A vehicle-mounted selective noise reduction method based on sound field reconstruction
By collaboratively collecting acoustic and vibration data through sensors, using blind source separation and convolutional neural networks for voiceprint recognition, and generating adaptive noise reduction parameters combined with physical vibration warnings, this technology solves the problem of insufficient selective protection of safety sounds in existing vehicle noise reduction technologies, and improves the perception and response capabilities of emergency sound sources.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- RIVOTEK TECH (JIANGSU) CO LTD
- Filing Date
- 2026-02-27
- Publication Date
- 2026-05-26
AI Technical Summary
Existing vehicle noise reduction technologies lack selective protection for safety sounds when reducing overall noise energy, making it difficult to achieve zoned adaptive control based on vehicle speed, sound source location, and safety level. Furthermore, the robustness of a single sensing mode is insufficient, leading to the masking of emergency sound sources and a decrease in recognition confidence.
Acoustic and vibration data are collected collaboratively by sensors. Blind source separation algorithm and convolutional neural network are used for voiceprint feature recognition to generate adaptive noise reduction parameters. Combined with physical vibration warning, selective noise reduction is achieved.
It improves the detection and recognition accuracy of critical safety sounds, avoids the obscuring of emergency warnings, and reduces the risk of missed hearing or misjudgment by using bone conduction vibration prompts, while taking into account both comfort and safety.
Smart Images
Figure CN122090858A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle safety warning technology, and in particular to a vehicle selective noise reduction method based on sound field reconstruction. Background Technology
[0002] In recent years, active noise cancellation, beamforming, and sound field reconstruction technologies for vehicles have made significant progress in the field of automotive cabin acoustic optimization. Active noise cancellation schemes based on adaptive filters and multi-microphone arrays have effectively suppressed low-frequency periodic noise, such as engine vibration and vehicle body resonance. Simultaneously, methods based on time-frequency analysis and blind source separation (such as ICA / FastICA) have been introduced to handle mixed non-stationary acoustic sources, improving noise separation capabilities under complex road conditions. In recent years, multimodal perception and deep learning technologies (including convolutional neural networks for voiceprint / acoustic event recognition) have rapidly developed in the fields of vehicle safety and human-machine interaction, making the detection and classification of critical safety sounds such as ambulance sirens and horns more robust. In parallel, physical cues such as bone conduction and seat vibration feedback are being explored to convey emergency information to drivers without interfering with the auditory environment, thus providing a new path for coupling noise reduction and safety cues.
[0003] Despite the rapid development of each of these technologies, existing solutions still have several shortcomings in comprehensive applications. Traditional ANC and sound field reconstruction focus on overall noise energy reduction, lacking selective protection for safety sounds, which can easily lead to the masking of emergency sound sources. Blind source separation has limited effectiveness in separating coherent noise (such as wind noise and tire noise) in pure acoustic channels, and its robustness under single channels or single sensing modes is insufficient. Existing voiceprint recognition mostly relies on offline training and single features, and its recognition confidence decreases when faced with multi-path, reverberation, and low signal-to-noise ratio situations in the vehicle. In addition, the cabin usually adopts a uniform noise reduction strategy, which fails to achieve zoned adaptive control based on vehicle speed, sound source spatial location, and safety level, and also lacks a strategy closely coupled with physical vibration warnings, making it difficult to ensure both comfort and driving safety. Summary of the Invention
[0004] In view of the problems existing in a current vehicle-mounted selective noise reduction method based on sound field reconstruction, this invention is proposed. Therefore, the problem to be solved by this invention is how to provide a vehicle-mounted selective noise reduction method based on sound field reconstruction.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an in-vehicle selective noise reduction method based on sound field reconstruction, which includes: collecting in-vehicle acoustic data and vibration data through sensors in collaboration; separating the in-vehicle acoustic data and vibration data through a blind source separation algorithm; and obtaining a voiceprint feature tensor. The voiceprint feature tensor is input into a convolutional neural network for voiceprint recognition, and the safety sound category and the safety sound hazard level are output. Adaptive noise reduction parameters are generated based on vehicle speed and safety sound category. The final adaptive noise reduction parameters are assigned to seats in different areas, and physical vibration warnings are issued in conjunction with the safety sound hazard level to complete the vehicle-mounted selective noise reduction.
[0006] As a preferred embodiment of the vehicle-mounted selective noise reduction method based on sound field reconstruction described in this invention, the sensor includes a microphone array and a wheel arch vibration sensor; the microphone array is used to collect in-vehicle acoustic data, and the wheel arch vibration sensor is used to collect vibration data.
[0007] As a preferred embodiment of the vehicle-mounted selective noise reduction method based on sound field reconstruction described in this invention, the step of separating in-vehicle acoustic data and vibration data using a blind source separation algorithm includes: The acoustic data from the microphone array and the vibration data from the wheel arch vibration sensor are normalized. All the normalized acoustic data and vibration data are combined into an observation vector. The covariance matrix is calculated, and the covariance matrix is decomposed into eigenvalues to obtain the eigenvector matrix. The whitening matrix is constructed based on the eigenvector matrix, and the whitened data is obtained, represented as follows: ; ; in: For whitening matrix, It is an eigenvalue diagonal matrix. The eigenvector matrix, It is the transpose symbol. for The whitened observation vector at time step [time]. for The observation vector at time; The separation vector is iteratively updated based on the whitened observation vector, as follows: ; in: For the updated number Separated vectors, For the original number Separated vectors, The sample mean. Let be the log-likelihood gradient function. The first derivative of the log-likelihood gradient function. The whitened observation vector; By assembling all the separating vectors into a separating vector matrix, the observed vectors can be recovered, represented as: ; in: for The recovered observation vector at time step, For the separation vector matrix; A short-time Fourier transform is performed on the recovered observation vector to generate the amplitude spectrum, and the Mel frequency cepstral coefficients are obtained as the voiceprint feature tensor.
[0008] As a preferred embodiment of the vehicle-mounted selective noise reduction method based on sound field reconstruction described in this invention, wherein the input convolutional neural network for voiceprint recognition includes: Voiceprint feature tensors are input into a convolutional neural network for voiceprint recognition. The convolutional neural network includes a feature extraction module and a classification and matching module. The feature extraction module includes, in sequence, a convolutional component 1, a residual structure 1, a residual structure 2, a residual structure 3, a residual structure 4, a two-layer residual structure 5, a residual structure 6, and a global average pooling layer; the convolutional component 1 is connected to the residual structure 1, the residual structure 1 is sequentially connected to the residual structure 4; the residual structure 4 is connected to the two-layer residual structure 5, the two-layer residual structure 5 is connected to the residual structure 6, and the residual structure 6 is connected to the global average pooling layer; Specifically: Convolutional Component 1 uses 64 5×5 convolutional kernels with a stride of 2, the input is the voiceprint feature tensor, and the activation function is ReLU; Residual Structure 1 and Residual Structure 2 use 64 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU; Residual Structure 3 and Residual Structure 4 use 128 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU; Two-Layer Residual Structure 5 uses 256 3×3 convolutional kernels in both convolutional layers with a stride of 1 and the activation function is ReLU; Residual Structure 6 uses 256 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU. The classification and matching module includes a fully connected layer 1, a fully connected layer 2, and an output layer; the fully connected layer 1 has 256 nodes and uses the ReLU activation function; the fully connected layer 2 has 128 nodes and uses the ReLU activation function; the number of nodes in the output layer is equal to the number of entries in the voiceprint feature library, and the output is processed by the Softmax activation function to generate the matching probability corresponding to each entry; the entry with the highest matching probability is matched to the predefined safety sound category and safety sound hazard level.
[0009] As a preferred embodiment of the vehicle-mounted selective noise reduction method based on sound field reconstruction described in this invention, the step of generating adaptive noise reduction parameters based on vehicle speed and voiceprint recognition results includes: Obtain real-time vehicle speed, basic noise reduction weight matrix, and safety sound recognition results; predetermine a vehicle speed threshold; when the real-time vehicle speed exceeds the threshold, activate high-speed mode and correct the basic weights of the noise sources, as shown below: ; in: To perform the first in high-speed mode Class of noise in the first Noise reduction weight for seat number 1 This is the gain coefficient. To perform the first in standard mode Class of noise in the first The noise reduction weight for seat number 1, i.e., the base noise reduction weight; If the real-time vehicle speed is less than the vehicle speed threshold, all basic noise reduction weights remain unchanged. For each identified safety tone category, calculate the distance from the seat to the sound source of the safety tone, expressed as:
[0010] in: For the first Seat number 1 to safety sound source distance, For the first The coordinates of seat number 1 in the cockpit coordinate system For identified safe sound sources The coordinates; Converting distance into a spatial weighting factor, expressed as: ; ; in: Spatial weighting factor, It is a radius constant. For safe sound sources For the The warning weight of seat number 1; The noise reduction weights corrected in high-speed mode are combined with the warning weights of all safety tones to obtain the final adaptive noise reduction parameters, expressed as: ; ; in: The noise reduction weights are corrected for high-speed mode. For the final adaptive noise reduction parameters, For indicator functions; The cabin is divided into multiple zones, and final adaptive noise reduction parameters are assigned to the seats in each zone. These parameters are then broadcast to the sound field control unit via an in-vehicle protocol, and noise reduction measures are implemented based on the final adaptive noise reduction parameters for each zone.
[0011] As a preferred embodiment of the vehicle-mounted selective noise reduction method based on sound field reconstruction described in this invention, the step of issuing a physical vibration warning based on the safety sound hazard level includes: Obtain the recovered observation vector matrix and the safety sound hazard level. Map the headrest vibration intensity according to the safety sound hazard level, and synthesize the vibration waveform based on the vibration intensity, as shown below: ; ; in: Safety sound hazard level vibration intensity, For the lowest vibration intensity, For unit increment, for The vibration waveform at time t, The vibration frequency, The envelope attenuation constant; The vibration waveform is transmitted to the vibration motor driver inside the headrest to issue a physical vibration warning.
[0012] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of an in-vehicle selective noise reduction method based on sound field reconstruction.
[0013] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements the steps of an in-vehicle selective noise reduction method based on sound field reconstruction.
[0014] The beneficial effects of this invention are as follows: This method significantly improves the detection and identification accuracy of critical safety sounds in road noise environments; achieves a dynamic balance of noise reduction strategies, avoiding obscuring emergency warnings; and provides additional cues beyond visual / auditory through bone conduction vibration, reducing the risk of drivers missing or misjudging sounds. It balances noise suppression and active safety, effectively enhancing the vehicle's perception and response capabilities to emergency sound sources while improving ride comfort. Attached Figure Description
[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart of an in-vehicle selective noise reduction method based on sound field reconstruction. Detailed Implementation
[0017] To make the above-mentioned objects, features, and advantages of the present invention more readily understood, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0019] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. An embodiment appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment that selectively excludes other embodiments.
[0020] Reference Figure 1 This is the first embodiment of the present invention, which provides an in-vehicle selective noise reduction method based on sound field reconstruction, including: S1: Acoustic and vibration data inside the vehicle are collected collaboratively by sensors. The acoustic and vibration data inside the vehicle are separated by a blind source separation algorithm, and the voiceprint feature tensor is obtained. Specifically, the microphone array uses a condenser array microphone with 32 units covering the 20Hz–20kHz frequency band and a built-in wind noise suppression filter. The vibration sensor is a broadband accelerometer mounted inside the wheel fender, capable of accurately collecting tire noise and vibration data.
[0021] The 32 microphones are evenly distributed under the roof lining, and the spacing and arrangement should take into account both the distribution of safe sound sources and the acoustic reflection characteristics inside the vehicle; they are fixed with wind shields and vibration-damping brackets to avoid direct coupling between wind noise and vehicle body vibration.
[0022] The accelerometer is securely attached to the inner wall of the wheel arch, near the tire tread, and led out to the data acquisition compartment inside the vehicle using shielded twisted-pair cables. All external sensors are connected to the central DAQ chassis via shielded cables.
[0023] The microphone array sampling rate was set to 48kHz, with a segmented buffer of 1024 points; the accelerometer sampling rate was set to 10kHz, with a buffer of 512 points. During the acquisition process, a second-order high-pass filter (cutoff frequency 20Hz) and a fourth-order low-pass inverse filter (cutoff frequency 20kHz) were applied in real time to the acoustic channel; the vibration channel used a 0.5Hz high-pass filter and a 5kHz low-pass filter.
[0024] Each data frame (acoustic + vibration) is appended with a unique timestamp and frame number and sent to the main control unit via the vehicle bus. Upon receiving the data, the main control unit pushes it to the real-time processing module.
[0025] The acoustic data from the 32-channel microphone array and the vibration data from the wheel arch vibration sensor were processed with zero mean, and normalized by the standard deviation of each channel. All the normalized acoustic data and vibration data were combined into an observation vector, the covariance matrix was calculated, and the covariance matrix was decomposed into eigenvalues to obtain the eigenvector matrix. The whitening matrix is constructed based on the eigenvector matrix, and the whitened data is obtained, represented as follows: ; ; in: For whitening matrix, It is an eigenvalue diagonal matrix. The eigenvector matrix, It is the transpose symbol. for The whitened observation vector at time step [time]. for The observation vector at time; The separation vector is iteratively updated based on the whitened observation vector, as follows: ; in: For the updated number Separated vectors, For the original number Separated vectors, The sample mean. Let be the log-likelihood gradient function. The first derivative of the log-likelihood gradient function. The whitened observation vector; By assembling all the separating vectors into a separating vector matrix, the observed vectors can be recovered, represented as: ; in: for The recovered observation vector at time step, For the separation vector matrix; A short-time Fourier transform is performed on the recovered observation vector to generate an amplitude spectrum. The Mel frequency cepstral coefficients are then calculated as the voiceprint feature tensor.
[0026] S2: Input the voiceprint feature tensor into a convolutional neural network to perform voiceprint recognition, and output the safety sound category and the safety sound hazard level; Specifically, the voiceprint feature tensor is input into a convolutional neural network for voiceprint recognition. The convolutional neural network includes a feature extraction module and a classification and matching module. The feature extraction module includes a convolutional component 1, a residual structure 1, a residual structure 2, a residual structure 3, a residual structure 4, a two-layer residual structure 5, a residual structure 6, and a global average pooling layer arranged sequentially. The convolutional component 1 is connected to the residual structure 1, and the residual structure 1 is sequentially connected to the residual structure 4. The residual structure 4 is connected to the two-layer residual structure 5, the two-layer residual structure 5 is connected to the residual structure 6, and the residual structure 6 is connected to the global average pooling layer. Specifically: Convolutional Component 1 uses 64 5×5 convolutional kernels with a stride of 2, the input is the voiceprint feature tensor, and the activation function is ReLU; Residual Structures 1 and 2 use 64 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU; Residual Structures 3 and 4 use 128 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU; Two-Layer Residual Structure 5 uses 256 3×3 convolutional kernels in each of its two convolutional layers with a stride of 1 and the activation function is ReLU; Residual Structure 6 uses 256 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU; All residual structures employ batch normalization to stabilize training and accelerate convergence. The classification and matching module includes a fully connected layer 1, a fully connected layer 2, and an output layer. The fully connected layer 1 has 256 nodes and uses the ReLU activation function. The fully connected layer 2 has 128 nodes and uses the ReLU activation function. The number of nodes in the output layer is equal to the number of entries in the voiceprint feature library. After the output is processed by the Softmax activation function, the matching probability corresponding to each entry is generated. The entry with the highest matching probability is directly matched to the predefined safety sound category and danger level.
[0027] The voiceprint feature database stores standardized voiceprint vectors, such as the feature distribution of ambulance sirens and car horns. The input voiceprint feature tensor is processed by a convolutional neural network to output a probability vector. ,in, Indicates the first in the voiceprint feature database Confidence of matching safe tone.
[0028] If the matching confidence score of a certain type of safety sound is greater than the confidence threshold, then that type of safety sound is determined to have occurred. The safety sound recognition result is output, and the pre-marked safety sound hazard level in the voiceprint feature database is queried based on the safety sound recognition result. S3: Generates adaptive noise reduction parameters based on vehicle speed and safety sound category, assigns final adaptive noise reduction parameters to seats in different areas, and issues physical vibration warnings based on the safety sound hazard level to complete vehicle-mounted selective noise reduction.
[0029] Specifically, real-time vehicle speed, basic noise reduction weight matrix, and safety tone recognition results are obtained from the vehicle's CAN bus; a predetermined vehicle speed threshold is set, and when the real-time vehicle speed exceeds the threshold, high-speed mode is activated, and the basic weights of the noise sources are corrected, as shown below: ; in: To perform the first in high-speed mode Class of noise in the first Noise reduction weight for seat number 1 This is the gain coefficient; the basic noise reduction weights remain unchanged for other noise sources. To perform the first in standard mode Class of noise in the first The noise reduction weight for seat number [number] is the basic noise reduction weight. If the real-time vehicle speed is less than the vehicle speed threshold, all basic noise reduction weights remain unchanged.
[0030] Based on spatial weighting of safety sounds, for each identified safety sound category, the distance from the seat to the sound source of the safety sound is calculated and represented as: ; in: For the first Seat number 1 to safety sound source distance, For the first The coordinates of seat number 1 in the cockpit coordinate system For identified safe sound sources The coordinates; Applying a Gaussian or exponential decay model, the distance is converted into a spatial weighting factor, expressed as: ; ; in: Spatial weighting factor, To control the radius constant of the decay rate, For safe sound sources For the The warning weight of seat number 1; The noise reduction weights corrected in high-speed mode are combined with the warning weights of all safety tones to obtain the final adaptive noise reduction parameters, expressed as: ; ; in: The noise reduction weights are corrected for high-speed mode. For the final adaptive noise reduction parameters, The indicator function is 1 when the noise source and the safety tone are of the same type, and 0 otherwise.
[0031] The cabin is divided into multiple areas (such as the driver's seat, front passenger seat, and rear left and right seats). The final adaptive noise reduction parameters are assigned to the seats in each area. The final adaptive noise reduction parameters are broadcast to the sound field control unit through the vehicle protocol. Each unit performs corresponding noise reduction measures according to the final adaptive noise reduction parameters in its area.
[0032] The system obtains the recovered observation vector matrix and the safety tone hazard level, performs channel assignment mapping, predefines the mapping relationship between speaker array channels and cabin areas, and sends the waveform packets of the corresponding channels to the DSP receivers on the doors and roof according to the recovered observation vector matrix. After receiving the data, the DSP uses its internal DAC to reconstruct the analog data and pushes it to the speaker units.
[0033] The headrest vibration intensity is mapped according to the safety sound hazard level, and the vibration waveform is synthesized based on the vibration intensity, as shown below: ; ; in: Safety sound hazard level vibration intensity, For the lowest vibration intensity, For unit increment, for The vibration waveform at time t, The vibration frequency, The envelope attenuation constant; The vibration waveform is transmitted to the vibration motor driver inside the headrest via the SPI interface to issue a physical vibration warning. The driver has a built-in current loop control that dynamically adjusts the input voltage according to the required vibration intensity.
[0034] When multiple safety tones are triggered simultaneously, vibration warnings are executed in order of hazard level; the vibration will automatically stop if the duration does not exceed the predetermined time, while the speaker warning will continue until the sound data ends.
[0035] This embodiment also provides a computer device applicable to a vehicle-mounted selective noise reduction method based on sound field reconstruction, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement all or part of the steps of the method described in the above embodiments of the present invention.
[0036] This embodiment also provides a storage medium storing a computer program thereon. When the computer program is executed by a processor, it performs the method in any optional implementation of the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0037] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.
[0038] In summary, this method significantly improves the detection and recognition accuracy of critical safety sounds in road noise environments; achieves a dynamic balance in noise reduction strategies, avoiding obscuring emergency warnings; and provides additional cues beyond visual / auditory through bone conduction vibration, reducing the risk of drivers missing or misjudging sounds. It balances noise suppression and active safety, effectively enhancing the vehicle's perception and response capabilities to emergency sound sources while improving ride comfort.
[0039] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A vehicle-mounted selective noise reduction method based on sound field reconstruction, characterized in that: include, Acoustic and vibration data inside the vehicle are collected collaboratively by sensors. The acoustic and vibration data inside the vehicle are separated by a blind source separation algorithm, and the voiceprint feature tensor is obtained. The voiceprint feature tensor is input into a convolutional neural network for voiceprint recognition, and the safety sound category and the safety sound hazard level are output. Adaptive noise reduction parameters are generated based on vehicle speed and safety sound category. The final adaptive noise reduction parameters are assigned to seats in different areas, and physical vibration warnings are issued in conjunction with the safety sound hazard level to complete the vehicle-mounted selective noise reduction.
2. The vehicle-mounted selective noise reduction method based on sound field reconstruction as described in claim 1, characterized in that: The sensor includes a microphone array and a wheel arch vibration sensor; the microphone array is used to collect acoustic data inside the vehicle, and the wheel arch vibration sensor is used to collect vibration data.
3. The vehicle-mounted selective noise reduction method based on sound field reconstruction as described in claim 2, characterized in that: The separation of in-vehicle acoustic and vibration data using a blind source separation algorithm includes: The acoustic data from the microphone array and the vibration data from the wheel arch vibration sensor are normalized. All the normalized acoustic data and vibration data are combined into an observation vector. The covariance matrix is calculated, and the covariance matrix is decomposed into eigenvalues to obtain the eigenvector matrix. The whitening matrix is constructed based on the eigenvector matrix, and the whitened data is obtained, represented as follows: ； ； in: For whitening matrix, It is an eigenvalue diagonal matrix. The eigenvector matrix, It is the transpose symbol. for The whitened observation vector at time step [time] for The observation vector at time; The separation vector is iteratively updated based on the whitened observation vector, as follows: ； in: For the updated number Separated vectors, For the original number Separated vectors, The sample mean. Let be the log-likelihood gradient function. The first derivative of the log-likelihood gradient function. The whitened observation vector; By assembling all the separating vectors into a separating vector matrix, the observed vectors can be recovered, represented as: ； in: for The recovered observation vector at time step, For the separation vector matrix; A short-time Fourier transform is performed on the recovered observation vector to generate the amplitude spectrum, and the Mel frequency cepstral coefficients are obtained as the voiceprint feature tensor.
4. The vehicle-mounted selective noise reduction method based on sound field reconstruction as described in claim 3, characterized in that: The input convolutional neural network for voiceprint recognition includes: Voiceprint feature tensors are input into a convolutional neural network for voiceprint recognition. The convolutional neural network includes a feature extraction module and a classification and matching module. The feature extraction module includes, in sequence, a convolutional component 1, a residual structure 1, a residual structure 2, a residual structure 3, a residual structure 4, a two-layer residual structure 5, a residual structure 6, and a global average pooling layer; the convolutional component 1 is connected to the residual structure 1, the residual structure 1 is sequentially connected to the residual structure 4; the residual structure 4 is connected to the two-layer residual structure 5, the two-layer residual structure 5 is connected to the residual structure 6, and the residual structure 6 is connected to the global average pooling layer; Specifically: Convolutional Component 1 uses 64 5×5 convolutional kernels with a stride of 2, the input is the voiceprint feature tensor, and the activation function is ReLU; Residual Structure 1 and Residual Structure 2 use 64 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU; Residual Structure 3 and Residual Structure 4 use 128 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU; Two-Layer Residual Structure 5 uses 256 3×3 convolutional kernels in both convolutional layers with a stride of 1 and the activation function is ReLU; Residual Structure 6 uses 256 3×3 convolutional kernels with a stride of 1 and the activation function is ReLU. The classification and matching module includes a fully connected layer 1, a fully connected layer 2, and an output layer; the fully connected layer 1 has 256 nodes and uses the ReLU activation function; the fully connected layer 2 has 128 nodes and uses the ReLU activation function; the number of nodes in the output layer is equal to the number of entries in the voiceprint feature library, and the output is processed by the Softmax activation function to generate the matching probability corresponding to each entry; the entry with the highest matching probability is matched to the predefined safety sound category and safety sound hazard level.
5. The vehicle-mounted selective noise reduction method based on sound field reconstruction as described in claim 4, characterized in that: The process of generating adaptive noise reduction parameters based on vehicle speed and voiceprint recognition results includes: Obtain real-time vehicle speed, basic noise reduction weight matrix, and safety sound recognition results; predetermine a vehicle speed threshold; when the real-time vehicle speed exceeds the threshold, activate high-speed mode and correct the basic weights of the noise sources, as shown below: ； in: To perform the first in high-speed mode Class of noise in the first Noise reduction weight for seat number 1 This is the gain coefficient. To perform the first in standard mode Class of noise in the first The noise reduction weight for seat number 1, i.e., the base noise reduction weight; If the real-time vehicle speed is less than the vehicle speed threshold, all basic noise reduction weights remain unchanged. For each identified safety tone category, calculate the distance from the seat to the sound source of the safety tone, expressed as: ； in: For the first Seat number 1 to safety sound source distance, For the first The coordinates of seat number 1 in the cockpit coordinate system For identified safe sound sources The coordinates; Converting distance into a spatial weighting factor, expressed as: ； ； in: Spatial weighting factor, It is a radius constant. For safe sound sources For the first The warning weight of seat number 1; The noise reduction weights corrected in high-speed mode are combined with the warning weights of all safety tones to obtain the final adaptive noise reduction parameters, expressed as: ； ； in: The noise reduction weights are corrected for high-speed mode. For the final adaptive noise reduction parameters, For indicator functions; The cabin is divided into multiple zones, and final adaptive noise reduction parameters are assigned to the seats in each zone. These parameters are then broadcast to the sound field control unit via an in-vehicle protocol, and noise reduction measures are implemented based on the final adaptive noise reduction parameters for each zone.
6. The vehicle-mounted selective noise reduction method based on sound field reconstruction as described in claim 5, characterized in that: The provision of physical vibration warnings based on the hazard level of the safety sound includes: Obtain the recovered observation vector matrix and the safety sound hazard level. Map the headrest vibration intensity according to the safety sound hazard level, and synthesize the vibration waveform based on the vibration intensity, as shown below: ； ； in: Safety sound hazard level vibration intensity, For the lowest vibration intensity, For unit increment, for The vibration waveform at time t, The vibration frequency, The envelope attenuation constant; The vibration waveform is transmitted to the vibration motor driver inside the headrest to issue a physical vibration warning.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the vehicle-mounted selective noise reduction method based on sound field reconstruction as described in any one of claims 1 to 6.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the vehicle-mounted selective noise reduction method based on sound field reconstruction as described in any one of claims 1 to 6.