to the sound field environment type of the current range hood audio signal frame. This avoids the slow convergence and howling problems caused by the fully adaptive system iterating from scratch, overcomes the defect that the fixed parameter system cannot adapt to environmental changes, and implements anti-howling processing to improve the user experience.

[0055] 2. This method features an upgraded self-inspection function for device oil contamination status, solving the problem that relying solely on audio signal detection can only provide a simple assessment of the presence or absence of oil contamination and is insufficient for accurate diagnosis. Since viscous and transparent oil contamination have different physical mechanisms of impact on electroacoustic devices, directly affecting the failure modes and maintenance strategies of the noise reduction system, this method uses visual methods to perform refined identification of oil contamination types and their adhesion states. This can be extended to assess pipeline contamination, enabling more precise and forward-looking performance evaluation and maintenance warnings for the noise reduction system's operating environment, greatly improving the system's long-term reliability and user experience.

[0056] 3. This method innovatively integrates visual and auditory information for sound field recognition. By performing a visual self-check of device oil stains at startup, prior information about the oil stain status is obtained, narrowing down the scope of the range hood sound field vector representation library. Based on this, the vector representation of the current range hood audio signal frame is compared with the narrowed range hood sound field vector representation library for similarity, thereby significantly reducing unnecessary calculation comparisons and making the judgment of sound field environment type faster and more accurate. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the semi-adaptive active noise reduction method and device for a range hood with self-testing function according to the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0060] like Figure 1As shown, a semi-adaptive active noise reduction method for range hoods with self-checking function is provided. This method is applied to a semi-adaptive active noise reduction system for range hoods, which includes an image acquisition device, electroacoustic components (including a reference microphone, secondary loudspeaker, and error microphone), and a controller. The image acquisition device is positioned near the range hood's smoke inlet. This position is the result of field-of-view (FOV) analysis and optimization. By selecting a lens with an appropriate focal length and adjusting its installation angle, it is ensured that a single frame image captured at this position can completely cover all electroacoustic components and their sound-transmitting structures (sound-transmitting mesh or grille structure). This layout design aims to achieve synchronous, global visual monitoring of the core components of the entire noise reduction system with minimal hardware cost (single camera), providing reliable and comprehensive image input for subsequent self-checking of component positions and determination of oil stain status. The reference microphone is typically located inside the range hood's cavity, and is considered to be affected only by the range hood's own noise. The error microphone is usually located near the range hood's smoke inlet, and is considered to represent the final noise emitted from the range hood. The secondary speaker is positioned between the reference and error microphones. All electroacoustic components undergo oil-proofing treatment to protect their sound-transmitting structures. Specifically:

[0061] Reference microphone and error microphone: The dustproof mesh or microphone mesh at the sound inlet must be a sound-transmitting structure to ensure that sound waves can enter smoothly.

[0062] Secondary loudspeakers: The sound outlet grille and the protective mesh at the front of the diaphragm are sound-transmitting structures, which must ensure the effective radiation of anti-phase sound waves.

[0063] The aforementioned acoustically transparent structure can be treated with oleophobic coatings, replaceable oil-proof mesh covers, etc., to reduce the adhesion and accumulation of oil at the physical level, thereby mitigating the deteriorating effect of oil on the acoustic sensitivity, frequency response, and other performance of the device, and ensuring the long-term effectiveness of the noise reduction system.

[0064] The semi-adaptive active noise reduction method for range hoods with self-testing function includes (the overall process of this method is as follows) Figure 1 As shown in the diagram, the process mainly includes two stages: pre-training and actual noise reduction. The following section provides a detailed explanation of each step in conjunction with the diagram.

[0065] Step 1, Pre-training Phase:

[0066] Establish multiple pre-training libraries: respectively establish a device position self-test vector expression library, a device oil stain self-test vector expression library, a range hood sound field database, a secondary channel estimation function library, a controller coefficient library, a range hood sound field feature library, and a range hood sound field vector expression library;

[0067] In step 1.1, during the process of establishing the device position self-check vector expression library, multiple first image frames containing all electroacoustic devices are acquired using an image acquisition device. All first image frames are divided into two categories, normal electroacoustic device position and abnormal electroacoustic device position, according to whether the position of the electroacoustic device is abnormal (such as skew, displacement, etc.). Each category contains W first image frames. Each first image frame is labeled with the label and number of the electroacoustic device position category (e.g., device position normal, number: 1). The electroacoustic device position category is the same for the W first image frames of the same category.

[0068] For each first image frame in each class, a vector representation is generated. The specific process is as follows: each first image frame is scaled to a size of 150x150, and then pixel values ​​are normalized from 0-255 to 0-1 to accelerate model convergence. The normalized first image frame is then input into the trained device location self-detection classification model. The vector representation of the first image frame is output from the flattening layer (after the last pooling layer and before the first fully connected layer) of the trained device location self-detection classification model. The mean of the vector representations of W first image frames in each class is then calculated to obtain the vector representations of the first image frames in both classes. These two vector representations constitute the device location self-detection vector representation library.

[0069] The training of the device location self-test classification model involves using all first image frames (with labels) of all categories to train the device location self-test classification model, which is a Convolutional Neural Network (CNN). The specific training process is as follows:

[0070] (1) Dataset preprocessing

[0071] The dataset uses the first image frame (with labels) of all categories. First, all first image frames are uniformly scaled to 150×150 pixels, and pixel values ​​are normalized (linearly scaling the range [0, 255] to the interval [0, 1] to accelerate model convergence). Then, a stratified sampling method is used to divide the dataset into training, validation, and test sets in a 7:2:1 ratio, used for model training, hyperparameter tuning, and final performance evaluation, respectively.

[0072] (2) CNN model structure

[0073] The device location self-check classification model is a custom convolutional neural network (CNN) with a hierarchical design. Its structure from input to output and key parameter settings are as follows:

[0074] Convolutional blocks: There are 3 blocks in total. Each block contains a convolutional layer and a pooling layer. The convolutional layers use 32, 64, and 128 3×3 convolutional kernels respectively, with a stride of 1, padding method of "same", and activation function of ReLU. The pooling layer is a 2×2 max pooling layer with a stride of 2.

[0075] Flattening layer: The output feature map of the last pooling layer is flattened into a one-dimensional vector, which is used as the vector representation of the first image frame.

[0076] Fully connected layer: A hidden layer containing 512 neurons with ReLU activation function, followed by a Dropout layer (dropout rate 0.5) to alleviate overfitting.

[0077] Output layer: The number of neurons is set to 2, which is the number of device location categories, and the activation function is Softmax.

[0078] (3) Model training configuration

[0079] Optimizer and loss function: The Adam optimizer is used (learning rate = 0.001, exponential decay rate of first-order momentum β1 = 0.9, exponential decay rate of second-order momentum β2 = 0.999), and the loss function is classification cross-entropy.

[0080] Training parameters: Batch size is set to 32, and training epochs are 50.

[0081] Regularization strategy: Early stopping is used to monitor the validation set loss. If the loss does not decrease for five consecutive training rounds, training is terminated early, and the model weights with the lowest validation set loss are automatically restored.

[0082] (4) Model compilation

[0083] Based on the aforementioned optimizer and loss function, and combined with evaluation metrics such as classification accuracy, precision, and recall, the CNN model is compiled, and the optimization goals and performance evaluation criteria for model training are clarified.

[0084] (5) Model training execution

[0085] The divided training and validation sets are input into the compiled CNN model, and the training process is started according to the set batch size and training epochs. During training, the changes in the validation set loss are monitored in real time, and training is terminated when the early stopping condition is triggered to ensure the generalization performance of the model.

[0086] (6) Model evaluation and storage

[0087] After training, the model's classification accuracy, precision, recall, and other performance metrics are calculated using the test set. The model with the lowest loss on the validation set is saved as the final device location self-testing classification model.

[0088] Step 1.2: The oil contamination status of electroacoustic devices includes oil contamination type and adhesion status. In establishing the device oil contamination self-check vector representation library, the classification criteria are first clarified. Since viscous oil contamination and transparent oil contamination have different physical impact mechanisms on the acoustic structure of electroacoustic devices (the former easily leads to blockage, while the latter mainly causes gradual changes in acoustic parameters), directly affecting the failure mode and maintenance strategy of the noise reduction system, it is necessary to refine the identification of oil contamination type and its adhesion status. Oil contamination types include viscous oil contamination and transparent oil contamination, and adhesion status includes clean, lightly covered, moderately covered, and heavily covered / blocked (where adhesion status in this embodiment refers to the adhesion status of the acoustic structure of the electroacoustic device).

[0089] Among them, sticky oil stains: have a viscous texture and are very easy to clog the microphone inlet or the speaker outlet mesh, directly and quickly blocking the sound wave channel, causing device failure or a sharp drop in performance.

[0090] Among them, clear oil stains: although they do not directly block, covering the surface of the diaphragm or mesh will change the acoustic quality and damping characteristics of the device, causing a gradual deterioration in performance such as frequency response drift and sensitivity reduction.

[0091] Based on this, multiple second image frames containing all electroacoustic devices (especially their acoustic structures) were acquired using an image acquisition device. All second image frames were categorized into H classes based on the type of oil contamination on the electroacoustic devices (mainly viscous and clear oil contamination) and their adhesion status to the acoustic structure (e.g., clean, lightly covered, moderately covered, heavily covered / blocked). Each class contained K second image frames with the same oil contamination type and adhesion status. Each second image frame was labeled with a tag and number indicating the oil contamination status of the electroacoustic device (e.g., viscous oil contamination - heavily blocked mesh, number: 2).

[0092] For each second image frame in each class, a vector representation is generated. The specific process is as follows: each second image frame is scaled to a size of 150x150, and then pixel values ​​are normalized from 0-255 to 0-1 to accelerate model convergence. The normalized second image frames are then input into the trained device oil contamination self-detection classification model. The vector representation of the second image frame is output from the flattening layer (after the last pooling layer and before the first fully connected layer) of the trained model. The mean of the vector representations of K second image frames in each class is calculated to obtain the vector representation of the second image frames in class H. The vector representations of the second image frames in class H constitute the device oil contamination self-detection vector representation library.

[0093] The training of the device oil contamination self-inspection classification model involves using all second image frames (with labels) of all categories to train the model, resulting in a trained model that can be used for feature extraction. This model is a Convolutional Neural Network (CNN). The specific training process is as follows:

[0094] (1) Dataset preprocessing

[0095] The dataset uses labeled second image frames from all categories. First, all second image frames are uniformly scaled to 150×150 pixels, and pixel values ​​are normalized (linearly scaling the range [0, 255] to the interval [0, 1] to accelerate model convergence). Then, a stratified sampling method is used to divide the dataset into training, validation, and test sets in a 7:2:1 ratio, used for model training, hyperparameter tuning, and final performance evaluation, respectively.

[0096] (2) CNN model structure

[0097] The device oil contamination self-inspection and classification model is a custom convolutional neural network (CNN) with a hierarchical design. Its structure from input to output and key parameter settings are as follows:

[0098] Convolutional blocks: There are 3 blocks in total. Each block contains a convolutional layer and a pooling layer. The convolutional layers use 32, 64, and 128 3×3 convolutional kernels respectively, with a stride of 1, padding method of "same", and activation function of ReLU. The pooling layer is a 2×2 max pooling layer with a stride of 2.

[0099] Flattening layer: The output feature map of the last pooling layer is flattened into a one-dimensional vector, which is used as the vector representation of the second image frame.

[0100] Fully connected layer: A hidden layer containing 512 neurons with ReLU activation function, followed by a Dropout layer (dropout rate 0.5) to alleviate overfitting.

[0101] Output layer: The number of neurons is set to H, the number of oily state categories in the device, and the activation function is Softmax.

[0102] (3) Model training configuration

[0103] Optimizer and loss function: The Adam optimizer is used (learning rate = 0.001, exponential decay rate of first-order momentum β1 = 0.9, exponential decay rate of second-order momentum β2 = 0.999), and the loss function is classification cross-entropy.

[0104] Training parameters: Batch size is set to 32, and training epochs are 50.

[0105] Regularization strategy: Early stopping is used to monitor the validation set loss. If the loss does not decrease for five consecutive training rounds, training is terminated early, and the model weights with the lowest validation set loss are automatically restored.

[0106] (4) Model compilation

[0107] Based on the aforementioned optimizer and loss function, and combined with evaluation metrics such as classification accuracy, precision, and recall, the CNN model is compiled, and the optimization goals and performance evaluation criteria for model training are clarified.

[0108] (5) Model training execution

[0109] The divided training and validation sets are input into the compiled CNN model, and the training process is started according to the set batch size and training epochs. During training, the changes in the validation set loss are monitored in real time, and training is terminated when the early stopping condition is triggered to ensure the generalization performance of the model.

[0110] (6) Model evaluation and storage

[0111] After training, the model's classification accuracy, precision, recall, and other performance metrics are calculated using the test set. The model with the lowest loss on the validation set is saved as the final device oil contamination self-inspection classification model.

[0112] Step 1.3: Establish the range hood sound field database, which contains Class M sound field environments where the device positions are normal;

[0113] A continuous I-frame of range hood audio signal is collected using a reference microphone as a single data point for the range hood.

[0114] Each type of sound field environment in the range hood sound field database contains J range hood data entries. Each range hood data entry is calibrated to obtain calibration information, which includes the sound field environment type, the oil stain status of electroacoustic devices, the fan speed, and the serial number (e.g., sound field environment type A, electroacoustic device oil stain status: transparent oil stain on the surface of the sound-transparent structure - moderate coverage, fan speed: level 1, serial number: data entry 3).

[0115] Step 1.4: Establish a secondary channel estimation function library and a controller coefficient library, including:

[0116] For each range hood data point, a secondary channel identification algorithm (such as the Least Mean Square (LMS) algorithm) is used to identify the secondary channels, obtaining the secondary channel estimation function. This secondary channel estimation function is then substituted into the semi-adaptive active noise cancellation system, and an adaptive algorithm (such as the Filtered-x Least Mean Square (FxLMS) algorithm) is used for active noise cancellation. Once the semi-adaptive active noise cancellation system reaches its optimal noise cancellation effect, the secondary channel estimation function and controller coefficients corresponding to the current range hood data point are recorded. The process for determining when the semi-adaptive active noise cancellation system reaches its optimal noise cancellation effect includes:

[0117] The residual noise e(n) and its mean square value are collected in real time using an error microphone. The calculation formula is as follows:

[0118] ;

[0119] Among them, when the range hood does not produce a whistling sound, the mean square value of the residual noise is... If the preset number of seconds remains unchanged and is less than the first threshold (which is determined based on a large amount of experimental data and the mean square value of residual noise when the system reaches a stable and optimal noise reduction effect), then the semi-adaptive active noise reduction system achieves the best noise reduction effect. Let n be the nth number in the residual noise e(n) of the range hood.

[0120] For the M-class sound field environment, the secondary channel estimation functions of J range hood data in each class are summed and averaged to obtain the secondary channel estimation functions for each type of sound field environment. The secondary channel estimation functions of the M-class range hood data are then used to construct a secondary channel estimation function library. The secondary channel estimation function library contains M-class secondary channel estimation functions that correspond one-to-one with the M-class sound field environment.

[0121] For the M-type sound field environment, the controller coefficients of the J range hood data in each type are summed and averaged to obtain the controller coefficients of each type of sound field environment, and the controller coefficients of the M-type range hood data are used to form a controller coefficient library; the controller coefficient library contains the M-type controller coefficients that correspond one-to-one with the M-type sound field environment.

[0122] Step 1.5: Establish a sound field feature library for range hoods, including:

[0123] For each range hood data point in the range hood sound field database, fifteen features are calculated for each range hood audio signal frame using sampling points: zero-crossing rate, short-time amplitude, energy, absolute mean, root mean square value, variance, standard deviation, kurtosis, skewness, peak index, peak factor, margin coefficient, spectral centroid, mean square frequency, and frequency variance. Each feature is calculated using the corresponding formulas from all sampling points within the range hood audio signal frame. This results in fifteen features of I-frame length for each feature.

[0124] Calculate the variance of fifteen features of length I for each frame (e.g., calculate the variance of short-time amplitude feature using the variance calculation formula; calculate the variance of zero-crossing rate feature using the variance calculation formula; and so on for other features). Select the ten features with the highest variance from the fifteen features, and then use the Pearson correlation coefficient to calculate the correlation between any two features among these ten features (the correlation value is between -1 and 1, where 1 indicates perfect positive correlation, -1 indicates perfect negative correlation, and 0 indicates no linear relationship):

[0125] When the correlation between two features is less than a preset value (which is 0.8), it means that the correlation between the two features is low, and both features are retained.

[0126] When the correlation between two features is greater than or equal to a preset value, it indicates that the correlation between the two features is high. In this case, the feature with the larger variance is selected from the two features and retained (to save computing resources, the feature with the higher variance is retained).

[0127] Based on the variance sorting and correlation analysis results between features (after verification of multiple range hood data), only five features are ultimately retained, which are called the five main features: short-term amplitude, kurtosis, peak index, margin coefficient, and spectral centroid. For each range hood data in the range hood sound field database, the five main features with a length of I frames are retained.

[0128] Normalization is performed on each frame of the five main features (the normalization in this application is all Min-Max Normalization, scaled to the [0, 1] interval), to obtain the five main features with a normalized length of I frames;

[0129] The range hood sound field feature library is constructed by normalizing five main features of length I frames corresponding to all range hood data in the range hood sound field database. The range hood sound field feature library contains M types of features that correspond one-to-one with M types of sound field environments.

[0130] Step 1.6, establishing the range hood sound field vector representation library, includes: In the pre-training stage, for each range hood data in each type of sound field environment, extracting five normalized main features of length I frames corresponding to the current range hood data in the current type of sound field environment from the range hood sound field feature library. Then, concatenating the five normalized main features and inputting them into the trained range hood noise classification model. In the feature layer of the trained range hood noise classification model (the feature layer is the last fully connected layer before the classification layer), the vector representation of the current range hood data in the current type of sound field environment is output. The mean of the J vector representations of the current type of sound field environment is calculated to obtain the vector representation of the current type of sound field environment. The vector representations of the M types of sound field environments constitute the range hood sound field vector representation library, which contains M-type vector representations that correspond one-to-one with the M types of sound field environments.

[0131] The training of the range hood noise classification model involves using all calibrated sound field features from the range hood sound field feature library to train the model, resulting in a trained model that is a deep neural network (DNN). The specific training process is as follows:

[0132] (1) Dataset preprocessing

[0133] The dataset uses calibrated sound field features from all categories of range hoods. First, five principal features of normalized length I frames are extracted from the range hood sound field feature library for each of the J range hood data points in M ​​sound field environments. These five principal features are then concatenated. Subsequently, a stratified sampling method is used to divide the dataset into training, validation, and test sets in a 7:2:1 ratio, used for model training, hyperparameter tuning, and final performance evaluation, respectively.

[0134] (2) DNN model structure

[0135] The range hood noise classification model is a custom deep neural network (DNN) with a hierarchical design. Its structure from input to output and key parameter settings are as follows:

[0136] Input layer: The dimension is the same as the dimension of the concatenated feature vector (5 by I), and it receives the preprocessed sound field feature data of the range hood;

[0137] Hidden layers: 3 fully connected layers with 256, 128 and 128 neurons respectively. All layers use the ReLU activation function to introduce non-linear mapping. A Dropout layer (dropout rate set to 0.2) is added after each layer to alleviate overfitting.

[0138] Feature layer: The third hidden layer (i.e., the last fully connected layer before the classification layer), with a dimension of 128, outputs a vector representation of the sound field features of the range hood;

[0139] Classification layer: The output dimension is the range hood sound field environment category M, and the activation function is Softmax.

[0140] (3) Model training configuration

[0141] Optimizer and loss function: The Adam optimizer is used (learning rate = 0.001, exponential decay rate of first-order momentum β1 = 0.9, exponential decay rate of second-order momentum β2 = 0.999), and the loss function is classification cross-entropy.

[0142] Training parameters: Batch size is set to 32, and training epochs are 50.

[0143] Regularization strategy: Early stopping is used to monitor the validation set loss. If the loss does not decrease for five consecutive training rounds, training is terminated early, and the model weights with the lowest validation set loss are automatically restored.

[0144] (4) Model compilation

[0145] Based on the aforementioned optimizer and loss function, and combined with evaluation metrics such as classification accuracy, precision, and recall, the DNN model is compiled, and the optimization objectives and performance evaluation criteria for model training are clarified.

[0146] (5) Model training execution

[0147] The divided training and validation sets are input into the compiled DNN model, and the training process is started according to the set batch size and training rounds. During training, the changes in the validation set loss are monitored in real time, and training is terminated when the early stopping condition is triggered to ensure the generalization performance of the model.

[0148] (6) Model evaluation and storage

[0149] After training, the model's classification accuracy, precision, recall, and other performance metrics are calculated using the test set. The model with the lowest loss on the validation set is saved as the final range hood noise classification model.

[0150] Step 2, Actual Noise Reduction Stage:

[0151] Step 2.1: Triggered Self-Check of Electroacoustic Device Status. At the moment the range hood starts, the image acquisition process is triggered. First, the lighting module (such as an LED fill light) of the image acquisition device is turned on, illuminating the area of ​​the electroacoustic devices inside the range hood. Then, the acquisition device captures a current image frame containing all electroacoustic devices (after which, the lighting module is turned off to save energy). Based on the device position self-check vector representation library, the position type of each electroacoustic device in the current image frame is determined. When the electroacoustic device position is normal, the range hood's semi-adaptive active noise reduction system proceeds normally to the next step. When the electroacoustic device position is abnormal (such as tilting or falling off), the operation of the range hood's semi-adaptive active noise reduction system is interrupted (it will not attempt to run again during the range hood's startup), and an alarm message is pushed to the user after the range hood is turned off for quick location and resolution of the problem. This method effectively avoids system howling and other problems caused by abnormal device positions and significantly reduces the system's continuous computational load.

[0152] The process of determining the location type of each electroacoustic device in the current image frame based on the device location self-check vector representation library is as follows: the current image frame is normalized, and the normalized current image frame is used to generate a vector representation of the electroacoustic device location (the specific process of this vector representation is: inputting the normalized current image frame into the trained device location self-check classification model to obtain the vector representation of the device location in the current image frame), comparing the similarity (such as Euclidean distance) between the vector representation of the electroacoustic device location in the current image frame and the vector representation in the device location self-check vector representation library, and taking the device location type corresponding to the vector representation with the highest similarity in the device location self-check vector representation library as the electroacoustic device location type of the current image frame;

[0153] Step 2.2: When the position of the electroacoustic device is normal, determine the oil stain status of the electroacoustic device in the current image frame according to the device oil stain self-check vector expression library. If the oil stain status of the electroacoustic device is abnormal, interrupt the operation of the range hood semi-adaptive active noise reduction system (otherwise the range hood semi-adaptive active noise reduction system will work normally).

[0154] The process of determining the oil contamination status of the electroacoustic device in the current image frame based on the device oil contamination self-inspection vector expression library is as follows: The normalized current image frame is used to generate a vector expression of the device oil contamination status (the specific process of this vector expression is: inputting the normalized current image frame into the trained device oil contamination self-inspection classification model to obtain a vector expression of the degree of device oil contamination in the current image frame); the vector expression of the electroacoustic device oil contamination status in the current image frame is compared with the vector expressions in the device oil contamination self-inspection vector expression library for similarity (e.g., Euclidean distance); the oil contamination status corresponding to the vector expression with the highest similarity in the device oil contamination self-inspection vector expression library is taken as the electroacoustic device oil contamination status of the current image frame (oil contamination type and adhesion status (e.g., "viscous oil contamination - heavily clogged mesh" or "transparent oil contamination - light surface coverage")).

[0155] It should be noted that, based on the determined oil contamination status of the electroacoustic components, the range hood's semi-adaptive active noise reduction system adopts a graded response strategy:

[0156] (1) Normal / Mild state: If the state is determined to be "clean" or "clear oil stains - light coverage", the impact on acoustic performance is negligible, the range hood semi-adaptive active noise reduction system will operate normally;

[0157] (2) Moderate impact status: If the status is determined to be "transparent oil stains - moderate coverage", the range hood semi-adaptive active noise reduction system records the status and continues to operate, prompting the user to perform daily cleaning after the range hood is turned off;

[0158] (3) Severe impact / high risk state: If it is determined to be "sticky oil stains" or any type of "severe coverage / blockage" state, the range hood semi-adaptive active noise reduction system will immediately (or after a short attempt to fine-tune the parameters is ineffective) interrupt the active noise reduction function and issue a maintenance alarm to the user after the range hood is turned off, prompting that the sound-transmitting structure of the electroacoustic device needs to be deeply cleaned or the components replaced; among them, the third situation belongs to the abnormal self-test situation of the electroacoustic device status.

[0159] Step 2.3: When the oil stain status of the electroacoustic device is normal (when both the position and oil stain status of the electroacoustic device are normal, the self-test of the electroacoustic device status is normal), acquire the current range hood audio signal frame collected in real time by the reference microphone, extract the features corresponding to the range hood audio signal frame according to the range hood sound field feature library (i.e., according to the five main features), normalize each extracted feature, and then splice the normalized features to obtain the spliced ​​features;

[0160] Step 2.4: Input the splicing features of the current range hood audio signal frame into the trained range hood noise classification model to obtain the vector representation of the current range hood audio signal frame. Compare the similarity of the vector representation of the current range hood audio signal frame with the range hood sound field vector representation library, and combine the oil stain status of the electroacoustic devices in the current image frame to determine the sound field environment type of the current range hood audio signal frame.

[0161] The process of comparing the vector representation of the current range hood audio signal frame with the range hood sound field vector representation library, and determining the sound field environment type of the current range hood audio signal frame based on the oil stain status of the electroacoustic devices in the current image frame, includes: first, reducing the range hood sound field vector representation library based on the oil stain status of the electroacoustic devices in the current image frame; then, comparing the vector representation of the current range hood audio signal frame with the reduced range hood sound field vector representation library, obtaining the sound field environment corresponding to the vector representation with the highest similarity in the reduced range hood sound field vector representation library, and using this sound field environment as the current range hood audio signal frame. Sound field environment type (First, based on the oil stain status of the electroacoustic device determined by the current image frame, some information of the current sound field environment is determined. Therefore, the range of the range hood sound field vector expression library that needs to be compared with the vector expression of the current range hood audio signal frame can be narrowed. That is, only the part of the range hood sound field vector expression library that meets the oil stain status of the electroacoustic device needs to be compared. Then, the similarity comparison between the vector expression of the current range hood audio signal frame and the narrowed range hood sound field vector expression library is performed to obtain the sound field environment corresponding to the vector expression with the highest similarity in the range hood sound field vector expression library, and this sound field environment is taken as the sound field environment type of the current range hood audio signal frame).

[0162] Step 2.5: Based on the sound field environment type of the current range hood audio signal frame, select the corresponding secondary channel estimation function and controller coefficient from the preset secondary channel estimation function library and controller coefficient library respectively;

[0163] Step 2.6: The controller adopts an adaptive algorithm (FxLMS algorithm). Based on the selected secondary channel estimation function and controller coefficients, as well as the current range hood audio signal frame collected by the reference microphone and the residual noise collected by the error microphone, it calculates an anti-phase sound wave signal that is opposite in phase to the current range hood audio signal frame, and transmits the calculated anti-phase sound wave signal to the secondary speaker for emission, thereby achieving noise reduction of the current range hood audio signal frame. If the sound field environment type of subsequent range hood audio signal frames less than Q consecutive frames differs from that of the current range hood audio signal frame, the current secondary channel estimation function remains unchanged, and the current controller coefficients are used as initial values. An adaptive algorithm (FxLMS algorithm) is used to continuously fine-tune the controller coefficients based on the noise reduction results until the semi-adaptive active noise cancellation system achieves optimal noise reduction or requires re-selection of the corresponding secondary channel estimation function and controller coefficients. If the sound field environment type of subsequent range hood audio signal frames greater than or equal to Q consecutive frames differs from that of the current range hood audio signal frame, the corresponding secondary channel estimation function and controller coefficients need to be re-selected for the first subsequent frame. The re-selected secondary channel estimation function and controller coefficients are then used to refine the noise reduction process. The controller coefficients are used to calculate the inverted acoustic signal (i.e., based on the selected secondary channel estimation function and controller coefficients, as well as the current range hood audio signal frame collected by the reference microphone and the residual noise collected by the error microphone, the inverted acoustic signal with the opposite phase to the current range hood audio signal frame is calculated, and the calculated inverted acoustic signal is transmitted to the secondary speaker for emission to achieve noise reduction). Then, the reselected secondary channel estimation function is kept unchanged, and the reselected control coefficients are used as the initial values. Based on the noise reduction results, the adaptive algorithm (FxLMS algorithm) is used to continuously fine-tune the controller coefficients until the semi-adaptive active noise reduction system achieves the best noise reduction effect or the corresponding secondary channel estimation function and controller coefficients need to be reselected.

[0164] When using an adaptive algorithm to fine-tune the controller parameters, the residual noise collected by the error microphone and corresponding to the current range hood audio signal frame is obtained, and the amplitude of the current residual noise is compared with the second threshold. If the amplitude of the residual noise of the range hood is greater than the second threshold, the fine-tuning of the controller parameters is stopped immediately to prevent howling.

[0165] The second threshold is a safety threshold pre-set based on acoustic safety margin to prevent the system from entering a positive feedback loop (whistling).

[0166] It should be noted that in the actual noise reduction stage, the selection of the secondary channel estimation function and the controller coefficients, or the updating of the controller coefficients, falls into the following categories:

[0167] (1) If the current range hood audio signal frame has the same sound field environment type as the previous range hood audio signal frame, then there is no need to reselect the corresponding secondary channel estimation function and controller coefficients (use the same secondary channel estimation function and controller coefficients), and the controller coefficients are fine-tuned using an adaptive algorithm based on the noise reduction result, and the calculation of the inverted sound wave signal continues (stop fine-tuning the controller coefficients when the semi-adaptive active noise reduction system reaches the best noise reduction effect).

[0168] (2) If, starting from the current range hood audio signal frame, the sound field environment type of the range hood audio signal frame less than Q consecutive frames is different from that of the previous range hood audio signal frame, then there is no need to reselect the corresponding secondary channel estimation function and controller coefficients, which is the same as in case (1).

[0169] (3) If, starting from the current range hood audio signal frame, there are more than or equal to Q consecutive range hood audio signal frames of a different type than the previous range hood audio signal frame in the pre-training library, then the corresponding secondary channel estimation function and controller coefficients need to be reselected, and the inverted sound wave signal is calculated based on the reselected secondary channel estimation function and controller coefficients. The controller coefficients are then fine-tuned using an adaptive algorithm based on the noise reduction results, and the calculation of the inverted sound wave signal continues (the fine-tuning of the controller coefficients is stopped when the semi-adaptive active noise reduction system reaches the best noise reduction effect).

[0170] (4) If the current frame category is not included in the pre-training library, the semi-adaptive active noise reduction system is interrupted and the 'incremental learning model' is started. When in actual use, if it is found that the sound field environment of the range hood cannot be highly matched with any category in the existing pre-training library, the user is prompted that the current noise reduction effect is not ideal and whether to record the abnormal data. If yes, the system will store the data and upload it to the server or cloud storage system. During the self-check period (such as after shutdown), the data will be reprocessed and a new category will be generated to complete the update of the pre-training library. If no, it may be instantaneous interference noise (such as spatula collision), and there is no need to record and generate a new type. If the user does not respond within a certain period of time, the system will automatically mark the data as 'pending review' and store it locally. During the next self-check, based on the frequency and feature consistency of the historical 'pending review' data, it will automatically decide whether to generate a new category or discard it.

[0171] This method constructs a secondary channel estimation function library and a controller coefficient library. During the actual noise reduction process, the corresponding secondary channel estimation functions and controller coefficients are selected from these libraries based on the sound field environment type of the current range hood audio signal frame. This avoids the slow convergence and howling problems caused by iterating from scratch in a fully adaptive system, and overcomes the limitation of fixed-parameter systems in adapting to environmental changes. Anti-howling processing is also implemented to improve the user experience. This method innovatively proposes a layout scheme for the image acquisition device. By placing the camera near the range hood's smoke inlet and adjusting the lens parameters based on field-of-view analysis, a single camera can completely cover all electroacoustic components and their key acoustic structures. This achieves global, synchronous visual monitoring of core components with minimal hardware cost, laying the image foundation for highly reliable determination of component location and oil stain status. This method also features an upgraded component oil stain status self-check function, solving the problem that relying solely on audio signal detection can only provide a simple assessment of the presence or absence of oil stains and is insufficient for accurate diagnosis. Because viscous and clear oil stains have different physical impact mechanisms on electroacoustic devices, directly affecting the failure modes and maintenance strategies of noise reduction systems, this method uses vision to perform refined identification of oil stain types and their adhesion states. This can be extended to assess pipe contamination, enabling more precise and forward-looking performance evaluation and maintenance warnings for the noise reduction system's operating environment, significantly improving the system's long-term reliability and user experience. This method innovatively integrates visual and auditory information for sound field recognition. Through visual self-checking of device oil stains at startup, prior information about the oil stain state is obtained, narrowing down the scope of the range hood's sound field vector representation library. Based on this, the vector representation of the current range hood audio signal frame is compared with the narrowed range hood sound field vector representation library, significantly reducing unnecessary computational comparisons and making the sound field environment type determination faster and more accurate. This method features an 'incremental learning mode,' continuously updating the pre-trained library of the range hood's sound field environment, effectively addressing the pain point in semi-adaptive active noise reduction systems for range hoods where "long-term effectiveness of noise reduction is difficult to guarantee due to device wear and tear." This method constructs a complete closed-loop system of "monitoring-noise reduction-learning". In actual operation, the system not only ensures noise reduction safety and effectiveness in real time through visual and audio monitoring, but also triggers an incremental learning mechanism when encountering unexpected sound field environments, transforming new scene data into updates to the pre-training library. This allows the system to continuously optimize itself according to user habits and environmental changes, fundamentally resolving the contradiction between the rigidity of fixed-parameter systems and the instability of fully adaptive systems. It achieves a leap from "one-time calibration" to "lifelong learning and evolution," significantly improving the product's long-term applicability and user satisfaction.

[0172] It should be understood that, although Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0173] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0174] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A semi-adaptive active noise reduction method for a range hood with self-testing function, characterized in that: The self-testing semi-adaptive active noise reduction method for range hoods is applied to a semi-adaptive active noise reduction system for range hoods. The semi-adaptive active noise reduction system includes an image acquisition device, electroacoustic components, and a controller. The electroacoustic components include a reference microphone, a secondary loudspeaker, and an error microphone. The self-testing semi-adaptive active noise reduction method for range hoods includes: Pre-training phase: A device position self-check vector expression library, a device oil stain self-check vector expression library, a range hood sound field database, a secondary channel estimation function library, a controller coefficient library, a range hood sound field feature library, and a range hood sound field vector expression library are established respectively. The range hood sound field database contains M-type sound field environments where the electroacoustic devices are in normal positions. The secondary channel estimation function library contains M-type secondary channel estimation functions that correspond one-to-one with each of the M-type sound field environments. The controller coefficient library contains M-type controller coefficients that correspond one-to-one with each of the M-type sound field environments. The range hood sound field vector expression library contains M-type vector expressions that correspond one-to-one with each of the M-type sound field environments. Actual noise reduction stage: At the moment the range hood is turned on, the current image frame of all electroacoustic devices is acquired by the image acquisition device. The position type of each electroacoustic device in the current image frame is determined according to the device position self-check vector expression library. When the position of an electroacoustic device is abnormal, the operation of the range hood's semi-adaptive active noise reduction system is interrupted. When the position of the electroacoustic device is normal, the oil stain status of the electroacoustic device in the current image frame is determined according to the device oil stain self-check vector expression library. When the oil stain status of the electroacoustic device is abnormal, the operation of the range hood semi-adaptive active noise reduction system is interrupted. When the oil stains on the electroacoustic device are in normal condition, the current range hood audio signal frame collected in real time by the reference microphone is obtained. The features corresponding to the range hood audio signal frame are extracted according to the range hood sound field feature library, and the extracted features are normalized. Then, the normalized features are spliced ​​together to obtain the spliced ​​features. The splicing features of the current range hood audio signal frame are input into the trained range hood noise classification model to obtain the vector representation of the current range hood audio signal frame. The similarity of the vector representation of the current range hood audio signal frame with the range hood sound field vector representation library is compared, and the sound field environment type of the current range hood audio signal frame is determined by combining the oil stain status of the electroacoustic devices in the current image frame. Based on the sound field environment type of the current range hood audio signal frame, the corresponding secondary channel estimation function and controller coefficient are selected from the preset secondary channel estimation function library and controller coefficient library, respectively. The controller uses an adaptive algorithm to calculate an anti-phase sound wave signal that is out of phase with the current range hood audio signal frame based on the selected secondary channel estimation function and controller coefficients, as well as the current range hood audio signal frame collected by the reference microphone and the residual noise collected by the error microphone. The calculated anti-phase sound wave signal is then transmitted to the secondary speaker for emission, thereby achieving noise reduction. The noise reduction effect of a semi-adaptive active noise reduction system is monitored based on the residual noise collected by the error microphone.

2. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 1, characterized in that: In the process of establishing the device position self-check vector expression library, multiple first image frames containing all electroacoustic devices are acquired using image acquisition equipment. All first image frames are divided into two categories according to whether the electroacoustic device position is abnormal: normal electroacoustic device position and abnormal electroacoustic device position. Each category contains W first image frames, and each first image frame is labeled with the electroacoustic device position category and number. For each first image frame in each class, a vector representation is generated. Then, the average of the vector representations of W first image frames in each class is calculated to obtain the vector representations of the first image frames in the two classes. The vector representations of the first image frames in the two classes constitute the device position self-test vector representation library.

3. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 1, characterized in that: The oil contamination status of the electroacoustic device includes the type of oil contamination and the adhesion status; In the process of establishing the device oil contamination self-inspection vector expression library, multiple second image frames containing all electroacoustic devices were acquired using image acquisition equipment. All second image frames were divided into H categories according to oil contamination type and adhesion state, with each category containing K second image frames. Second image frames of the same category had the same oil contamination type and adhesion state. Each second image frame was labeled with the label and number of the electroacoustic device oil contamination state. For each second image frame in each class, a vector representation is generated. Then, the mean of the vector representations of K second image frames in each class is calculated to obtain the vector representations of the second image frames of class H. The vector representations of the second image frames of class H constitute the device oil stain self-inspection vector representation library.

4. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 3, characterized in that: Each type of sound field environment in the range hood sound field database contains J range hood data. A reference microphone is used to collect I consecutive frames of range hood audio signal as one range hood data. Each range hood data is then calibrated to obtain calibration information, which includes the sound field environment type, the oil stain status of electroacoustic devices, the fan speed, and the serial number.

5. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 4, characterized in that: The establishment of the secondary channel estimation function library and controller coefficient library includes: For the sound field environment corresponding to each range hood data, a secondary channel identification algorithm is used to identify the secondary channel and obtain the secondary channel estimation function. The secondary channel estimation function is substituted into the semi-adaptive active noise reduction system, and an adaptive algorithm is used for active noise reduction. When the semi-adaptive active noise reduction system reaches the best noise reduction effect, the secondary channel estimation function and controller coefficient corresponding to the current range hood data are recorded. For the M-class sound field environment, the secondary channel estimation functions of the J range hood data in each class are summed and averaged to obtain the secondary channel estimation functions of each type of sound field environment, and the secondary channel estimation functions of the M-class range hood data are used to form a secondary channel estimation function library; For the M-type sound field environment, the controller coefficients of the J range hood data in each type are summed and averaged to obtain the controller coefficients of each type of sound field environment, and the controller coefficients of the M-type range hood data are used to form a controller coefficient library.

6. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 5, characterized in that: The establishment of the range hood sound field feature library includes: For each range hood data point in the range hood sound field database, the following 15 features are calculated for each range hood audio signal frame: zero crossing rate, short-time amplitude, energy, absolute mean, root mean square value, variance, standard deviation, kurtosis, skewness, peak index, peak factor, margin coefficient, spectral centroid, mean square frequency, and frequency variance. Thus, 15 features with a length of I frames are obtained for each feature. Calculate the variance of each of the fifteen features of length I-frame. Select the ten features with the highest variance from the fifteen features, and then use the Pearson correlation coefficient to calculate the correlation between any two features among these ten features: When the correlation between two features is less than a preset value, it means that the correlation between the two features is low, and both features are retained. When the correlation between two features is greater than or equal to a preset value, it indicates that the correlation between the two features is high, and the feature with the larger variance is selected from the two features for retention. Ultimately, only five features are retained, referred to as the five main features: short-term amplitude, kurtosis, peak index, margin coefficient, and spectral centroid. For each I-frame of range hood audio signal in the range hood sound field database, the five main features with a length of I frames are retained. Normalize each frame of the five main features to obtain the five main features with a normalized length of I frames; The five main features of the range hood sound field are normalized to a length of I frames and constitute the range hood sound field feature library.

7. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 6, characterized in that: In the pre-training stage, the range hood noise classification model is trained using all range hood sound field features in the range hood sound field feature library to obtain a trained range hood noise classification model, and the range hood noise classification model is a deep neural network. During the pre-training phase, for each range hood data point in each sound field environment, five normalized main features of length I frames corresponding to the current range hood data point in the current sound field environment of the current category are extracted from the range hood sound field feature library. The five normalized main features are then concatenated and input into the trained range hood noise classification model. In the feature layer of the trained range hood noise classification model, the vector representation of the current range hood data point in the current sound field environment of the current category is output. The mean of the J vector representations of the current sound field environment of the current category is calculated to obtain the vector representation of the current sound field environment of the current category. The vector representations of the M sound field environments constitute the range hood sound field vector representation library.

8. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 5, characterized in that: The monitoring of the noise reduction effect of the semi-adaptive active noise cancellation system based on residual noise collected by the error microphone includes: The residual noise e(n) and its mean square value are collected in real time using an error microphone. The calculation formula is as follows: ； Among them, when the range hood does not produce a whistling sound, the mean square value of the residual noise is... If the preset number of seconds remains unchanged and is less than the first threshold, the semi-adaptive active noise cancellation system will achieve the best noise reduction effect. This represents the nth number in the residual noise e(n); When the amplitude of the residual noise exceeds the second threshold, it is determined that the range hood is whistling and the noise reduction is immediately interrupted.

9. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 4, characterized in that: The process of determining the position type of each electroacoustic device in the current image frame based on the device position self-test vector expression library is as follows: normalize the current image frame, and express the position of the electroacoustic device in the normalized current image frame as a vector expression. Compare the similarity between the current image frame's electroacoustic device position vector expression and the vector expression in the device position self-test vector expression library. The device position type corresponding to the vector expression with the highest similarity in the device position self-test vector expression library is taken as the electroacoustic device position type of the current image frame. The process of determining the oil contamination status of the electroacoustic device in the current image frame based on the device oil contamination self-inspection vector expression library is as follows: the normalized current image frame is expressed as a vector of the device oil contamination status; the similarity of the vector of the current image frame's electroacoustic device oil contamination status with the vector expression in the device oil contamination self-inspection vector expression library is compared; and the oil contamination status corresponding to the vector expression with the highest similarity in the device oil contamination self-inspection vector expression library is taken as the current image frame's electroacoustic device oil contamination status. The process of comparing the similarity of the vector representation of the current range hood audio signal frame with the range hood sound field vector representation library, and determining the sound field environment type of the current range hood audio signal frame in conjunction with the oil stain status of the electroacoustic devices in the current image frame, includes: first, reducing the range hood sound field vector representation library based on the oil stain status of the electroacoustic devices in the current image frame; then, comparing the similarity of the vector representation of the current range hood audio signal frame with the reduced range hood sound field vector representation library, obtaining the sound field environment corresponding to the vector representation with the highest similarity in the reduced range hood sound field vector representation library, and taking this sound field environment as the sound field environment type of the current range hood audio signal frame.

10. The semi-adaptive active noise reduction method for a range hood with self-testing function as described in claim 1, characterized in that: The self-testing semi-adaptive active noise reduction method for range hoods further includes: during the actual noise reduction stage, if the sound field environment type of the current range hood audio signal frame is the same as that of the previous range hood audio signal frame, then there is no need to reselect the corresponding secondary channel estimation function and controller coefficients. The controller coefficients are fine-tuned using an adaptive algorithm based on the noise reduction results, and the calculation of the inverted sound wave signal continues. If, starting from the current range hood audio signal frame, the sound field environment type of the range hood audio signal frames greater than or equal to Q consecutive frames is different from that of the previous range hood audio signal frame, then the corresponding secondary channel estimation function and controller coefficients need to be reselected.

Images