Loudness compensation method for vehicle-mounted sound system

Abstract

The invention discloses a loudness compensation method for a vehicle-mounted audio system, relates to the technical field of vehicle-mounted audio processing, and solves the technical problem of low loudness compensation precision of the audio system caused by interference of multiple noise sources in the prior art. The method comprises the following steps: collecting vehicle environment data through a multi-mode sensor array; performing mixed noise separation processing on the vehicle environment data, and outputting separated independent noise source features; on the basis of independent noise source features and the vehicle environment data, scene tags are generated through a scene cognition model; inputting the scene label into a reinforcement learning strategy network to generate a loudness compensation strategy; and implementing multi-band independent compensation and sound field control at the output end of the loudspeaker based on a loudness compensation strategy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of vehicle-mounted audio processing, and particularly relates to a loudness compensation method for a vehicle-mounted audio system. BACKGROUND

[0002] With the continuous improvement of the intelligence and comfort requirements of automobiles, the vehicle-mounted audio system has become one of the core configurations for improving the driving and riding experience, and the core goal of loudness compensation as a key function of the vehicle-mounted audio system is to offset the interference of environmental noise on the auditory experience by dynamically adjusting the audio signal parameters during vehicle driving, so as to ensure clear and balanced audio content.

[0003] The existing loudness compensation method is generally simplified in terms of the processing of vehicle noise, and does not fully consider the mixed scene of multiple noise sources actually existing during vehicle driving, so it is unable to distinguish the independent characteristics of various types of noise from the mixed noise, resulting in that the compensation parameter adjustment can only be controlled in a rough way based on the overall noise intensity, and the problems of over-compensation or insufficient compensation are prone to occur; and since the independent noise source characteristics are not separated, the existing scheme is prone to mistakenly suppressing or excessively amplifying part of the frequency bands in the audio signal when adjusting the gain to offset the noise, resulting in distortion of the audio content. Meanwhile, the existing compensation scheme mostly adopts a preset fixed compensation curve, and does not dynamically adjust in combination with the real-time collected noise source characteristics and vehicle environment data, so that the compensation strategy cannot match the dynamic change trend of the noise in real time. SUMMARY

[0004] The present application aims to at least solve one of the technical problems existing in the prior art; for this purpose, the present application proposes a loudness compensation method for a vehicle-mounted audio system, which is used to solve the technical problem of low loudness compensation accuracy of the audio system caused by multiple noise sources in the prior art.

[0005] To achieve the above-mentioned purpose, the first aspect of the present application provides a loudness compensation method for a vehicle-mounted audio system, comprising: acquiring vehicle environment data through a multi-modal sensor array; performing mixed noise separation processing on the vehicle environment data to output separated independent noise source characteristics; generating a scene label through a scene cognition model based on the independent noise source characteristics and the vehicle environment data; the scene label comprises an environmental scene, a vehicle state, a riding state, audio content, a time attribute, and user preferences; inputting the scene label into a reinforcement learning strategy network to generate a loudness compensation strategy; implementing multi-frequency band independent compensation and sound field control at the output end of the loudspeaker based on the loudness compensation strategy.

[0006] Based on the above technical scheme, the loudness compensation method of the vehicle-mounted sound equipment provided by the application comprehensively collects vehicle environment data through a multi-modal sensor array, covers multi-dimensional information such as environment, vehicle, and personnel, provides complete and actual data sources for subsequent compensation decision-making, accurately extracts independent noise source features through mixed noise separation processing, can effectively distinguish different interference sources such as engine noise, wind noise, and external environmental sound, and reduces the interference of mixed noise on the compensation effect, generates multi-dimensional scene labels using a scene cognition model, can accurately match the actual use scene, and generates personalized solutions that fit specific scenes and user needs, then generates a loudness compensation strategy through a reinforcement learning strategy network, can rely on a learning mechanism to continuously optimize the compensation logic, continuously adapt to new situations as the use scene increases, can improve the flexibility and adaptability of the strategy, and finally through multi-frequency band independent compensation and sound field control, can accurately adjust the loudness according to the sound characteristics of different frequency bands, avoid excessive compensation of a certain frequency band to affect the sound quality, optimize the distribution of the sound field in the vehicle, ensure that passengers at different positions in the vehicle can obtain balanced and comfortable auditory experiences, and comprehensively improve the accuracy, personalization, and user experience of the loudness compensation of the vehicle-mounted sound equipment.

[0007] Further, the multi-modal sensor array includes an in-vehicle microphone array, an out-vehicle microphone array, a millimeter wave radar, an air pressure sensor, and a vehicle CAN bus sensor; wherein the millimeter wave radar is deployed at the four corners of the vehicle to monitor the position of the noise source, the air pressure sensor is installed on the window guide rail to detect the opening and closing state, and the CAN bus sensor collects vehicle speed, engine speed, and air conditioner compressor state data.

[0008] Further, the vehicle environment data includes original audio signals, noise source three-dimensional polar coordinates, vehicle speed, engine speed, air conditioner compressor state, and air density.

[0009] Further, the mixed noise separation processing of the vehicle environment data includes: obtaining the three-dimensional polar coordinates of the noise source through the millimeter wave radar point cloud; calculating the wind noise feature template of the current environment based on the wind noise model established based on the fluid mechanics formula; calculating the tire noise feature template of the current environment based on the tire noise model established based on the frequency spectrum feature formula; calculating the air conditioner noise template of the current environment based on the air conditioner noise model coupled by the compressor speed-pressure; obtaining the human voice noise template of the current environment based on the human voice noise model based on the sound pressure-direction joint detection; constructing a separation matrix W = [w ij ] m×n between the microphone array and the noise source; wherein m is the number of microphones, n is the number of noise source types, and w ij represents the weight coefficient of the i th microphone to the j th noise source. The feature templates are input into a separation matrix, and the independent noise source features after separation are output by iteratively solving to maximize the signal-to-noise ratio of each frequency band; the independent noise source features include a wind dryness feature, a tire noise feature, an air conditioner noise feature, and a human voice noise feature.

[0010] Further, the wind dryness model P wind The mathematical expression of (f, t) is as follows: ; wherein f represents a sound frequency, t represents time, is a time-domain variable of a sound pressure signal, p represents air density, v represents vehicle speed, C d represents a turbulent flow coefficient, A represents a gas flow disturbance projection area, refers to an effective projection area of a local structure of a vehicle body in a gas flow direction that generates vortex shedding, f vortex represents a vortex shedding frequency, f c represents a center frequency corresponding to a maximum sound pressure level, and s represents a frequency distribution dispersion degree. The tire noise model S tire The mathematical expression of (f) is as follows: ; wherein K represents a tire coefficient, a represents a vehicle speed influence coefficient, b represents a frequency attenuation index, g represents a road surface attenuation coefficient, a k represents a harmonic amplitude of the kth harmonic, k represents a harmonic index, f rotation represents a tire rotation frequency, and d represents a Dirac delta function used to describe a discrete frequency component. The air conditioner noise model S ac The mathematical expression of (f) is as follows: ; wherein h represents a mechanical vibration conversion efficiency, P represents a refrigerant pressure, f fund represents a compressor operating frequency, b m represents a harmonic attenuation rate of the mth harmonic. The human voice noise model V speech The mathematical expression of (t, f) is as follows: ; wherein L0 represents a human voice triggering threshold, L p represents a sound intensity value, m represents an activation slope, q represents a sound source horizontal azimuth angle, N represents a number of microphones in the vehicle, and y VAD represents a human voice existence probability.

[0011] Further, the human voice existence probability is obtained through a voice activity detection algorithm, and the voice activity detection algorithm comprises the following steps. Frame processing is performed on the original audio signal, and energy and spectral features of each frame are extracted. According to the energy and spectral features of each frame, a human voice existence probability is output through a pre-trained SNM classifier.

[0012] Furthermore, the weight coefficients are optimized using the Gradient Boosting Decision Tree (GBDT) algorithm, including: The initial weight matrix and feature templates of each noise source are input into the GBDT for iterative calculation. In each iteration, the signal-to-noise ratio of each frequency band is calculated as the objective function, and a Frobenius norm regularization term is introduced to prevent overfitting. A gradient boosting strategy is used to construct a set of regression trees. The weight coefficients are gradually optimized through residual fitting to obtain the optimal weight matrix that maximizes the objective function.

[0013] Furthermore, the scene cognition model is used to generate six-dimensional scene environment labels, including: Millimeter-wave radar point cloud data is processed by a convolutional neural network to identify road types; the road types include, but are not limited to, highways, urban roads, tunnels, etc., to reflect the environmental scene. The bit map of the CAN bus protocol frame is analyzed to extract engine speed and air conditioning status, thus obtaining the vehicle status. The YOLO network was used to analyze infrared camera images to detect the number and location of passengers. Audio content is classified based on Mel-frequency cepstral coefficient features to obtain audio types; the audio types include music, speech, and prompts. Time attributes are marked using a dual calendar system of Gregorian and Lunar calendars; these time attributes include day and night periods, holidays, and weekdays. User preferences are modeled using a gated loop unit; the gated loop unit includes an explicit learning channel and an implicit learning channel, the explicit learning channel is used to record manually adjusted parameters, and the implicit learning channel is used to analyze user emotions.

[0014] Furthermore, the generation process of the loudness compensation strategy includes: Input historical scene labels into the reinforcement learning policy network; The encoder compresses the historical scene labels into multi-dimensional feature vectors; A dual-delay deep deterministic policy gradient algorithm is employed to search for the optimal combination of compensation parameters within the action space; the action space includes the low-frequency gain B. L ∈[-10,10]dB, intermediate frequency gain B M ∈[-5,5]dB, high-frequency gain B H ∈[-8,8]dB, sound field width factor W s ∈[0.8,1.2], the sound field width factor is used to control the expansion of the three-dimensional sound field; The value function of a state-action pair is evaluated using a reinforcement learning value network; the state represents a multi-dimensional feature vector, and the action represents an adjustment instruction for the optimal combination of compensation parameters. Offline iterative training is performed based on the output of the value function to obtain a pre-trained basic policy library; The pre-trained basic policy library is fine-tuned online based on real-time six-dimensional scene labels to obtain an adaptive compensation policy.

[0015] Furthermore, the implementation of multi-band independent compensation and sound field control at the speaker output includes: According to the compensation strategy, the low-frequency gain B L Adjust the equalizer slope in the 50-200Hz frequency band; According to the compensation strategy, the intermediate frequency gain B M Adjust the quality factor Q value of the bandpass filter in the 200-5000Hz frequency band; According to the compensation strategy, the high-frequency gain B H Adjust the gain parameters in the 5000-20000Hz frequency band; Through the sound field width factor W s Controlling the phase difference of multiple channels, when W s When W > 1, it increases the difference in delay between the left and right channels. s When the value is ≤1, the difference in delay between the left and right channels is reduced.

[0016] Furthermore, the control commands for the multi-band independent compensation and sound field control are executed synchronously by multiple DSP processors, and each speaker unit performs differentiated compensation based on its spatial position; the differentiated compensation includes: Perform high-frequency compensation on the front speakers: B H+ =B H +ΔH front Among them, B H+ The high-frequency gain after compensation is represented by ΔH. front This represents the gain correction for the front row, taken as 1.2dB. Low-frequency compensation for rear speakers: B L+ =B L +ΔL rear Among them, B L+ ΔL represents the compensated low-frequency gain. rear This represents the gain correction for the rear row, taken as 2.0dB. Enhance the sound field width of the side speakers: W s+ =W s ×k side Among them, W s+ k represents the compensated sound field width factor. side This represents the lateral enhancement factor, which is set to 1.15.

[0017] Compared with the prior art, the beneficial effects of the present invention are: First, this application demonstrates greater comprehensiveness and accuracy in environmental perception and noise processing. It employs a multi-modal sensor array comprised of in-vehicle and external microphone arrays, millimeter-wave radar, a barometric pressure sensor, and a vehicle CAN bus sensor. This array can simultaneously collect multi-dimensional vehicle environmental data, including raw audio signals, three-dimensional polar coordinates of noise sources, vehicle speed, engine speed, and air conditioning status, enabling a more complete capture of various factors affecting the loudness of the in-vehicle audio system. Simultaneously, by establishing a wind noise model based on fluid dynamics, a tire noise model based on spectral characteristics, an air conditioning noise model coupled with compressor speed, and a human voice noise model combining sound pressure and direction, and utilizing the GBDT algorithm to optimize the weighting coefficients of the separation matrix, it achieves precise separation of wind noise, tire noise, air conditioning noise, and human voice noise. This effectively solves the problems of mixed noise interference compensation and difficulty in targeted adjustment in existing technologies.

[0018] Secondly, this invention is more personalized and practical in scene adaptation and compensation execution. The scene recognition model generates six-dimensional scene labels covering environmental scenes, vehicle status, and riding status by using CNN to identify road types, YOLO to detect the number and distribution of passengers, and GRU (Gated Recurrent Unit) combined with explicit and implicit channel modeling of user preferences. This enables more accurate matching of different usage scenarios. The reinforcement learning policy network adopts a dual-delay deep deterministic policy gradient algorithm to search for optimal parameters in an action space that includes low-frequency, mid-frequency, and high-frequency gain and sound field width factors. Multi-band independent compensation and differentiated sound field control are executed by a multi-channel DSP processor. This achieves adaptive optimization of the compensation strategy and meets the auditory needs of passengers in different positions in the vehicle, overcoming the shortcomings of insufficient personalization of compensation strategies and uneven sound field distribution in existing technologies. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0020] Figure 1 A flowchart illustrating a loudness compensation method for a vehicle audio system provided by the present invention; Figure 2 A flowchart illustrating another loudness compensation method for a vehicle audio system provided by the present invention; Figure 3 A flowchart illustrating another loudness compensation method for a vehicle audio system provided by the present invention; Figure 4A flowchart illustrating another loudness compensation method for a vehicle audio system provided by the present invention; Figure 5 A flowchart illustrating another loudness compensation method for a vehicle audio system provided by the present invention; Figure 6 This is a schematic diagram of the loudness compensation system for a car audio system provided by the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0022] To address the technical problem of low loudness compensation accuracy caused by multiple noise sources in existing technologies, this application provides a loudness compensation method for vehicle audio systems, which includes: Vehicle environmental data is collected using a multimodal sensor array; Perform mixed noise separation processing on vehicle environmental data and output the characteristics of the separated independent noise sources; Based on the characteristics of independent noise sources and vehicle environment data, scene labels are generated through a scene cognition model, including environmental scene, vehicle status, riding status, audio content, time attribute, and user preference. The scene labels are input into the reinforcement learning policy network to generate a loudness compensation policy. Based on a loudness compensation strategy, multi-band independent compensation and sound field control are implemented at the speaker output.

[0023] Based on this, this application, through the synergistic effect of multi-dimensional data collection, precise noise separation, scenario-based strategy generation, and targeted compensation control, can effectively adapt to diverse in-vehicle scenarios and improve the accuracy of loudness compensation and the user's auditory experience.

[0024] like Figure 1 As shown in the embodiment of this application, a loudness compensation method for a vehicle audio system includes: S1. Collect vehicle environmental data through a multimodal sensor array, perform mixed noise separation processing on the vehicle environmental data, and output the characteristics of the separated independent noise sources.

[0025] The vehicle environmental data includes, but is not limited to, raw audio signals, vehicle speed data, engine operating parameters, occupant activity signals, and external environmental sound signals. This data comprehensively reflects the internal and external environmental conditions during vehicle operation, providing fundamental data support for subsequent mixed noise separation and scene recognition. Since vehicle environmental data contains characteristic information of different types of noise, data beneficial for mixed noise separation can be extracted. Mixed noise can interfere with the loudness output accuracy and auditory experience of in-vehicle audio systems. Therefore, through mixed noise separation processing, independent noise source characteristics can be extracted from complex data, providing a clear basis for subsequent scene judgment and compensation strategy generation.

[0026] In some implementations, mixed noise separation processing can employ common signal separation techniques such as Independent Component Analysis (ICA), Principal Component Analysis (PCA), and Wavelet Transform (WT). These techniques can effectively separate different noise sources by decomposing and reconstructing the mixed data.

[0027] It should be noted that the core of hybrid noise separation processing is to accurately extract the characteristics of various noise types that may exist in the vehicle environment, such as engine noise, wind noise, tire noise, and conversation noise of people inside the vehicle, so as to avoid mutual interference between different noise sources and affect the processing accuracy of subsequent steps.

[0028] For example, when a vehicle is driving at high speed, the environmental data collected by the multimodal sensor array will simultaneously include wind noise, tire noise, and engine noise. In this case, the ICA method is used to decompose the mixed audio signal. By analyzing the statistical independence of the signals, the low-frequency characteristics of wind noise, the mid-frequency fluctuation characteristics of tire noise, and the regular frequency band characteristics of engine noise can be gradually separated, providing a clear noise reference for subsequent scene label generation.

[0029] S2. Based on the characteristics of independent noise sources and vehicle environment data, scene labels are generated through a scene cognition model.

[0030] The scene tags include, but are not limited to, environmental scene, vehicle status, passenger status, audio content, time attributes, and user preferences. These tags reflect the comprehensive scene information of the vehicle's current location, providing accurate scene basis for the subsequent generation of loudness compensation strategies and ensuring that the compensation strategies can match actual usage needs.

[0031] In some implementations, the scene cognition model is used to integrate the features of independent noise sources with vehicle environment data to achieve multi-dimensional scene information judgment. It can be built through traditional machine learning algorithms such as Support Vector Machine (SVM) and Random Forest (RF), or it can be built through ensemble learning that integrates multiple basic models. The core is to use the key features in the data to achieve accurate classification and judgment of different scene dimensions.

[0032] It should be noted that the scene labels generated by the scene cognition model must cover all six dimensions of information and cannot be limited to a single or partial dimension. For example, labels cannot be generated based solely on the environmental scene and vehicle status while ignoring user preferences. Otherwise, the subsequent loudness compensation strategy will lack personalized adaptability and will fail to meet the auditory needs of different users.

[0033] For example, when the independent noise source characteristics show that low-frequency tire noise is the main feature, and the vehicle environment data shows that the vehicle speed is low and the road surface is bumpy, the environment scene can be judged as an urban unpaved road; combined with the engine idling speed parameter in the vehicle environment data, the vehicle status is marked as low-speed driving; if the environmental data detects sound signals in multiple different locations inside the vehicle, the riding status is determined to be multiple people riding; and combined with the audio content being news broadcast, the time attribute being morning rush hour, and the user preference being clear mid-to-high frequency information, a complete scene label covering six dimensions is finally generated.

[0034] S3. Input the scene labels into the reinforcement learning policy network to generate a loudness compensation policy.

[0035] Reinforcement learning (RL) is a machine learning method that uses an agent to interact with its environment to maximize cumulative rewards. It enables policy networks to optimize their decision-making abilities through continuous learning, gradually mastering the optimal loudness adjustment rules for different scenarios. By inputting scenario labels into the policy network, the network can learn the loudness adjustment requirements under different scenario combinations, resulting in loudness compensation strategies that match the characteristics of the current scenario and the user's needs.

[0036] In some implementations, loudness compensation strategies may include gain adjustment parameters for different frequency bands, sound field distribution control parameters, such as the gain adjustment range for low frequency, mid frequency, and high frequency bands, as well as the width control coefficient and focus position parameters of the in-vehicle sound field. These parameters together determine the final loudness output and sound field effect of the audio system.

[0037] It should be noted that the core of reinforcement learning strategy networks lies in self-optimization through continuous scene data input and feedback adjustment, rather than using fixed compensation parameters. For example, when the scene label "night + solo ride + soothing music" is received multiple times, the network will gradually optimize the compensation strategy based on the user's potential low volume preference, reducing the overall loudness and increasing the low-frequency warmth.

[0038] For example, when the scene is labeled "highway + single driver + rock music + daytime", the reinforcement learning strategy network will generate a compensation strategy based on historical learning experience. This strategy will increase the high-frequency gain by 4-6 dB, keep the mid-frequency gain stable, increase the low-frequency gain by 2-3 dB, and expand the sound field width to cover the driver's seat and the passenger seat. This will adapt to wind noise interference during high-speed driving and match the auditory needs of rock music.

[0039] S4. Based on the loudness compensation strategy, multi-band independent compensation and sound field control are implemented at the speaker output.

[0040] Among them, multi-band independent compensation and sound field control are designed to achieve precise loudness adjustment and balanced sound field distribution based on the propagation characteristics of different frequency bands and the auditory differences at different locations in the vehicle. This can typically be achieved using digital signal processing (DSP) technology, equalizer (EQ) adjustment methods, multi-channel phase control technology, etc. These technologies can transform loudness compensation strategies into actual audio output effects through real-time processing of audio signals.

[0041] In some implementations, the gain of the low-frequency, mid-frequency, and high-frequency bands can be independently adjusted by EQ based on the frequency band gain parameters in the loudness compensation strategy. At the same time, multi-channel phase control technology can be used to adjust the phase difference between speakers in different positions. For example, the phase difference between the left and right door speakers can be increased to expand the sound field width, or the phase difference between the front and rear speakers can be decreased to focus the sound field to the front row.

[0042] It should be noted that multi-band independent compensation and sound field control must ensure that the adjustment of different frequency bands does not interfere with each other, and that the sound field effect can cover all major seating positions in the vehicle, avoiding situations where the sound is too strong or too weak in some areas or that specific frequency bands are missing, so as to ensure that all passengers can have a balanced listening experience.

[0043] For example, when the loudness compensation strategy requires increasing the mid-frequency gain and reducing the sound field width, the gain value of the mid-frequency band (200-5000Hz) can be increased by 3 dB through EQ, while adjusting the phase difference between the front and rear speakers, so that the sound field is focused on the front driver and passenger seats. This not only improves the clarity of mid-frequency sounds such as human voices and musical instruments, but also avoids the decline in the rear listening experience caused by excessive sound diffusion.

[0044] Based on the above technical solutions, the loudness compensation method for vehicle audio provided in this application achieves comprehensive acquisition of vehicle environmental data through a multi-modal sensor array. Combined with existing noise separation technologies such as independent component analysis and wavelet transform, the accuracy of independent noise source feature extraction is ensured. Common algorithms such as support vector machines and random forests are used to construct a scene cognition model, realizing the complete generation of multi-dimensional scene labels. An adaptive loudness compensation strategy is generated by a reinforcement learning policy network, and multi-band compensation and sound field control are implemented through digital signal processing, equalizer adjustment, and other technologies. This method can be adapted to different types of vehicle sensors, audio systems, and usage scenarios, effectively improving the accuracy of loudness compensation and the user's auditory experience.

[0045] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2 As shown, the above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below: S101. Collect vehicle environmental data through a multimodal sensor array.

[0046] The multimodal sensor array includes an in-vehicle microphone array, an external microphone array, millimeter-wave radar, a barometric pressure sensor, and a vehicle CAN bus sensor. The in-vehicle microphone array is evenly distributed in the roof lining, while the external microphone array is integrated into the rearview mirrors and bumpers. The millimeter-wave radar is deployed at the four corners of the vehicle to monitor the location of noise sources. The barometric pressure sensor is installed on the window rails to detect the opening and closing status of the windows. The vehicle CAN bus sensor collects real-time data on vehicle speed, engine speed, and air conditioning compressor status. The collected vehicle environmental data includes raw audio signals, three-dimensional polar coordinates of noise sources, vehicle speed, engine speed, air conditioning compressor status, and air density.

[0047] In some implementations, the number of sensors in the multimodal sensor array needs to be determined based on the vehicle model and size: the number of in-vehicle microphone arrays is based on the vehicle length, with 6-8 microphones for sedans with a length of 4-5 meters and 8-10 microphones for SUVs with a length of 5-6 meters, ensuring no blind spots in the acquisition of in-vehicle sound signals; the external microphone array is configured in a combination of "rearview mirror + bumper", with one microphone integrated in each rearview mirror and two microphones integrated in each of the front and rear bumpers, for a total of 6 microphones to cover noise around the vehicle; the millimeter-wave radar is fixed at 4 units, deployed at the four corners of the vehicle (front left, front right, rear left, and rear right) to ensure 360-degree noise source location monitoring; the barometric pressure sensor is configured according to the number of windows, with one sensor installed on each window rail to achieve independent detection of the opening and closing status of a single window.

[0048] It should be noted that the data acquisition of each sensor must be synchronized in time, which is achieved through a unified timestamp calibration mechanism to avoid subsequent noise source feature matching deviations caused by data acquisition time differences. For example, the vehicle speed data acquired by the CAN bus sensor and the noise source location data acquired by the millimeter-wave radar must be ensured to correspond to the vehicle status at the same point in time.

[0049] For example, for a sedan with a length of 4.6 meters, the in-car microphone array is evenly distributed in the roof lining with 2 in the front row, 2 in the middle row, and 2 in the rear row, for a total of 6; there is 1 microphone on each of the left and right side mirrors, 1 on each of the left and right front bumpers, and 1 on each of the left and right rear bumpers, for a total of 6; there is 1 millimeter-wave radar at each of the four corners of the vehicle; there is 1 air pressure sensor on each of the 4 windows, namely the front left, front right, rear left, and rear right; the CAN bus sensors collect vehicle speed 10 times per second, engine speed 20 times per second, and air conditioning compressor status 5 times per second. All sensor data carries a timestamp with a precision of 1ms. Finally, the raw audio signals containing in-car conversation, outside wind noise, engine running sound, and air conditioning operating sound are collected, as well as vehicle environmental data at the corresponding time points: vehicle speed 80km / h, engine speed 2000rpm, air conditioning compressor operating status, windows fully closed, three-dimensional polar coordinates of the noise source (30° forward, 5 meters away), and air density 1.18kg / m³.

[0050] S102. Based on the collected vehicle environment data, construct feature templates for various types of noise.

[0051] The noise feature templates include wind noise, tire noise, air conditioning noise, and human voice noise, each constructed based on different parameters in the vehicle's environmental data: the wind noise feature template is based on parameters such as air density and vehicle speed, and is established using fluid dynamics formulas; the tire noise feature template is based on vehicle speed and tire parameters, and is established using spectral feature formulas; the air conditioning noise feature template is based on air conditioning compressor speed and refrigerant pressure, and is established using a compressor speed-pressure coupling model; and the human voice noise feature template is based on the sound intensity and sound source location of the original audio signal inside the vehicle, and is established using a sound pressure-direction joint detection model.

[0052] In some implementations, the wind-drying model P wind The mathematical expression for (f,t) is: Where f represents the sound frequency, t represents time, is the time-domain variable of the sound pressure signal, ρ represents air density, v represents vehicle speed, and C... d The turbulence coefficient is represented by A, which represents the projected area of ​​the airflow disturbance, referring to the effective projected area of ​​the local vehicle structure that generates vortex shedding in the direction of the airflow. vortex f represents the vortex shedding frequency. c σ represents the center frequency corresponding to the maximum sound pressure level, and σ represents the frequency distribution dispersion.

[0053] It should be noted that wind noise originates from the interaction between airflow and the vehicle's surface, especially from protruding parts such as rearview mirrors and windows, where vortices are generated and detached, as well as pressure fluctuations caused by airflow turbulence. These fluctuations propagate as sound waves, forming noise. In the formula, This is the dynamic pressure expression in fluid mechanics, directly reflecting the influence of vehicle speed v and air density ρ on airflow pressure—that is, the higher the vehicle speed and the greater the air density, the stronger the pressure disturbance of the airflow on the vehicle body, and the higher the basic strength of wind noise. Turbulence coefficient C d The projected area A of airflow disturbance further refines the impact of vehicle structure on wind noise, and C for different models d Unlike model A, by introducing these two parameters, the model can adapt to the wind noise characteristics of different vehicle models, thus improving its versatility.

[0054] also, The simulation describes the periodicity of vortex shedding, where vortex shedding generates periodic pressure pulsations with a frequency f. vortex Related to airflow speed and vehicle size; while The term is used to simulate the spectral distribution of wind noise, meaning that wind noise energy is not uniformly distributed across the entire frequency band, but rather concentrated at a certain center frequency f. c Nearby, the frequency distribution dispersion σ can quantify the energy diffusion range, reflecting the acoustic characteristics of wind noise spectrum exhibiting a bell-shaped distribution.

[0055] In some implementations, the tire noise model S tire The mathematical expression for (f) is: Where K represents the tire coefficient, α represents the vehicle speed influence coefficient, β represents the frequency attenuation index, γ represents the road surface attenuation coefficient, and a k f represents the harmonic amplitude of the k-th harmonic, where k represents the harmonic index. rotation δ represents the tire rotation frequency, and δ represents the Dirac delta function, used to describe discrete frequency components.

[0056] It should be noted that the main sources of tire noise include frictional vibration between the tire and the road surface, air compression vibration of the tire tread, and resonance of the tire structure. These mechanisms are all represented by model parameters. Because the higher the vehicle speed, the greater the contact frequency and frictional energy between the tire and the road surface, the tire noise intensity increases non-linearly. Therefore, a model parameter is used... It is directly related to vehicle speed v and tire noise intensity. The vehicle speed influence coefficient α can be determined through experimental fitting (usually taken as 2.5-3.0) to match the measured law that "for every 10km / h increase in vehicle speed, tire noise intensity increases by 3-5dB". The term describes the frequency attenuation characteristics of tire noise, that is, in the low-frequency range (200-1000Hz), tire noise mainly comes from tire structural vibration, and the intensity increases with frequency. βThe intensity is dominant in the high-frequency range (above 1000Hz), but decreases with increasing frequency due to air absorption and tire damping. At this point, e -γf The item is the main focus.

[0057] Harmonic terms in the model This further reflects the direct correlation between tire rotation frequency and vehicle speed; the periodic rotation of the tire causes vibrations that generate a fundamental frequency f. rotation With each order harmonic kf rotation The Dirac delta function can accurately describe these discrete frequency components—in actual measurements, tire noise spectra often show obvious "fundamental frequency-harmonic" peaks, and the model uses harmonic amplitude a k This characteristic can be fully reproduced by quantifying the intensity of each harmonic.

[0058] In some implementations, the air conditioner noise model S ac The mathematical expression for (f) is: Where η represents the mechanical vibration conversion efficiency, P represents the refrigerant pressure, and f fund Indicates the compressor's operating frequency, b m This represents the attenuation rate of the m-th harmonic.

[0059] It should be noted that the core source of air conditioning noise is the mechanical vibration of the compressor. The compressor works by periodically compressing the refrigerant, and the reciprocating motion of its crankshaft and piston generates periodic vibrations. These vibrations are transmitted to the vehicle interior through the pipes and the vehicle body, forming noise. This periodic vibration determines the narrow-band periodicity of air conditioning noise.

[0060] In the formula, the mechanical vibration conversion efficiency η and the refrigerant pressure P together determine the intensity of air conditioner noise: physically, the higher the refrigerant pressure, the greater the load on the compressor, and the more intense the vibration; η quantifies the conversion efficiency of vibration into noise, and the product of the two directly reflects the basic intensity of the noise. The term is used to describe the frequency composition of air conditioner noise, where the compressor operating frequency f fund It is the fundamental frequency of noise, while the nonlinear vibration of mechanical parts will generate harmonics of various orders mf. fund Harmonic attenuation rate b m This quantifies the intensity attenuation of higher-order harmonics, reflecting the decrease in intensity of higher-order harmonics as the order increases due to vibration energy loss.

[0061] In some implementations, the human voice noise model V speech The mathematical expression for (t,f) is: Where L0 represents the human voice trigger threshold, L p The value represents the sound intensity, μ represents the activation slope, θ represents the horizontal azimuth angle of the sound source, N represents the number of microphones in the vehicle, and Ψ represents the sound intensity value. VADIt represents the probability of human voice presence, obtained by outputting a pre-trained SNM classifier based on the energy and spectral characteristics of each frame in the audio signal.

[0062] It should be noted that, from the perspective of speech acoustics, human voice has a definite intensity threshold; only when the sound intensity L... p Only when the voice exceeds a certain threshold L0 can it be identified as human voice; this is the Sigmoid function in the model. This decision-making process can be smoothly simulated: when L p When L ≫0, the function value approaches 1, which is then confirmed as human voice; when L p When L < 0, the function value approaches 0, which is then determined to be noise.

[0063] From the perspective of array signal processing principles The term represents the beamforming gain factor of the microphone array, where N is the number of microphones in the vehicle, and θ is the horizontal azimuth angle of the human voice source. The gain of the microphone array can be adjusted according to the direction of the sound source—the gain is maximum when θ corresponds to the target direction, which can enhance the human voice signal in that direction; when θ corresponds to other directions, the gain is minimum, which can suppress interference signals. Additionally, the probability of human voice presence... The intermittent nature of human voice can be quantified—human voice is not continuous, but rather alternates between utterance and pause. The probability of human voice presence is output in real time through a speech activity detection algorithm, which can dynamically match the intermittent characteristics of human voice.

[0064] In some implementations, key parameters of each noise model need to be determined through specialized testing and data fitting: air density ρ is calculated by combining air pressure values ​​collected by a pressure sensor with the ideal gas law PV=nRT, where temperature T is taken from data from an in-vehicle temperature sensor; the turbulence coefficient C in the wind noise model... d Referring to wind tunnel test data of similar vehicle models, the coefficient is 0.28-0.35 for sedans and 0.35-0.42 for SUVs. The tire coefficient K in the tire noise model is determined by consulting the technical manuals provided by the tire manufacturers. For example, the K value of Michelin Primacy 4 tires is 1.2×10^-6, and the K value of Goodyear Eagle F1 tires is 1.1×10^-6. The human voice trigger threshold L0 in the human voice noise model is determined by collecting the normal conversation sound intensity data of multiple groups of users of different genders and ages, taking the average value and subtracting 3dB, which is usually between 55-60dB.

[0065] It should be noted that each noise feature template needs to be dynamically updated as the vehicle's environmental data changes in real time. For example, when the vehicle speed increases from 60km / h to 100km / h, the vehicle speed parameter v in the wind noise model and the parameter v in the tire noise model need to be updated simultaneously. α item.

[0066] S103. Perform mixed noise separation processing on the vehicle environment data and output the characteristics of the separated independent noise sources.

[0067] Hybrid noise separation processing first constructs a separation matrix W=[w ij ] m×n Then, the noise feature templates are input into the separation matrix, and the signal-to-noise ratio of each frequency band is maximized through iterative solution. Finally, four independent noise source features are output: wind noise, tire noise, air conditioning noise, and human voice noise. Here, m is the number of microphones, n is the number of noise source types, and w... i,j This represents the weighting coefficient of the i-th microphone for the j-th noise source.

[0068] In some implementations, the weight coefficients of the separation matrix are optimized using the Gradient Boosting Decision Tree (GBDT) algorithm, with the following specific steps: The first step is to initialize the weight matrix W0, where each w... i,j The initial value is set to 1 / n, where n is the number of noise source types. Here, n=4, so the initial value is 0.25. The second step involves inputting W0 and each noise feature template into GBDT, using the signal-to-noise ratio (SNR) of each frequency band (SNR = 10lg(signal power / noise power)) as the objective function, and simultaneously introducing the Frobenius norm regularization term ||W||. Fro =√(ΣΣw i,j ²) To prevent overfitting, the regularization coefficient λ is determined through 5-fold cross-validation, with a value range of 0.01-0.05, and is usually taken as 0.03; The third step is to calculate the objective function for each w in each iteration. i,j The gradient is used to construct a set of regression trees. The depth of each tree in the set is set to 3-5 layers and the number of leaf nodes is set to 8-12. The weight coefficients are adjusted by fitting the residuals. The number of iterations is set to 100-200. The iteration stops when the improvement of the objective function is less than 0.1dB for 5 consecutive iterations. Fourth step: After the iteration is complete, output the optimal weight matrix W. opt Furthermore, in the signal-to-noise ratio calculation, the signal power is calibrated using the reference power of the original audio signal collected by the microphone in a noise-free environment, and the noise power is calculated by integrating the power spectral density of each noise source characteristic after separation in the corresponding frequency band (20-20000Hz divided into 1 / 3 octave bands).

[0069] In some implementations, the objective function is calculated using the following formula: Where λ is the regularization coefficient, and N freq For the total number of frequency bands, SNR k (W) represents the signal-to-noise ratio of the k-th frequency band under the action of the separation matrix W.

[0070] It should be noted that the dimension of the separation matrix must match the total number of microphones and the number of noise source types. For example, if there are 12 microphones in total (6 inside the car and 6 outside), and 4 types of noise sources, the separation matrix will have a dimension of 12×4. At the same time, the iterative solution must be performed independently by frequency band to ensure that noise in different frequency bands can be effectively separated and that no frequency band is not completely separated.

[0071] For example, the total number of microphones is 12, namely 6 inside the vehicle and 6 outside the vehicle. There are 4 types of noise sources. The separation matrix W is initially a 12×4 matrix, and each w... i,j =0.25; the regularization coefficient λ was set to 0.03 after 5-fold cross-validation, the regression tree depth was set to 4 layers, the number of leaf nodes was 10, and the number of iterations was 150; in the 200Hz frequency band, the signal reference power for anechoic chamber calibration was 0.1W, the mixed noise power was 0.08W, and after GBDT optimization, the wind noise characteristic power was 0.03W, tire noise was 0.025W, air conditioning noise was 0.015W, and human voice noise was 0.01W. The signal-to-noise ratio in this frequency band was improved to 10lg(0.1 / 0.08)=0.97dB. Finally, four independent noise source characteristics were output, of which the wind noise characteristic was a continuous fluctuation in the low frequency band 50-200Hz, the tire noise characteristic was a regular pulse in the mid frequency band 200-1000Hz, the air conditioning noise characteristic was a harmonic of a specific frequency (such as 500Hz), and the human voice noise characteristic was an intermittent signal in the mid-high frequency band 1000-5000Hz.

[0072] Based on the above technical solution, in this embodiment, S1 not only ensures the comprehensiveness and synchronization of vehicle environmental data collection, but also achieves accurate separation of mixed noise through targeted noise feature template construction and GBDT-optimized separation matrix. This provides high-quality independent noise source feature support for subsequent scene label generation and loudness compensation strategy formulation, avoiding compensation deviation problems caused by incomplete data or incomplete noise separation.

[0073] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 3 As shown, the above S2 can be implemented through the following S201, S202 and S203, which are explained in detail below: S201. Preprocess the independent noise source characteristics and vehicle environment data to construct the input dataset for the scene cognition model.

[0074] Preprocessing is crucial for the collaborative adaptation of multi-source data, ensuring that the features of independent noise sources and vehicle environmental data remain consistent in time and dimensions, thus laying the foundation for subsequent multi-dimensional scene feature extraction. The preprocessing process encompasses three main operations: data time synchronization, outlier removal, and feature standardization.

[0075] In some implementations, outlier removal uses the 3σ criterion, calculating the mean μ and standard deviation σ for each data dimension and removing data exceeding the range [μ-3σ, μ+3σ]. Feature standardization uses the Min-Max standardization method, mapping the data to the [0,1] interval, as shown in the formula. , where x min and x max The value of x was determined by collecting and statistically analyzing historical data under multiple vehicle conditions, such as vehicle speed. min Set to 0km / h, x max Set to 200 km / h, sound pressure level x min Set to 30dB, x max Set it to 120dB.

[0076] It should be noted that the preprocessing stage needs to retain the correlation between the features of independent noise sources and the vehicle environment data. For example, wind noise features should be bound to vehicle speed and the three-dimensional polar coordinates of the noise source, and tire noise features should be bound to vehicle speed and road surface type. This will prevent isolated data from causing the inability to form a complete logical chain when extracting scene features in the later stages.

[0077] S202. Based on the preprocessed dataset, extract six types of scene features by dimension, covering all the constituent dimensions of the scene label.

[0078] The six dimensions of the scene label need to be obtained through targeted feature extraction methods. Each method needs to match the characteristics of the corresponding data. The environmental scene depends on the spatial features of millimeter-wave radar point cloud data, the vehicle status depends on the operating parameters of the CAN bus, the riding status depends on the image features of the infrared camera, the audio content depends on the spectral features of the original audio signal, the time attribute depends on the time stamp of the calendar system, and the user preference depends on the behavioral features of historical interaction data. All dimensional features need to be output in vector form to facilitate subsequent model fusion.

[0079] In some implementations, environmental scene features are extracted using a convolutional neural network (CNN). The input is three-dimensional rasterized data of millimeter-wave radar point clouds, and the output is an environmental scene feature vector, such as [1,0,0] for highways, [0,1,0] for urban roads, and [0,0,1] for tunnels. Vehicle status features are obtained by parsing the bit map of the CAN bus protocol frame. Three core parameters are extracted: engine speed, air conditioning compressor status, and window opening / closing status. The engine speed is quantified into three levels: [low speed 0-1500rpm, medium speed 1501-3500rpm, and high speed above 3501rpm]. The air conditioning status is divided into on / off, and the window status is divided into fully closed / partially open / fully open. The final combination is a 3-dimensional vehicle status feature vector. The passenger status features are extracted through a YOLO network. The YOLO input is the in-vehicle image captured by an infrared camera, and the output is the number of passengers and their distribution location. Audio content features are extracted using Mel-Frequency Cepstral Coefficients (MFCCs). The original audio signal is framed, Hanning windows are added, and fast Fourier transform is performed. After passing through 26 Mel filter banks, the 13th-order MFCCs are calculated to form a 13-dimensional audio content feature vector. The time attribute features are marked by the Gregorian and Lunar calendar systems, and two types of information are extracted: day and night time period (0-6 am, 6-12 am, 12-6 pm, 6-11 pm) and date type (weekday / weekend / holiday), which are encoded into a 5-dimensional time attribute feature vector. User preference features are extracted through a gated recurrent unit (GRU). The GRU includes an explicit learning channel, which records parameters such as frequency band gain and volume that users manually adjust over 30 days and weights them according to the frequency of adjustment, and an implicit learning channel, which analyzes the user's audio selection type and listening duration, and determines the preference style by combining audio content features, and outputs a user preference feature vector.

[0080] In some implementation methods, the methods for obtaining user preference features may specifically include: The first step is to construct a parameter acquisition and weighted calculation system for the explicit learning channel.

[0081] The explicit learning channel needs to collect user manual adjustments to the car audio system within 30 days in real time. The collected parameters include low-frequency gain adjustments (50-200Hz), mid-frequency gain adjustments (200-5000Hz), high-frequency gain adjustments (5000-20000Hz), volume, and sound field width factor. Each adjustment record must be bound to a corresponding timestamp and the scene tag at that time to ensure the relevance of the adjustment behavior to the usage scenario. Weight calculation uses a time decay strategy, accumulating a weighted sum based on "adjustment value × corresponding weight coefficient." Then, Min-Max standardization maps the weighted sum to the [0,1] interval, ultimately forming a 5-dimensional explicit feature sub-vector containing low-frequency gain preference, mid-frequency gain preference, high-frequency gain preference, volume preference, and sound field width preference.

[0082] The second step is to build an audio interaction analysis mechanism for implicit learning channels.

[0083] The implicit learning channel first categorizes and statistically analyzes users' audio selection behavior over a 30-day period, classifying audio types into five categories: rock, classical, pop, news or talk shows, and notification sounds. The listening time for each category is calculated daily, and the proportion of each category's listening time to the total listening time is calculated. Simultaneously, Mel-frequency cepstral coefficients (MFCCs) are extracted for each audio category, and the 13th-order mean and variance of the MFCCs are calculated as audio style features. Combined with scene tags, the channel analyzes users' audio selection tendencies in different scenarios. For example, it counts the frequency of users choosing rock audio in a "highway scenario" and the frequency of choosing news audio in a "city traffic jam scenario," converting the scene-audio selection frequency into a quantified value. The audio duration proportion, MFCC style features, and scene-audio selection quantified values ​​are concatenated to form a 12-dimensional implicit feature sub-vector.

[0084] The third step is to use GRU to fuse explicit and implicit features and output the preference vector.

[0085] The input layer of GRU concatenates explicit and implicit feature vectors into a 17-dimensional input vector, which is then fed into the GRU network in 30 time steps (in days). The number of hidden layer units in GRU can be set to 64. The update gate controls the degree of retention of historical features, and the reset gate controls the degree of influence of the current input. The hidden layer output of each time step integrates the user interaction features and historical preference features of that time step.

[0086] The output layer of the GRU maps the hidden layer output to a 64-dimensional user preference feature vector through a fully connected layer, where each dimension corresponds to a quantized value for a different preference dimension. During model training, user interaction data from several days of historical data, including explicit adjustment records and implicit audio selection records, is used as training samples. The deviation between the user's subsequent adjustment behavior and the model's predicted preferences is used as the loss function. The Adam optimizer is used for iterative training until the model's prediction deviation falls below a preset threshold, ensuring that the output user preference feature vector accurately reflects the user's long-term usage habits.

[0087] S203. Input the six types of scene features into the scene cognition model, and output complete scene labels through feature fusion and classification.

[0088] The core function of the scene cognition model is to effectively fuse multi-dimensional features, integrating six scattered feature vectors into a unified scene representation, and then mapping it to human-understandable scene labels through a classification module. The model needs to have strong generalization ability and be adaptable to the scene differences of different car models and users. The feature fusion stage needs to make full use of the complementarity of features in each dimension, and the classification stage needs to ensure that the label output covers all six dimensions without any missing or redundant labels.

[0089] In some implementations, the scene cognition model adopts an architecture of "feature concatenation + attention mechanism + fully connected classification": First, the six feature vectors output by S202 are concatenated into a total feature vector of 221 dimensions; Then, a multi-head attention mechanism is introduced to assign weights to different dimensions of the total feature vector. The weights for user preference features are set to 0.3, environmental scene features to 0.25, vehicle state features to 0.2, and other dimensions to 0.08-0.09. The weights are iteratively optimized through training data to ensure that the key dimensions play a dominant role in label generation. Finally, the attention-weighted feature vector is input into a 3-layer fully connected layer, with ReLU as the activation function. The output layer uses the Softmax function to predict the probability of candidate categories for each scene dimension, and takes the category with the highest probability as the label result for that dimension.

[0090] During model training, the dataset uses multiple sets of labeled data from different scenarios, the loss function is cross-entropy loss, and the optimizer is Adam. Training stops when the accuracy on the validation set reaches 97% or higher.

[0091] It should be noted that the scene cognition model needs to have online update capabilities. When new scenes appear during vehicle use, the model parameters can be updated through incremental learning mechanisms to avoid the model's adaptation lag to new scenes. At the same time, the label output needs to adopt a standardized format, and the label descriptions for each dimension need to be consistent. For example, environmental scenes should only use five categories of descriptions: "highway, urban road, rural road, tunnel, and parking lot" to avoid confusion caused by synonyms such as "expressway" and "highway".

[0092] Based on the above technical solution, in this embodiment, S2 ensures the quality of input data through preprocessing in S201, guarantees the accuracy of scene features through dimensional extraction in S202, and achieves the integrity of scene labels through model fusion in S203, thus constructing a transformation link from multi-source data to complete scene labels. This process fully utilizes the adaptability advantages of technologies such as convolutional neural networks, YOLO, and gated recurrent units to different types of data. At the same time, it highlights the influence of key scene dimensions through attention mechanisms. The generated scene labels can accurately reflect the current comprehensive environment of the vehicle and user needs, providing a high-quality input basis for subsequently inputting the labels into the reinforcement learning strategy network and generating personalized loudness compensation strategies.

[0093] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 4 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Construct a historical scene label dataset and compress it into a multi-dimensional feature vector using an encoder.

[0094] The historical scene label dataset serves as the training foundation for the reinforcement learning policy network. It comprises scene labels recorded over several months of vehicle usage under different scenarios, covering a full range of dimensions including environmental scene, vehicle status, passenger status, audio content, time attributes, and user preferences. The encoder's core function is to compress high-dimensional historical scene labels into low-dimensional, lossless feature vectors, reducing subsequent network computation while preserving key scene information.

[0095] In some implementations, the encoder aims to minimize the error in reconstructing the original label from the compressed feature vector during training, and uses mean squared error as the loss function. The sample size of the historical scene label dataset needs to be determined through "scene coverage verification," that is, collection stops when new samples no longer add uncovered scene combinations; the encoder output dimension is determined through "information retention rate test," requiring that the compressed feature vector retains at least 95% of the information of the original label; where information retention rate = 1 - reconstruction error / original label variance.

[0096] S302. The basic policy library is trained using the dual-delay deep deterministic policy gradient algorithm.

[0097] The Dual-Delay Deep Deterministic Policy Gradient (DDPG) algorithm is the core algorithm of the reinforcement learning policy network, used to search for the optimal combination of compensation parameters in the action space. The action space contains four key parameters: low-frequency gain (corresponding to the 50-200Hz frequency band), mid-frequency gain (corresponding to the 200-5000Hz frequency band), high-frequency gain (corresponding to the 5000-20000Hz frequency band), and sound field width factor. The reinforcement learning value network is used to evaluate the value of "state-action pairs," where the state is the multi-dimensional feature vector output by S301, and the action is the combination of compensation parameters. The value evaluation results are used to guide policy optimization. The basic policy library is a set of "scene-compensation parameter" mappings generated after offline training, providing a benchmark for subsequent online fine-tuning.

[0098] In some implementations, the DDPG algorithm includes an Actor network (policy network) and a Critic network (value network): The Actor network receives the scene label multi-dimensional feature vector compressed by the encoder, which is the "state" in reinforcement learning. It searches for the optimal loudness compensation parameter combination that fits the current state within a preset action space and outputs specific compensation parameters that conform to the action space range, including low-frequency gain, mid-frequency gain, high-frequency gain, and sound field width factor. The Critic network receives the "state-action pair" composed of the "scene label multi-dimensional feature vector + compensation parameter combination output by the Actor network" and evaluates the adaptation value of the "state-action pair" in the current scene. That is, it judges whether the compensation parameter combination can effectively improve the loudness adaptation of the car audio system and the user's listening experience, and outputs a single-dimensional value score. This score will serve as the basis for the Actor network to optimize the compensation parameter combination and drive the Actor network to continuously generate better strategies. The action space parameter range was determined through hardware testing and user listening experiments, resulting in a low-frequency gain range of [-10, 10] dB, a mid-frequency gain range of [-5, 5] dB, a high-frequency gain range of [-8, 8] dB, and a sound field width factor range of [0.8, 1.2]. Training was stopped when the evaluation error of the value network was below 0.02 for 10 consecutive iterations, and a basic policy library was generated.

[0099] For example, during training, the 32-dimensional feature vector of a scene "highway + 2 people in the front row + rock music" is input into the Actor network, and the output compensation parameter combination is "low frequency gain +4dB, mid frequency gain +2dB, high frequency gain +3dB, sound field width factor 1.1". The Critic network evaluates the value score of this combination as 0.92 (out of 1.0), which is higher than other parameter combinations. After multiple iterations of training, the optimal compensation parameters corresponding to this scene in the basic policy library are fixed, and subsequent similar scenes can be directly called.

[0100] S303. Based on real-time scene labels, the basic strategy library is fine-tuned online to generate the final loudness compensation strategy.

[0101] The real-time scene label, which is the comprehensive scene label of the current vehicle output by S2, is used for online fine-tuning. The core of this fine-tuning is to adjust the compensation parameters based on the differences between the real-time scene and similar scenes in the basic strategy library, so that the strategy is more suitable for the current actual situation. The fine-tuning process must ensure real-time performance, with the delay from receiving the real-time label to outputting the compensation strategy not exceeding 100ms, to avoid affecting the user's auditory experience.

[0102] In some implementations, the specific process of online fine-tuning may include: The first step is to calculate the similarity between the real-time scene label and the scenes in the basic policy library. The cosine similarity algorithm is used, and the similarity calculation formula is "Similarity = (Real-time feature vector · Feature vector in library) / (||Real-time feature vector|| × ||Feature vector in library||)". The compensation parameters corresponding to the three scenes in the library with the highest similarity are selected as the benchmark parameters.

[0103] The second step is to determine the fine-tuning step size, which is inversely proportional to the similarity: when the similarity is ≥0.9, the step size is set to 0.5dB (gain parameter) and 0.05 (sound field width factor); when the similarity is 0.8≤similarity<0.9, the step size is set to 1dB and 0.1; when the similarity is <0.8, the step size is set to 1.5dB and 0.15. The upper limit of the step size is determined by the "auditory comfort test" to ensure that a single fine-tuning will not cause a sudden change in sound.

[0104] The third step is to adjust the parameters according to the specific dimensions of the real-time scenario: for example, if the real-time scenario is "3 people in the back row," while the basic library mostly contains "2 people in the front row," then the low-frequency gain needs to be increased, with an adjustment range of +1.5dB to the basic low-frequency gain parameter; if the real-time scenario is "news audio," while the basic library mostly contains "music," then the high-frequency gain needs to be reduced to avoid harsh audio, with an adjustment range of -1dB to the basic high-frequency gain parameter. After fine-tuning, it is necessary to verify whether the parameters meet the hardware constraints. If they exceed them, they are truncated to the limit value, and finally, a real-time loudness compensation strategy is generated.

[0105] It should be noted that online fine-tuning should avoid over-adjustment. When the similarity between the real-time scene and the scene in the base library is ≥0.95, the base parameters can be used directly without fine-tuning. At the same time, the fine-tuned parameters should be recorded in the experience pool. Every 10,000 sets of fine-tuning data should be accumulated to perform an incremental update on the base strategy library to improve the adaptability of the library.

[0106] For example, the real-time scene label is "city road (vehicle speed 60km / h) + mid-speed (engine 2500rpm) + 3 people in the back seat + news voice + weekday afternoon + low-frequency preference". Its 32-dimensional feature vector has a similarity of 0.88 with the scene in the basic strategy library "city road + mid-speed + 2 people in the front seat + news voice + weekday afternoon + low-frequency preference". The corresponding baseline parameters are "low-frequency gain +3dB, mid-frequency gain +1dB, high-frequency gain +1dB, sound field width factor 1.0". Based on the similarity, the fine-tuning step size is determined to be 1dB and 0.1. Combining the "3 people in the back seat" dimension, the low-frequency gain is adjusted by +1dB, and the high-frequency gain is adjusted by -1dB in combination with the "news voice" dimension. The final output compensation strategy is "low-frequency gain +4dB, mid-frequency gain +1dB, high-frequency gain 0dB, sound field width factor 1.05". This strategy takes only 60ms from receiving the real-time label to output, which meets the real-time requirements.

[0107] Based on the above technical solution, in this embodiment, S3 constructs a basic policy library covering multiple scenarios through DDPG offline training in S302 to ensure the universality of the policy; through online fine-tuning in S303, it adapts to the special characteristics of real-time scenarios and ensures the accuracy of the policy, realizing a closed loop of "offline accumulation-online adaptation". It not only uses historical data to optimize the policy foundation, but also adjusts in real time to cope with changes in the scenario. At the same time, it strictly controls the parameter range and latency to ensure that the loudness compensation policy takes into account effectiveness, safety and real-time performance, providing a scientific and reasonable parameter basis for subsequent compensation execution at the speaker end.

[0108] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 5 As shown, the above S4 specifically includes the following S401 to S403: S401, analyze loudness compensation strategy, and extract core parameters of multi-band gain and sound field control.

[0109] The loudness compensation strategy includes low-frequency gain, mid-frequency gain, and high-frequency gain required to achieve independent compensation of multiple frequency bands, as well as the sound field width factor required to achieve sound field control. The analysis process needs to accurately extract these core parameters and ensure that the parameter format matches the execution interface of the vehicle audio control system to provide clear instructions for subsequent compensation operations.

[0110] In some implementations, parameter parsing employs a "strategy instruction structured parsing" process: first, the instruction data packet of the compensation strategy is read, and the low-frequency gain B is extracted according to a preset field separator. L Intermediate frequency gain B M High-frequency gain B H Sound field width factor W s Four types of parameters; then, hardware compatibility verification is performed on the extracted parameters to determine B. L Is it within the range of [-10, 10] decibels, B? M Is it within the range of [-5, 5] dB? H Is it within the range of [-8, 8] dB? W s If the value is within the range of [0.8, 1.2], it will be automatically truncated to the corresponding limit value if it exceeds the range. Finally, the verified parameters are converted into a binary instruction format that the audio control system can recognize. The conversion rules are determined by the communication protocol agreed upon with the audio hardware manufacturer to ensure that the instruction transmission is without deviation.

[0111] For example, in a loudness compensation strategy instruction data packet, the extracted parameter is B. L =+4 decibels, B M =+2 dB, B H =+3 decibels, W s=1.1, all values ​​are within a reasonable range after verification. After conversion to binary instructions, a checksum of 0x5A is appended. The final instruction output to the audio control system is "01 04 02 03 0B 5A", where "01" represents the instruction type, and "04", "02", and "03" correspond to B respectively. L B M B H The binary value, "0B" corresponds to W. s =1.1 is the binary value, and "5A" is the check code.

[0112] S402. Perform multi-band independent compensation based on frequency band gain parameters.

[0113] Among them, the multi-band independent compensation adjusts the characteristics of different frequency bands of the car audio separately to ensure that the sound of each frequency band remains clear and balanced in a noisy environment. Specifically, it covers three core frequency bands: low frequency, mid frequency, and high frequency. Each frequency band has its own dedicated adjustment component and adjustment logic, which do not interfere with each other.

[0114] In some implementations, the specific methods for compensation in each frequency band are as follows: Low-frequency band compensation is achieved by adjusting the equalizer slope, and the slope adjustment formula is: "Slope value = base slope × (1 + B)". L / 20), where the base slope is determined through audio hardware testing—in an anechoic chamber environment, a standard sine wave signal from 50 to 200 Hz is played, and the equalizer slope is gradually adjusted until the sound pressure level of this frequency band deviates by ≤1 dB at various locations inside the car. This slope is the base slope, typically taken as 0.5 dB / octave. Mid-frequency compensation is achieved by adjusting the quality factor Q of the bandpass filter. The Q adjustment rule is "Q=Q0+B". M ×0.1”, Q0 is the base quality factor, which is the average score of mid-frequency speech clarity at different Q values ​​determined through user listening experiments, with a default value of 1.2; high-frequency compensation is achieved by directly adjusting the high-frequency gain parameter, which will adjust the calibrated B... H The gain control module of the high-frequency amplifier is input to the value, and the amplifier controls the gain according to B. H The output voltage is changed in real time, thereby adjusting the sound pressure level in the high-frequency band.

[0115] It should be noted that multi-band compensation should be performed in the order of "low frequency → mid frequency → high frequency"; at the same time, the output sound pressure level of each frequency band should be monitored in real time. When the sound pressure level exceeds 85 dB, the gain of the corresponding frequency band should be automatically reduced to ensure hearing safety.

[0116] For example, parameter B extracted based on S401 L =+4 decibels, B M =+2 dB, B H=+3dB, the basic slope of the low frequency band is 0.5dB / octave, and the calculated adjusted slope is 0.5×(1+4 / 20)=0.6dB / octave; the basic Q0 of the mid frequency band is 1.2, and the calculated adjusted Q is 1.2+2×0.1=1.4; the high frequency band directly increases the amplifier gain by 3dB, ultimately achieving a multi-band independent compensation effect of increasing the slope of the 50 to 200 Hz band, optimizing the Q value of the 200 to 5000 Hz band, and increasing the sound pressure level of the 5000 to 20000 Hz band.

[0117] S403 performs sound field control and differential compensation based on the sound field width factor and speaker position.

[0118] Among them, sound field control achieves the expansion degree of three-dimensional sound field by adjusting the phase difference of multi-channel, while differential compensation makes targeted parameter adjustments based on the differences in auditory characteristics of speakers in different positions in the vehicle. The two work together to ensure that all passengers in the vehicle can obtain a balanced and comfortable auditory experience. Moreover, the operation needs to be executed synchronously by multiple DSP processors to ensure the consistency of signal processing.

[0119] In some implementation methods, the specific execution process of sound field control and differential compensation is as follows: First, based on the sound field width factor W s Adjust the phase difference between the left and right channels: When W s When the value is greater than 1, increase the delay difference between the left and right channels. The delay difference adjustment formula is: "Delay difference = Base delay difference × W". s The fundamental delay difference is determined by sound field uniformity testing, and is usually taken as 0.2 milliseconds. When Ws≤1, reduce the delay difference between the left and right channels. The adjustment formula is "delay difference = basic delay difference × Ws".

[0120] Next, implement differentiated compensation: High-frequency compensation is applied to the front speakers, and the high-frequency gain B after compensation is... H+ =B H +ΔH front ΔH front This is the gain correction for the front row, determined through front row hearing sensitivity testing, with a default value of 1.2 dB. Low-frequency compensation is applied to the rear speakers, and the low-frequency gain B after compensation is... L+ =B L +ΔL rear ΔL rear This is the gain correction amount for the rear seats, determined through rear-seat hearing sensitivity testing, with a default value of 2.0 dB. The sound field width is enhanced for the side speakers, and the enhanced sound field width factor W s+ =W s ×kside k side This is the lateral enhancement factor, determined through lateral sound field coverage testing, with a default value of 1.15.

[0121] All compensation instructions are executed synchronously by a multi-channel DSP processor, with the number of DSPs matching the number of speakers, and are synchronized using a clock synchronization signal.

[0122] It should be noted that the parameter of differential compensation (ΔH) front ΔL rear k side The method needs to be fixed according to the vehicle model. Different vehicle models have different interior space structures and need to be re-determined through listening tests to avoid directly using the same method, which may result in poor compensation effect.

[0123] For example, based on Ws=1.1 extracted by S401 and the base delay difference=0.2 milliseconds, the left and right channel delay difference is calculated to be 0.2×1.1=0.22 milliseconds; combined with S401's B... H =+3dB, BL=+4dB, B after front speaker compensation H+ =3 + 1.2 = 4.2 dB, after compensation of the rear speakers B L+ =4 + 2.0 = 6.0 dB, W after side speaker compensation s+ =1.1×1.15=1.265; Under the control of a 48kHz clock synchronization signal, the 6-channel DSP synchronously executes the above parameter adjustments, ultimately achieving an enhanced in-car sound field width, clear high frequencies in the front row, and full low frequencies in the rear row.

[0124] It should be noted that since multi-band independent compensation can only achieve overall equalization optimization of the entire vehicle's audio frequency bands, it cannot take into account the differences in auditory characteristics of different spatial positions in the vehicle. Front passengers are more sensitive to high-frequency sounds, while rear passengers are less sensitive to low-frequency sounds due to their distance from the speakers and the influence of the vehicle's structure. There is also a problem of insufficient sound field coverage in the lateral area. Therefore, after multi-band independent compensation, differentiated compensation is applied to the front speakers, rear speakers, and lateral speakers.

[0125] Based on the above technical solution, in the embodiment of this application, S4 achieves balanced optimization of sound in each frequency band through multi-band independent compensation. By controlling the sound field and using differentiated compensation to adapt to the differences in the in-vehicle space and passenger position, it not only ensures the accuracy and real-time performance of loudness compensation, but also solves the problem of poor auditory experience in some positions, and finally achieves high-quality auditory effects in all scenarios and positions in the vehicle.

[0126] The loudness compensation method for vehicle audio provided in this application embodiment can be applied to, for example... Figure 6 In the car audio loudness compensation system shown, such as Figure 6As shown, the system includes: a data acquisition module, a data processing module, a strategy generation module, and a compensation execution module; among which, The data acquisition module is used to collect vehicle environmental data through a multimodal sensor array, including raw audio signals, three-dimensional polar coordinates of noise sources, vehicle speed, engine speed, air conditioning compressor status, and air density. The data processing module is used to perform mixed noise separation processing on the vehicle environment data acquired by the data acquisition module. By constructing a separation matrix and iteratively solving to maximize the signal-to-noise ratio of each frequency band, it outputs independent noise source features such as wind noise features, tire noise features, air conditioning noise features, and human voice noise features. At the same time, based on the independent noise source features and vehicle environment data, it generates a six-dimensional scene label containing environmental scene, vehicle status, riding status, audio content, time attribute, and user preference through a scene cognition model, providing accurate scene basis for the generation of compensation strategy. The policy generation module is used to input the six-dimensional scene labels output by the data processing and scene cognition module into the reinforcement learning policy network. First, the scene labels are compressed into multi-dimensional feature vectors by the encoder. Then, the optimal combination of compensation parameters is searched in the action space by the dual-delay deep deterministic policy gradient algorithm. Combined with value network evaluation and offline iterative training, a pre-trained basic policy library is formed. Finally, the basic policy library is fine-tuned online according to the real-time scene labels to generate a loudness compensation policy that is adapted to the current scene. The compensation execution module is used to generate a loudness compensation strategy based on the output of the compensation strategy generation module. It implements multi-band independent compensation and sound field control at the speaker output end to ensure that passengers in different positions in the vehicle can obtain a balanced and comfortable listening experience.

[0127] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0128] Working principle of the invention: First, a multimodal sensor array comprehensively collects vehicle environmental data, acquiring key information such as raw audio signals, noise source location, vehicle speed, engine operating status, and air conditioning operating status. Next, the collected vehicle environmental data undergoes mixed noise separation processing. By constructing feature templates adapted to different noise types, combined with matrix operations and iterative optimization, the characteristics of independent noise sources such as wind noise, tire noise, air conditioning noise, and human voice noise are accurately separated from the mixed data, eliminating the interference of various noises on the judgment of sound volume.

[0129] Subsequently, based on the isolated independent noise source features and vehicle environment data, a six-dimensional scene label covering environmental scene, vehicle status, riding status, audio content, time attribute, and user preferences is generated using a scene cognition model, thereby achieving a comprehensive depiction of the current vehicle usage scenario.

[0130] The generated scene labels are then input into the reinforcement learning policy network. The scene features are first compressed by the encoder, and then the optimal solution is searched in the preset compensation parameter space using a specific algorithm. The basic policy library is formed by training with historical data, and the policy library is fine-tuned online according to the real-time scene labels. Finally, a loudness compensation policy adapted to the current scene is generated.

[0131] Finally, based on this loudness compensation strategy, multi-band independent compensation and sound field control are implemented at the speaker output end. Audio parameters are adjusted for different frequency bands to ensure sound balance in each frequency band. At the same time, differentiated compensation is performed according to the characteristics of speakers in different locations in the vehicle. Instructions are executed synchronously by multiple processors to ensure that passengers in all positions in the vehicle can obtain a high-quality and suitable listening experience.

[0132] The above embodiments are only used to illustrate the technical methods 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 methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method of loudness compensation for a car audio, characterized by, The method comprises the following steps: Collecting vehicle environment data through a multi-modal sensor array; Performing mixed noise separation processing on the vehicle environment data to output independent noise source features after separation; Based on the independent noise source features and the vehicle environment data, generating a scene label through a scene cognition model; the scene label includes an environmental scene, a vehicle state, a passenger state, an audio content, a time attribute, and a user preference; Inputting the scene label into a reinforcement learning policy network to generate a loudness compensation strategy; Based on the loudness compensation strategy, implementing multi-band independent compensation and sound field control at the output end of the loudspeaker.

2. The loudness compensation method of claim 1, wherein The mixed noise separation processing on the vehicle environment data comprises the following steps: Obtaining the three-dimensional polar coordinates of the noise source through the millimeter wave radar point cloud; Calculating the wind noise feature template of the current environment based on the wind noise model established based on the fluid mechanics formula; Calculating the tire noise feature template of the current environment based on the tire noise model established based on the frequency spectrum feature formula; Calculating the air conditioner noise template of the current environment based on the air conditioner noise model coupled with the compressor speed-pressure; Obtaining the human voice noise template of the current environment based on the sound pressure-direction joint detection human voice noise model; Construct a separation matrix W = [w ij ] m×n between the microphone array and the noise source; where m is the number of microphones, n is the number of noise source types, w ij represents the weight coefficient of the i-th microphone to the j-th type of noise source; Inputting each of the feature templates into a separation matrix to maximize the signal-to-noise ratio of each frequency band through iterative solution, and outputting the independent noise source features after separation; the independent noise source features include wind noise features, tire noise features, air conditioner noise features, and human voice noise features.

3. The loudness compensation method of claim 2, wherein, The wind dry model P wind The mathematical expression of (f, t) is: ; wherein f represents sound frequency, t represents time, is a time domain variable of sound pressure signal, p represents air density, v represents vehicle speed, C d represents turbulent flow coefficient, A represents air flow disturbance projection area, refers to the effective projection area of local structure of vehicle body in the direction of air flow which produces vortex shedding, f vortex represents vortex shedding frequency, f c represents the center frequency corresponding to the maximum sound pressure level, and sigma represents frequency distribution dispersion. The tire noise model S tire The mathematical expression of (f) is: ; wherein K represents a tire coefficient, a represents a vehicle speed influence coefficient, β represents a frequency attenuation index, γ represents a road surface attenuation coefficient, a k represents a harmonic amplitude of the kth harmonic, k represents a harmonic index, f rotation represents a tire rotation frequency, and δ represents a Dirac delta function for describing a discrete frequency component. The air conditioner noise model S ac The mathematical expression of (f) is: ; wherein, η represents mechanical vibration conversion efficiency, P represents refrigerant pressure, f fund represents compressor working frequency, b m represents the mth harmonic attenuation rate; The human voice noise model V speech The mathematical expression of (t, f) is: ; wherein L0 represents a human voice trigger threshold, L p represents a sound intensity value, μ represents an activation slope, θ represents a sound source horizontal azimuth angle, N represents the number of in-vehicle microphones, and Ψ VAD represents a human voice presence probability.

4. The loudness compensation method of claim 3, wherein, The probability of the existence of human voice is obtained through a voice activity detection algorithm, which comprises the following steps: Frame processing the original audio signal to extract the energy and frequency spectrum features of each frame; According to the energy and frequency spectrum features of each frame, outputting the probability of the existence of human voice through a pre-trained SNM classifier.

5. The loudness compensation method of claim 2, wherein, The weight coefficient is optimized through a gradient boosting decision tree algorithm (GBDT), which comprises the following steps: Inputting the initialized weight matrix and the feature templates of each noise source into the GBDT for iterative operation; In each round of iteration, the signal-to-noise ratio of each frequency band is calculated as the objective function, and a Frobenius norm regularization term is introduced to prevent overfitting; A set of regression trees is constructed using the gradient boosting strategy, and the weight coefficient is gradually optimized through residual fitting to obtain the optimal weight matrix that maximizes the objective function.

6. The loudness compensation method of claim 1, wherein, The scene cognition model is used to generate a six-dimensional scene environment label, which comprises the following steps: Processing the millimeter wave radar point cloud data through a convolutional neural network to identify the road type; the road type includes but is not limited to expressway, urban road, tunnel, etc., which is used to reflect the environmental scene; Analyzing the bit bitmap of the CAN bus protocol frame to extract the engine speed and air conditioner state to obtain the vehicle state; Analyzing the infrared camera image using a YOLO network to detect the number and distribution position of passengers; Classifying the audio content based on the mel-frequency cepstral coefficient feature to obtain the audio type; the audio type includes music, voice, and prompt tone; Marking the time attribute through a Gregorian calendar and lunar calendar dual calendar system; the time attribute includes day and night period, holiday, and weekday; Modeling the user preference through a gated recurrent unit; the gated recurrent unit includes a display learning channel and an implicit learning channel, the display learning channel is used to record the manual adjustment parameter, and the implicit learning channel is used to analyze the user emotion.

7. The loudness compensation method of claim 1, wherein, The generation process of the loudness compensation strategy comprises: inputting historical scene labels into a reinforcement learning strategy network; compressing the historical scene labels into a multi-dimensional feature vector through an encoder; The double-delay deep deterministic policy gradient algorithm is adopted to search the policy in the action space to obtain an optimal compensation parameter combination; the action space includes a low-frequency gain B L ∈[-10,10]dB, a medium-frequency gain B M ∈[-5,5]dB, a high-frequency gain B H ∈[-8,8]dB, and a sound field width factor W s ∈[0.8,1.2], and the sound field width factor is used to control the spread of the three-dimensional sound field. evaluating a value function of a state-action pair using a reinforcement learning value network; the state represents the multi-dimensional feature vector, and the action represents an adjustment instruction of an optimal compensation parameter combination; performing offline iterative training according to an output result of the value function to obtain a pre-trained basic strategy library; performing online fine-tuning on the pre-trained basic strategy library according to real-time six-dimensional scene labels to obtain an adaptive compensation strategy.

8. The loudness compensation method of claim 1, wherein, The loudspeaker output end implements multi-frequency band independent compensation and sound field control, comprising: Low frequency gain B according to compensation strategy L Adjust equalizer slope in 50-200 Hz band; The intermediate frequency gain B according to the compensation strategy M Adjusting the quality factor Q value of the 200-5000 Hz band pass filter; High frequency gain B according to compensation strategy H Adjust the gain parameter of the 5000-20000 Hz band; By a sound field width factor W s Controlling the multi-channel phase difference, when W s >1, increasing the left-right channel delay difference, when W s ≤1, decreasing the left-right channel delay difference.

9. The method of loudness compensation for a car audio as claimed in claim 8, wherein, Control instructions of the multi-frequency band independent compensation and sound field control are synchronously executed by a multi-path DSP processor, and each loudspeaker unit performs differential compensation according to a spatial position; the differential compensation comprises: High frequency compensation is performed on the front loudspeaker: B H+ =B H +ΔH front ; wherein, B H+ represents the compensated high frequency gain, ΔH front represents the front gain correction amount, which is 1.2dB; Low frequency compensation is performed on the back row speakers: B L+ =B L +ΔL rear ; wherein, B L+ represents the compensated low frequency gain, ΔL rear represents the back row gain correction amount, and is 2.0 dB; Sound field width enhancement for lateral speakers: W s+ = s × k side ; where W s+ represents the compensated sound field width factor, k side represents the lateral enhancement coefficient, taken as 1.

15.

10. The loudness compensation method of claim 1, wherein, The multi-modal sensor array comprises an in-vehicle microphone array, an out-vehicle microphone array, a millimeter wave radar, an air pressure sensor and a vehicle CAN bus sensor; wherein the millimeter wave radar is deployed at four corners of the vehicle to monitor the direction of a noise source, the air pressure sensor is installed on a vehicle window guide rail to detect an opening and closing state, and the CAN bus sensor collects vehicle speed, engine speed and air conditioner compressor state data.

Images