In-vehicle sound pressure signal prediction method and device, computer equipment and storage medium

Abstract

The invention discloses an in-vehicle sound pressure signal prediction method and device, computer equipment and a storage medium, relates to the technical field of in-vehicle noise control, and is used for solving the problem that sound pressure prediction based on a fixed transfer function often generates significant deviation in related technologies. The method comprises the following steps: acquiring a target working condition parameter during vehicle driving, and determining N paths of target sound pressure time domain signals; coding the target working condition parameter to obtain a working condition vector, and performing feature extraction on the N paths of target sound pressure time domain signals to obtain N paths of acoustic feature vectors; determining attention weights corresponding to the N acoustic feature vectors based on the working condition vectors, and fusing the N acoustic feature vectors based on the attention weights to obtain a target acoustic vector; and decoding the target acoustic vector to obtain a predicted sound pressure time domain signal at the target position in the vehicle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of in-vehicle noise control, and in particular to an in-vehicle sound pressure signal prediction method and device, computer equipment and a storage medium. BACKGROUND

[0002] In the process of vehicle NVH (Noise, Vibration, and Harshness) development and optimization, it is crucial to accurately predict the sound pressure response at the ear of the occupant. In the related art, the method of transfer path analysis based on physical modeling is mainly relied on: first, identify the main noise source; then, determine the acoustic transfer function from the sound source to the ear of the occupant through experiment or simulation; finally, convolve the measured source signal with the acoustic transfer function or superimpose in the frequency domain to synthesize the target sound pressure at the ear of the occupant.

[0003] The accuracy of this method depends heavily on the pre-established acoustic transfer function. However, this function is measured under a specific static working condition and cannot adapt to real-time changes in vehicle working conditions (such as different vehicle speeds, loads, window opening and closing states, etc.). Therefore, in the actual driving process, due to the change of the acoustic transfer path, the sound pressure prediction based on the fixed transfer function often produces significant deviations. SUMMARY

[0004] Therefore, an in-vehicle sound pressure signal prediction method, device, computer equipment and storage medium are provided to solve the problem that the sound pressure prediction based on the fixed transfer function in the related art often produces significant deviations.

[0005] In a first aspect, the present application provides an in-vehicle sound pressure signal prediction method, comprising: obtaining a target working condition parameter in vehicle driving, and determining N target sound pressure time domain signals, wherein N is an integer greater than 1, and N represents the number of transfer paths of data collected by the vehicle microphone; encoding the target working condition parameter to obtain a working condition vector, and extracting features from the N target sound pressure time domain signals to obtain N acoustic feature vectors; wherein the working condition vector represents the current working condition type; determining the attention weight corresponding to the N acoustic feature vectors based on the working condition vector, and fusing the N acoustic feature vectors based on the attention weight to obtain a target acoustic vector; decoding the target acoustic vector to obtain a predicted sound pressure time domain signal at the target position in the vehicle.

[0006] Through the above method, the acoustic characteristics of the multi-channel microphone and the real-time working condition parameters are fused, and an acoustic vector adaptive to a dynamic working condition type is constructed, so that the sound pressure time domain signal at the target position in the vehicle is accurately predicted through decoding the acoustic vector, and high-precision and working condition adaptive real-time prediction of the sound pressure signal in the vehicle under various complex driving conditions is realized.

[0007] In one embodiment, the target working condition parameters in vehicle driving are acquired, and N target sound pressure time domain signals are determined, including: The initial sound pressure time domain signals collected by the N vehicle-mounted microphones are subjected to first preprocessing to obtain the N target sound pressure time domain signals, wherein the first preprocessing at least includes time alignment, first resampling and filtering processing; The acquired initial working condition parameters are subjected to second preprocessing to obtain the target working condition parameters, wherein the second preprocessing at least includes time alignment, normalization and second resampling processing with the target sound pressure time domain signals.

[0008] Through the above method, the multi-channel microphone signals are respectively subjected to time alignment, resampling and band-pass filtering processing, and the vehicle bus working condition signals are subjected to time alignment, normalization and resampling processing, ensuring the consistency and usability of the acoustic and working condition data in the time domain and scale, and providing a high-precision and strictly synchronized input data basis for subsequent multi-source signal fusion and sound pressure prediction.

[0009] In one embodiment, before the feature extraction of the N target sound pressure time domain signals, it further includes: According to the acoustic transfer function from the target position in the vehicle to the N vehicle-mounted microphones, the N target sound pressure time domain signals are pre-compensated to obtain the N target sound pressure time domain signals after compensation.

[0010] Through the above method, before feature extraction, the multi-channel microphone signals are pre-compensated based on the acoustic transfer function, which can effectively correct the signal distortion caused by the difference in the acoustic transfer path in the vehicle, improve the consistency and physical accuracy of the acoustic feature expression of the multi-channel acoustic signals at the target position, and provide a more pure and more aligned acoustic input for subsequent fusion modeling.

[0011] In one embodiment, the feature extraction of the N target sound pressure time domain signals to obtain N acoustic feature vectors includes: The N target sound pressure time domain signals are input into a feature extraction network with shared weights in parallel; Through the feature extraction network, the local time domain structure features and frequency domain spectrum features of the N target sound pressure time domain signals are extracted, and N acoustic feature vectors are output; The local time domain structure feature is used for representing impact and periodic vibration events in the target sound pressure time domain signal, and the frequency domain spectrum feature is used for representing frequency spectrum components related to engine, tire and wind noise in the target sound pressure time domain signal. The differences between the N acoustic feature vectors jointly represent the spatial difference features of the sound field perceived by different spatial positions of the vehicle-mounted microphones.

[0012] Through the above method, the feature extraction network based on shared weights uniformly processes the multiple time domain sound pressure signals, effectively retains the spatial difference information of the sound field perceived by different spatial positions of the microphones while extracting the local time domain structure and the frequency domain spectrum features, and provides acoustic feature representation with consistent expression and spatial resolution for subsequent fusion modeling.

[0013] In one embodiment, based on the working condition vector, the attention weights corresponding to the N acoustic feature vectors are determined, and the N acoustic feature vectors are fused based on the attention weights to obtain a target acoustic vector, including: The working condition vector is converted into a query vector through linear projection, and a matrix composed of the N acoustic feature vectors is converted into a key matrix and a value matrix through linear projection respectively; Based on the working condition vector, the correlation between the query vector and the N key vectors in the key matrix is calculated to obtain N attention weights matched with the current working condition represented by the working condition vector; According to the N attention weights, the N value vectors in the value matrix are weighted and fused to obtain the target acoustic vector.

[0014] Through the above method, the working condition vector is taken as a query signal, the correlation between the working condition and the multiple acoustic features is dynamically calculated through the attention mechanism, and the features are weighted and fused based on the attention weights matched with the current working condition, so that the acoustic features are dynamically and adaptively spatially fused, thereby constructing a target acoustic vector that can accurately reflect the acoustic characteristics of the current working condition type.

[0015] In one embodiment, based on the working condition vector, the correlation between the query vector and the N key vectors in the key matrix is calculated to obtain N attention weights matched with the current working condition represented by the working condition vector, including: Based on the working condition vector, the dominant noise source in the current working condition is identified, and the spatial correlation coefficients of the N acoustic feature vectors are determined according to the strength of the spatial coupling relationship between the N vehicle-mounted microphones and the dominant noise source; The strength of the spatial coupling relationship is characterized by an acoustic transmission path between the dominant noise source and N paths of the vehicle-mounted microphones, and the stronger the strength of the spatial coupling relationship, the greater the corresponding spatial correlation coefficient; On a frequency domain representation of the N paths of the acoustic feature vectors, a target frequency band region corresponding to a noise component of the dominant noise source is determined to generate a frequency domain correlation coefficient; On a time domain representation of the N paths of the acoustic feature vectors, a target time segment is determined based on a preset acoustic event window in the current working condition to generate a time domain correlation coefficient; The N paths of attention weights are fused based on the spatial correlation coefficient, the frequency domain correlation coefficient, and the time domain correlation coefficient.

[0016] Through the above method, the strength of the spatial coupling relationship between the noise source and the microphone, the noise component frequency band feature, and the acoustic event occurrence timing are analyzed based on the working condition vector, which can realize adaptive and collaborative modulation of the multi-path acoustic features in the space, frequency domain, and time domain three dimensions, thereby generating fine attention weights that are highly matched with the current working condition, and significantly improving the physical reasonableness and working condition adaptability of acoustic feature fusion.

[0017] In one embodiment, the target working condition parameters include at least one of noise source information and vehicle driving state information.

[0018] Through the above method, the noise source information and the vehicle driving state information are selected as the target working condition parameters, so that the model can directly obtain and encode dynamic working condition information that is strongly related to the in-vehicle noise generation and propagation mechanism, providing an input with clear physical meaning and control correlation for the model to understand the acoustic feature changes in different driving states, thereby significantly improving the modeling capability of real-time working condition changes and the accuracy of sound pressure prediction.

[0019] In a second aspect, the present application provides an in-vehicle sound pressure signal prediction device, the device comprising: A determination module is configured to obtain target working condition parameters in vehicle driving and determine N paths of target sound pressure time domain signals, wherein N is an integer greater than 1, and N paths represent the number of transmission paths of data collected by vehicle-mounted microphones; A feature extraction module is configured to encode the target working condition parameters to obtain a working condition vector, and extract features from N paths of the target sound pressure time domain signals to obtain N paths of acoustic feature vectors; wherein the working condition vector represents the current working condition type; A fusion module is configured to determine attention weights corresponding to N paths of the acoustic feature vectors based on the working condition vector, and fuse N paths of the acoustic feature vectors based on the attention weights to obtain a target acoustic vector; a prediction module, configured to decode the target acoustic vector to obtain a predicted time-domain sound pressure signal at a target position in the vehicle.

[0020] In a third aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the in-vehicle sound pressure signal prediction method of the first aspect when executing the computer program.

[0021] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a computer program, and the computer program is executable on a processor to implement the in-vehicle sound pressure signal prediction method of the first aspect.

[0022] The technical effects of each of the second to fourth aspects and each aspect that can be achieved are described above with reference to the technical effects that can be achieved by the first aspect or various possible schemes in the first aspect, and will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 a flowchart of an in-vehicle sound pressure signal prediction method in an embodiment; Figure 2 a flowchart of a sound pressure signal prediction model training method in an embodiment; Figure 3 a structural diagram of an in-vehicle sound pressure signal prediction device in an embodiment; Figure 4 a structural diagram of a sound pressure signal prediction model training device in an embodiment; Figure 5 an internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0024] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings. The specific operation methods in the method embodiment can also be applied to the device embodiment or the system embodiment. It should be noted that in the description of the present application, “multiple” is understood as “at least two”. “And / or” describes the association relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases: A alone, A and B exist, and B alone. A and B are connected, which means that A and B are directly connected and A and B are connected through C. In addition, in the description of the present application, “first”, “second”, etc. are used only for the purpose of distinguishing the description, and cannot be understood as indicating or implying relative importance, nor can it be understood as indicating or implying order.

[0025] Before introducing the in-vehicle sound pressure signal prediction method provided by the embodiments of the present application, the technical background of the present application is introduced in detail below.

[0026] In the related art, the sound pressure response at the ear of the occupant is mainly predicted by a transfer path analysis method based on physical modeling, which includes first identifying the main noise source; then determining the acoustic transfer function from the sound source to the ear of the occupant through experiment or simulation; finally, convolving the measured source signal with the acoustic transfer function or superimposing in the frequency domain to synthesize the target sound pressure at the ear.

[0027] The accuracy of this method depends heavily on the pre-established acoustic transfer function. However, this function is measured under a specific static operating condition and cannot adapt to real-time changing vehicle operating conditions (such as different vehicle speeds, loads, window opening and closing states, etc.). Therefore, during actual driving, due to the change of the acoustic transfer path, the sound pressure prediction based on the fixed transfer function often produces significant deviations.

[0028] In view of this, the present application provides an in-vehicle sound pressure signal prediction method, device, computer equipment and storage medium, by fusing the acoustic characteristics of multiple microphones and real-time operating condition parameters, an acoustic vector adaptive to dynamic operating condition types is constructed, so that the sound pressure time domain signal at the target position in the vehicle is accurately predicted by decoding the acoustic vector, realizing high-precision, operating condition adaptive real-time prediction of the in-vehicle sound pressure signal under various complex driving conditions.

[0029] The technical solutions provided by the present application will be described in detail below in conjunction with the drawings and specific embodiments.

[0030] Figure 1 The flowchart of the in-vehicle sound pressure signal prediction method in one embodiment includes the following steps: S101, obtaining target operating condition parameters in vehicle driving, and determining N target sound pressure time domain signals; S102, encoding the target operating condition parameters to obtain an operating condition vector, and extracting features from the N target sound pressure time domain signals to obtain N acoustic feature vectors; S103, determining the attention weights corresponding to the N acoustic feature vectors based on the operating condition vector, and fusing the N acoustic feature vectors based on the attention weights to obtain a target acoustic vector; S104, decoding the target acoustic vector to obtain the predicted sound pressure time domain signal at the target position in the vehicle.

[0031] Wherein, N is an integer greater than 1, N represents the number of transmission paths of data collected by the vehicle-mounted microphone, the operating condition vector represents the current operating condition type, and the target position in the vehicle is mainly the ear position of the occupant in the vehicle, such as the positions of the driver, the front passenger and the rear passengers.

[0032] Exemplarily, first, during the vehicle driving, two types of data are synchronously acquired: target working condition parameters are acquired through a vehicle bus (such as a controller area network bus), and meanwhile, multi-channel target sound pressure time domain signals are collected through N vehicle-mounted microphones arranged at different physical positions in the vehicle cabin.

[0033] Then, the above data is input into a trained sound pressure signal prediction model. The model divides the input data into a series of continuous short-time prediction samples in a preset time period (such as 100 milliseconds) and a preset overlap rate (such as 50%) sliding time window. Each 100-millisecond short-time prediction sample is a basic processing unit for single feature extraction and fusion calculation of the model.

[0034] The model encodes the target working condition parameters in each 100-millisecond time window through a working condition encoder to obtain a working condition vector that can comprehensively represent the current working condition type (such as high rotating speed, large load, acceleration, etc.). Meanwhile, the N-channel target sound pressure time domain signals in the same 100-millisecond time window are processed in parallel through a feature extraction network to extract the time-frequency features of each channel of signals and output N acoustic feature vectors. The N acoustic feature vectors are adaptively fused through a feature fusion module guided by the working condition vector to obtain a target acoustic vector.

[0035] Finally, the target acoustic vector corresponding to each 100-millisecond time window is decoded and reconstructed into a single-channel time domain waveform with a length of 100 milliseconds through a sound pressure signal decoder. The single-channel time domain waveforms of all continuous 100-millisecond time windows are spliced through methods such as overlap-add to reconstruct a complete prediction sound pressure time domain signal with the same length as the original input.

[0036] Through the above method, multi-channel microphone acoustic signals and vehicle working condition parameters are synchronously collected, and based on the trained sound pressure signal prediction model, end-to-end mapping from multi-dimensional input to target position sound pressure time domain signal is realized. Meanwhile, the method uses the working condition vector as a guide to dynamically fuse multiple acoustic features using an attention mechanism, which significantly improves the high-precision and adaptive prediction capability of the sound pressure signal at the passenger's ear under different driving conditions.

[0037] In one embodiment, it is exemplarily illustrated that the target working condition parameters in the vehicle driving are acquired, and the N-channel target sound pressure time domain signals are determined, including but not limited to: The initial sound pressure time domain signals collected by the N vehicle-mounted microphones are first pre-processed to obtain N-channel target sound pressure time domain signals, wherein the first pre-processing at least includes time alignment, first resampling, and filtering processing.

[0038] For example, first, the N initial sound pressure time domain signals are time-aligned to ensure that there is a strictly synchronized time starting point between the N microphone signals, and the sound field spatial information is preserved. A fixed time length of valid data segment is intercepted as an independent prediction sample to obtain the N initial sound pressure time domain signals after interception.

[0039] Then, through anti-aliasing filtering and sampling rate conversion, the sampling rate of the N initial sound pressure time domain signals after interception is uniformly resampled to the target sampling rate (such as 20 kHz, the specific sampling rate is subject to the case, and is not limited here) set by the model to ensure the consistency of the data format, and the N resampled initial sound pressure time domain signals are obtained.

[0040] Finally, through a pre-set band-pass filter (such as 5 Hz to 10 kHz, etc., the specific passband is subject to the case, and is not limited here), the N resampled initial sound pressure time domain signals are band-pass filtered to remove DC offset, power frequency interference and ultrahigh frequency noise in the signals, and the main acoustic components related to in-vehicle NVH are retained to obtain the N target sound pressure time domain signals.

[0041] Correspondingly, the obtained initial working condition parameters are second-processed to obtain target working condition parameters, wherein the second processing at least includes time alignment, normalization and second resampling processing of the target sound pressure time domain signal.

[0042] For example, first, based on the accurate time stamp synchronously collected with the target sound pressure time domain signal, the time sequence of the initial working condition parameters is time-aligned with the target sound pressure time domain signal. And according to the same fixed time length, the synchronized initial working condition parameters are intercepted to ensure that the physical time period covered by the multi-modal signals in each prediction sample is completely consistent, and the intercepted initial working condition parameters are obtained.

[0043] Then, the values of each working condition parameter in the intercepted initial working condition parameters are normalized to eliminate the dimensional difference and numerical scale difference between different parameters, and the normalized initial working condition parameters are obtained, wherein the normalization processing includes but is not limited to Z-score standardization, minimum-maximum value normalization, etc., and the specific normalization method is subject to the case, and is not limited here.

[0044] Finally, since the inherent sampling rate of the controller area network bus may be lower than the audio sampling rate, the normalized initial working condition parameters need to be resampled to the target sampling rate through an interpolation method (such as linear interpolation or cubic spline interpolation, etc.) to realize point-to-point accurate alignment on the time axis with the target sound pressure signal, and the target working condition parameters are obtained.

[0045] Through the above method, systematic multi-source data preprocessing is implemented, strict alignment and effective purification of multi-path acoustic signals and working condition parameters in time domain, frequency domain and numerical scale are realized, and consistency of input signals in time starting point, sampling rate, frequency band range and numerical scale is ensured, thereby providing high-quality and high-precision multi-modal synchronous input for subsequent model prediction.

[0046] In one embodiment, it is exemplarily illustrated that, in order to improve the prediction accuracy of the model for the acoustic environment of the target position, before feature extraction is performed on the N-path target sound pressure time domain signals, the following steps are further included: According to the acoustic transfer function from the target position in the vehicle to the N-path vehicle-mounted microphones, pre-compensation is performed on the N-path target sound pressure time domain signals to obtain N-path compensated target sound pressure time domain signals.

[0047] For example, first, under the condition that the vehicle is in a stationary state, a standard sound source (such as a loudspeaker) is placed at the target position in the vehicle, and a known wideband excitation signal is played. The excitation signal has flat spectrum characteristics, such as maximum length sequence, linear frequency modulation signal or white noise, etc. The response signals of the N vehicle-mounted microphones to the excitation signal are synchronously collected to form N groups of excitation-response signals.

[0048] Then, based on each group of excitation-response signals, a system identification algorithm is used to estimate the acoustic transfer function from the target position to each vehicle-mounted microphone, wherein the system identification algorithm includes but is not limited to least mean square algorithm, frequency domain H1 estimation method or time domain identification method based on adaptive filtering, and the obtained acoustic transfer function is a complex function, the amplitude-frequency characteristic of which represents the amplitude attenuation or enhancement in the sound propagation process, and the phase-frequency characteristic represents the propagation delay and phase distortion. And based on the acoustic transfer function corresponding to each vehicle-mounted microphone, a corresponding compensation filter is constructed.

[0049] Finally, each path target sound pressure time domain signal is converted to the frequency domain through short-time Fourier transform to obtain its frequency domain representation. The frequency domain representation is subjected to complex multiplication operation with the corresponding compensation filter to obtain the compensated frequency domain signal. The compensated frequency domain signal is subjected to inverse short-time Fourier transform to reconstruct N-path compensated target sound pressure time domain signals. The subsequent feature extraction operation takes the N-path compensated target sound pressure time domain signals as input, and these signals are theoretically closer to the sound pressure waveform directly generated or received at the target position.

[0050] Through the above method, the multi-path microphone signal is pre-compensated in the frequency domain based on the measured acoustic transfer function before feature extraction, which effectively offsets the frequency response distortion and phase delay caused by the path difference in the sound propagation process, so that each path signal is physically closer to the original acoustic state at the target position, thereby providing a spatially consistent input for the model, and significantly improving the accuracy and physical interpretability of the target position sound pressure prediction.

[0051] In one embodiment, the feature extraction of the N-channel target sound pressure time-domain signals is exemplarily illustrated to obtain N-channel acoustic feature vectors, including but not limited to: The N-channel target sound pressure time-domain signals are input into a feature extraction network sharing weights in parallel, and the network is a one-dimensional deep convolutional neural network. Through the feature extraction network, the local time-domain structure features and the frequency domain spectral features of the N-channel target sound pressure time-domain signals are extracted layer by layer, and N-channel acoustic feature vectors are output.

[0052] Among them, the local time-domain structure features are used to represent the impact and periodic vibration events in the target sound pressure time-domain signal, including but not limited to transient impact features: identifying non-periodic impact events caused by road unevenness, structure impact, etc., which are represented as short-time high-energy pulses in the signal; periodic vibration features: capturing periodic pulse sequences generated by engine periodic combustion, rotating mechanical motion, etc., which are characterized by repeated waveform patterns with stable or slowly varying periods in the time domain.

[0053] The frequency domain spectral features are used to represent the frequency spectrum components in the target sound pressure time-domain signal related to the engine, tire and wind noise, including but not limited to engine order features: narrowband harmonic components corresponding to integer multiples of engine speed frequencies, used to represent the order structure and energy distribution of engine noise; tire noise features: wideband noise corresponding to the interaction between the tire and the road, usually containing specific passband frequencies and their sideband modulation information; wind noise band features: high-frequency wideband turbulent noise corresponding to the interaction between airflow and vehicle structure.

[0054] In addition, although the weights of the feature extraction network are shared among the N-channel target sound pressure time-domain signals, due to the fact that each channel signal itself comes from a microphone at a different spatial location in the vehicle, the sound field information (such as amplitude, phase, frequency response, etc.) perceived by each microphone has physical differences. When the network processes these spatially distributed signals, it will generate a unique set of feature activation patterns for each channel. Therefore, the differences between the N-channel acoustic feature vectors output ultimately represent the sound field spatial difference features perceived by different spatial locations of the vehicle-mounted microphones, such as the directivity of the sound source, the propagation attenuation, and the spatial distribution characteristics of the vehicle cavity modes.

[0055] Through the above method, the one-dimensional convolution network based on weight sharing extracts the time-domain transient and periodic features, frequency-domain order and wideband noise components of multiple sound pressure signals in parallel, and naturally retains the sound field difference information determined by the spatial position of each microphone in the unified feature representation, thereby realizing comprehensive and compact feature representation of the vehicle acoustic environment in time, frequency and space dimensions.

[0056] In one embodiment, it is exemplarily illustrated that based on the working condition vector, the attention weight corresponding to the N acoustic feature vectors is determined, and based on the attention weight, the N acoustic feature vectors are fused to obtain the target acoustic vector, including but not limited to: In the feature fusion module based on the attention mechanism, the working condition vector is converted into a query vector through linear projection, and the matrix composed of the N acoustic feature vectors is converted into a key matrix and a value matrix through linear projection respectively.

[0057] For example, the working condition vector is multiplied by a first weight matrix to obtain the query vector; the matrix composed of the N acoustic feature vectors is multiplied by a second weight matrix to obtain the key matrix; and the matrix composed of the N acoustic feature vectors is multiplied by a third weight matrix to obtain the value matrix. The first, second and third weight matrices are all parameters that can be learned by the model during the training process, which map the original features to different semantic subspaces to realize the functions of query representation, feature matching and information aggregation respectively.

[0058] Then, based on the working condition vector, the correlation between the query vector and the N key vectors in the key matrix is calculated to obtain the N attention weights matched with the current working condition represented by the working condition vector.

[0059] For example, through the scaled dot-product attention mechanism, the correlation coefficient between the query vector and each key vector in the key matrix is calculated. And the obtained correlation coefficient sequence is normalized by Softmax to obtain the N attention weights matched with the current working condition.

[0060] The size of the attention weight represents the correlation intensity between the acoustic feature of the i-th microphone and the current working condition. For example, in the high-speed working condition, the feature vector of the microphone close to the engine has high correlation with the working condition vector, and will obtain a larger weight; in the high-speed cruising working condition, the feature vector of the microphone close to the A-pillar or the rearview mirror may obtain higher attention.

[0061] Finally, according to the N attention weights, the N value vectors in the value matrix are weighted and fused to obtain the target acoustic vector. The target acoustic vector z can be calculated by the following formula: wherein, v i represents the i-th value vector in the value matrix, and a i represents the attention weight corresponding to the i-th value vector.

[0062] Through the above method, the feature fusion module based on the attention mechanism adaptively calculates and weightedly fuses the multiple acoustic features based on the working condition vector as the calculation signal, so that the model can highlight the acoustic information most related to the current driving state according to the real-time working condition, thereby constructing a target acoustic vector that is tightly coupled to the current working condition and contains key spatial sound field information.

[0063] In one embodiment, it is exemplarily illustrated that, based on the working condition vector, the correlation between the query vector and the N-way key vector in the key matrix is calculated to obtain the N-way attention weight matched with the current working condition represented by the working condition vector, including but not limited to: Based on the working condition vector, the dominant noise source (such as engine, tire or wind noise, etc.) in the current working condition is identified. And according to the strength of the spatial coupling relationship between the N-way on-board microphone and the dominant noise source, the spatial correlation coefficient of the N-way acoustic feature vector is determined; wherein the strength of the spatial coupling relationship is jointly represented by the distance, directionality, and shielding of the acoustic transmission path between the dominant noise source and the N-way on-board microphone, and the stronger the strength of the spatial coupling relationship, the greater the corresponding spatial correlation coefficient.

[0064] For example, the Euclidean distance between the microphone and the dominant noise source is calculated, the closer the distance, the stronger the strength of the spatial coupling relationship; the relative azimuth angle between the microphone and the dominant noise source is calculated, the more the azimuth angle is towards the radiation direction of the dominant noise source, the stronger the strength of the spatial coupling relationship; it is judged whether there is structural shielding in the acoustic transmission path, the less the shielding, the stronger the strength of the spatial coupling relationship. The spatial correlation coefficient can be calculated according to the strength of the above factors by weighted combination or pre-set mapping function.

[0065] In the frequency domain representation (obtained by short-time Fourier transform) of the N-way acoustic feature vector, the target frequency band region corresponding to the noise component of the dominant noise source is determined to generate the frequency domain correlation coefficient.

[0066] For example, when the working condition vector identifies the dominant noise source as a high-speed engine, the target frequency band region will focus on the main order frequency of the engine (such as 25 Hz, 100 Hz, 200 Hz, etc. for a four-cylinder engine at 3000 rpm).

[0067] The first ratio between the frequency energy of the target frequency band region (in a pre-set frequency band window centered on the main order frequency) in the frequency domain representation of each acoustic feature vector and the total energy of the full frequency band is calculated, and used as the frequency domain correlation coefficient of the acoustic feature vector. The larger the first ratio, the more prominent the engine order component related to the current working condition in the acoustic feature vector.

[0068] Taking the pre-set acoustic event window (such as the engine combustion pulse peak period) in the current working condition as the reference, the target time segment is determined on the time domain representation of the N-way acoustic feature vector to generate the time domain correlation coefficient.

[0069] For example, a second ratio between the time domain energy of the target time segment and the total time domain energy of the entire analysis period (such as the current 100 ms data frame) is calculated on the time domain representation of each acoustic feature vector, and is taken as the time domain correlation coefficient of the acoustic feature vector.

[0070] In this application, the time domain energy of a signal refers to the sum of the squares of its sample values within a specified time period; the frequency domain energy of a signal refers to the sum of the squares of the amplitudes of its spectral coefficients within a specified frequency band range.

[0071] Therefore, based on the spatial correlation coefficient, the frequency domain correlation coefficient and the time domain correlation coefficient, N attention weights are fused to generate. The fusion generation method can be weighted product or weighted sum, and the specific method is determined according to the situation, which is not limited here.

[0072] Through the above method, spatial, frequency domain and time domain three-dimensional physical modulation are introduced into the attention mechanism, realizing fine modulation of multiple acoustic features and the current working condition. The spatial correlation coefficient captures the strength of the spatial coupling relationship between the microphone and the dominant noise source, the frequency domain correlation coefficient focuses on the frequency band distribution of the dominant noise source, and the time domain correlation coefficient strengthens the timing information of the acoustic event. Finally, attention weights with clear physical interpretability and high matching to real-time working conditions are generated, thereby significantly improving the pertinence, robustness and prediction accuracy of feature fusion.

[0073] In one embodiment, it is exemplarily illustrated that the target working condition parameter includes at least one of noise source information (such as the speed and torque of the engine or driving motor), vehicle driving state information (such as vehicle speed, throttle opening, brake pedal opening, gear, etc.).

[0074] Through the above method, signals such as engine / motor speed and torque directly representing the excitation state of the noise source are selected, and combined with real-time driving state parameters such as vehicle speed and throttle opening, a multi-dimensional working condition representation system is constructed, providing a working condition input with clear physical correlation for the model, thereby enhancing the modeling ability of the model for the dynamic coupling relationship between the excitation characteristics of the noise source and the vehicle running state, and improving the adaptability and accuracy of sound pressure prediction to working condition changes.

[0075] Figure 2 A flowchart of the sound pressure signal prediction model training method in one embodiment is shown in the figure, which includes the following steps: S201, obtaining multi-modal synchronous data of a vehicle; the multi-modal synchronous data at least includes N sample time domain sound pressure signals, sample working condition parameters, and a real sound pressure time domain signal at a target position in the vehicle; S202, training the sound pressure signal prediction model by taking the N-channel sample time-domain sound pressure signals and the sample working condition parameters as the model input and taking the real sound pressure time-domain signal as the supervised output.

[0076] wherein N represents the number of transmission paths of the data collected by the vehicle-mounted microphones, and N is an integer greater than 1.

[0077] Exemplarily, on a test vehicle, N-channel vehicle-mounted microphones (located at positions such as the instrument panel, the roof, the door, the rear shelf, etc.) are arranged, and an artificial head or a high-precision probe microphone is placed at a target position (such as the right ear of the driver) as a reference microphone. Then, N-channel sample sound pressure time-domain signals are obtained through the N-channel vehicle-mounted microphones; sample working condition parameters are obtained through the vehicle controller local area network bus; and a real sound pressure time-domain signal at the target position in the vehicle is obtained through the reference microphone. All the collected data are preprocessed, and the specific preprocessing methods are described above and will not be repeated here.

[0078] The collection process covers multiple driving conditions from idle speed to high speed, different throttle openings (such as slow acceleration and rapid acceleration), different road surfaces (such as smooth asphalt roads, rough cement roads, and deceleration strips), etc., to form a sample data set with strictly aligned time within a preset time period (such as several hours, the specific time period is determined according to the situation, which is not limited here).

[0079] A sound pressure signal prediction model is constructed by using a mathematical fitting method (such as a multivariate linear regression equation, a neural network, etc.), which includes the following in sequence: a feature extraction network: a one-dimensional convolutional neural network is used to process the N-channel sample time-domain sound pressure signals in parallel, to extract local time-domain structural features, frequency domain spectral features, and spatial difference features of the sound field between multiple microphones, and to output N-channel sample acoustic feature vectors.

[0080] a working condition semantic encoder: a hybrid network including a one-dimensional convolutional layer and a recurrent neural network layer is used to encode the working condition parameters, to extract high-level semantic information representing the working condition types, and to output a sample working condition vector.

[0081] a feature fusion module: a cross-attention mechanism is used to convert the sample working condition vector into a sample query vector through linear projection, and to convert a matrix composed of N-channel sample acoustic feature vectors into a sample key matrix and a sample value matrix through linear projection; by calculating the correlation coefficients of the sample query vector and each sample key vector in the sample key matrix, N-channel sample attention weights are generated, and corresponding sample value vectors in the sample value matrix are weighted and fused according to the N-channel sample attention weights to obtain a sample fusion feature; the sample attention weights can: Identify the dominant noise source of the sample, and enhance the contribution of the microphone signals from the nearby spatial position; emphasize the key frequency band components related to the sample working condition; focus on the key time segment related to the sample working condition.

[0082] Sound pressure signal decoder: using a network containing a one-dimensional transpose convolution layer (for up-sampling) and a fully connected layer, the fusion feature output by the feature fusion module is decoded and reconstructed into a single-channel time-domain sound pressure waveform, i.e. the initial predicted sound pressure signal of the target position in the vehicle.

[0083] The prediction error between the initial predicted sound pressure signal and the real sound pressure time-domain signal is used to construct a composite loss function. The loss function can include time-domain error (such as root mean square error of prediction error) and frequency domain error (such as amplitude spectrum loss). And using the back propagation algorithm, based on the sample data set, the sound pressure signal prediction model is iteratively trained (the model processes the sample data in the same way as described above, which is not repeated here), until the initial predicted sound pressure signal meets the preset convergence condition, and the current model is used as the trained sound pressure signal prediction model.

[0084] Wherein, the preset convergence condition can be that the root mean square error of the prediction error is less than a first preset threshold (such as 2 decibels, the specific threshold is determined according to the situation, which is not limited here) and the determination coefficient between the initial predicted sound pressure signal and the real sound pressure time-domain signal is greater than a second preset threshold (such as 0.95, the specific threshold is determined according to the situation, which is not limited here), indicating that the model has very high prediction accuracy and goodness of fit.

[0085] The trained sound pressure signal prediction model is solidified and lightened, and integrated into a vehicle hardware-in-the-loop simulation system or a cloud NVH analysis platform. In real-time or offline scenarios, input real-time multi-channel vehicle microphone signals and controller area network bus signals, and the model can output high-precision predicted sound pressure time-domain signals of the target position online, which can be used for: Provide feedforward reference signals for active noise control systems; Real-time evaluation and optimization of the acoustic quality of the cockpit in the virtual prototype stage; Realize the acoustic contribution analysis and fault diagnosis based on the working condition.

[0086] The embodiment provides a training method of an end-to-end vehicle sound pressure signal prediction model, which constructs a deep neural network capable of fusing multi-channel acoustic features and working condition semantic information through end-to-end supervised training of multi-modal synchronous data. The model realizes dynamic adaptation of feature fusion through attention mechanism, and can learn complex sound field mapping relationship under different driving conditions in a unified framework. The final model can accurately and in real time predict the sound pressure signal of the target position in the vehicle under different working conditions, and provide high-reliability input for active noise control, NVH evaluation and diagnosis, etc.

[0087] It should be understood that, althoughFigure 1 and 2 The steps in the flowcharts of Figure 1 and 2 may be displayed in sequence according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover,

[0088] In one embodiment, as shown in Figure 3 , an in-vehicle sound pressure signal prediction device is provided, comprising: a determination module 301, a feature extraction module 302, a fusion module 303, and a prediction module 304, wherein: The determination module 301 is configured to obtain a target working condition parameter in vehicle driving, and determine N target sound pressure time domain signals, wherein N is an integer greater than 1, and N represents the number of transmission paths of data collected by a vehicle microphone; The feature extraction module 302 is configured to encode the target working condition parameter to obtain a working condition vector, and extract features from the N target sound pressure time domain signals to obtain N acoustic feature vectors; wherein the working condition vector represents the current working condition type; The fusion module 303 is configured to determine attention weights corresponding to the N acoustic feature vectors based on the working condition vector, and fuse the N acoustic feature vectors based on the attention weights to obtain a target acoustic vector; The prediction module 304 is configured to decode the target acoustic vector to obtain a predicted sound pressure time domain signal at a target position in the vehicle.

[0089] For specific limitations of the in-vehicle sound pressure signal prediction device, see the limitations of the in-vehicle sound pressure signal prediction method in the above, which will not be repeated here. Each module in the above in-vehicle sound pressure signal prediction device can be realized by software, hardware and their combinations. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor calls and executes the operations corresponding to the above modules.

[0090] In one embodiment, as shown in Figure 4 , a sound pressure signal prediction model training device is provided, comprising: an acquisition module 401 and a training module 402, wherein: The acquisition module 401 is configured to acquire multi-modal synchronous data of the vehicle, and the multi-modal synchronous data at least includes N-channel sample time-domain sound pressure signals, sample working condition parameters, and a real time-domain sound pressure signal of a target position in the vehicle. The training module 402 is configured to take the N-channel sample time-domain sound pressure signals and the sample working condition parameters as model inputs, take the real time-domain sound pressure signal as a supervised output, train the sound pressure signal prediction model, and obtain a trained sound pressure signal prediction model.

[0091] For specific limitations of the in-vehicle sound pressure signal prediction device, refer to the limitations of the in-vehicle sound pressure signal prediction method in the foregoing, which will not be repeated here. Each module in the above in-vehicle sound pressure signal prediction device can be realized by software, hardware, and combinations thereof, in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor calls and executes the operations corresponding to each of the above modules.

[0092] In one embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 5 The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is configured to store in-vehicle sound pressure signal prediction data and sound pressure signal prediction model training data. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement an in-vehicle sound pressure signal prediction method and a sound pressure signal prediction model training method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, trackball, or touchpad arranged on the shell of the computer device, or can be an external keyboard, touchpad, or mouse, etc.

[0093] Those skilled in the art can understand that Figure 5 the structure shown in the above

[0094] In one embodiment, a computer device is provided, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the following steps when executing the computer program: obtaining a target working condition parameter in vehicle driving, and determining N-path target sound pressure time domain signals, wherein N is an integer greater than 1, and N paths represent the number of transmission paths of data collected by vehicle-mounted microphones; encoding the target working condition parameter to obtain a working condition vector, and extracting features of the N-path target sound pressure time domain signals to obtain N-path acoustic feature vectors; wherein the working condition vector represents the current working condition type; determining attention weights corresponding to the N-path acoustic feature vectors based on the working condition vector, and fusing the N-path acoustic feature vectors based on the attention weights to obtain a target acoustic vector; decoding the target acoustic vector to obtain a predicted sound pressure time domain signal at a target position in the vehicle.

[0095] In one embodiment, the processor further implements the following steps when executing the computer program: performing first preprocessing on N-path initial sound pressure time domain signals collected by vehicle-mounted microphones to obtain N-path target sound pressure time domain signals, wherein the first preprocessing at least includes time alignment, first resampling, and filtering processing; performing second preprocessing on the obtained initial working condition parameter to obtain the target working condition parameter, wherein the second preprocessing at least includes time alignment, normalization, and second resampling processing with respect to the target sound pressure time domain signals.

[0096] In one embodiment, the processor further implements the following steps when executing the computer program: performing pre-compensation on the N-path target sound pressure time domain signals according to acoustic transfer functions from the target position in the vehicle to the N-path vehicle-mounted microphones to obtain N-path compensated target sound pressure time domain signals.

[0097] In one embodiment, the processor further implements the following steps when executing the computer program: parallel inputting the N-path target sound pressure time domain signals into a feature extraction network with shared weights; extracting local time domain structure features and frequency domain spectral features of the N-path target sound pressure time domain signals through the feature extraction network to output N-path acoustic feature vectors; wherein the local time domain structure features are used to represent impact and periodic vibration events in the target sound pressure time domain signals; and the frequency domain spectral features are used to represent frequency spectrum components related to engine, tire, and wind noise in the target sound pressure time domain signals; the differences between the N-path acoustic feature vectors collectively represent spatial difference features of the sound field perceived by different vehicle-mounted microphones.

[0098] In one embodiment, the processor, when executing the computer program, also implements the following steps: The working condition vector is converted into a query vector through linear projection, and a matrix composed of N acoustic feature vectors is converted into a key matrix and a value matrix through linear projection respectively; Based on the working condition vector, the correlation between the query vector and the N key vectors in the key matrix is calculated to obtain N attention weights matched with the current working condition represented by the working condition vector; According to the N attention weights, the corresponding N value vectors in the value matrix are weighted and fused to obtain a target acoustic vector.

[0099] In one embodiment, the processor, when executing the computer program, also implements the following steps: Based on the working condition vector, the dominant noise source in the current working condition is identified, and the spatial correlation coefficient of the N acoustic feature vectors is determined according to the strength of the spatial coupling relationship between the N vehicle-mounted microphones and the dominant noise source; The strength of the spatial coupling relationship is represented by the acoustic transmission path between the dominant noise source and the N vehicle-mounted microphones, and the stronger the strength of the spatial coupling relationship, the greater the corresponding spatial correlation coefficient; On the frequency domain representation of the N acoustic feature vectors, a target frequency band region corresponding to the noise component of the dominant noise source is determined to generate a frequency domain correlation coefficient; Taking a preset acoustic event window in the current working condition as a reference, a target time segment is determined on the time domain representation of the N acoustic feature vectors to generate a time domain correlation coefficient; Based on the spatial correlation coefficient, the frequency domain correlation coefficient and the time domain correlation coefficient, the N attention weights are fused and generated.

[0100] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the following steps: Obtain target working condition parameters in vehicle driving, and determine N target sound pressure time domain signals, wherein N is an integer greater than 1, and N represents the number of transmission paths of data collected by vehicle-mounted microphones; Encode the target working condition parameters to obtain a working condition vector, and extract features from the N target sound pressure time domain signals to obtain N acoustic feature vectors; wherein the working condition vector represents the current working condition type; Based on the working condition vector, determine the attention weights corresponding to the N acoustic feature vectors, and based on the attention weights, fuse the N acoustic feature vectors to obtain a target acoustic vector; Decode the target acoustic vector to obtain a predicted sound pressure time domain signal at a target position in the vehicle.

[0101] In one embodiment, the computer program, when executed by the processor, further implements the following steps: The initial sound pressure time domain signals collected by the N-channel vehicle-mounted microphones are pre-processed to obtain N-channel target sound pressure time domain signals, wherein the first pre-processing at least includes time alignment, first resampling and filtering processing; The obtained initial working condition parameters are secondly pre-processed to obtain target working condition parameters, wherein the second pre-processing at least includes time alignment, normalization and second resampling processing with the target sound pressure time domain signals.

[0102] In one embodiment, the computer program, when executed by the processor, further implements the following steps: According to the acoustic transfer function from the target position in the vehicle to the N-channel vehicle-mounted microphones, the N-channel target sound pressure time domain signals are pre-compensated to obtain N-channel compensated target sound pressure time domain signals.

[0103] In one embodiment, the computer program, when executed by the processor, further implements the following steps: The N-channel target sound pressure time domain signals are input into the feature extraction network with shared weights in parallel; Through the feature extraction network, the local time domain structure features and the frequency domain spectral features of the N-channel target sound pressure time domain signals are extracted, and N-channel acoustic feature vectors are output; Wherein, the local time domain structure features are used to represent the impact and periodic vibration events in the target sound pressure time domain signal; the frequency domain spectral features are used to represent the frequency spectrum components related to engine, tire and wind noise in the target sound pressure time domain signal; The differences between the N-channel acoustic feature vectors collectively represent the spatial difference features of the sound field perceived by different vehicle-mounted microphones.

[0104] In one embodiment, the computer program, when executed by the processor, further implements the following steps: The working condition vector is converted into a query vector through linear projection, and the matrix composed of the N-channel acoustic feature vectors is converted into a key matrix and a value matrix through linear projection respectively; Based on the working condition vector, the correlation between the query vector and the N-channel key vectors in the key matrix is calculated to obtain N-channel attention weights matched with the current working condition represented by the working condition vector; According to the N-channel attention weights, the corresponding N-channel value vectors in the value matrix are weighted and fused to obtain a target acoustic vector.

[0105] In one embodiment, the computer program, when executed by the processor, further implements the following steps: Based on the working condition vector, the dominant noise source under the current working condition is identified, and the spatial correlation coefficients of the N-channel acoustic feature vectors are determined according to the strength of the spatial coupling relationship between the N-channel vehicle-mounted microphones and the dominant noise source. The strength of the spatial coupling relationship is represented by an acoustic transmission path between the dominant noise source and the N-channel vehicle microphone, and the stronger the strength of the spatial coupling relationship, the greater the corresponding spatial correlation coefficient; On the frequency domain representation of the N-channel acoustic feature vector, a target frequency band region corresponding to the noise component of the dominant noise source is determined to generate a frequency domain correlation coefficient; On the time domain representation of the N-channel acoustic feature vector, a target time segment is determined based on the preset acoustic event window under the current working condition to generate a time domain correlation coefficient; Based on the spatial correlation coefficient, the frequency domain correlation coefficient and the time domain correlation coefficient, the N-channel attention weight is fused and generated.

[0106] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0107] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present disclosure.

[0108] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific and detailed manner, but should not be construed as limiting the scope of the patent. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these are all within the scope of the present application. Therefore, the scope of protection of the patent of the present application should be subject to the appended claims.

Claims

1. A method of predicting an in-vehicle sound pressure signal, characterized by, The method comprises: acquiring a target working condition parameter in vehicle driving, and determining N target sound pressure time domain signals, wherein N is an integer greater than 1, and N represents the number of transmission paths of data collected by vehicle-mounted microphones; encoding the target working condition parameter to obtain a working condition vector, and extracting features of the N target sound pressure time domain signals to obtain N acoustic feature vectors; wherein the working condition vector represents a current working condition type; based on the working condition vector, determining attention weights corresponding to the N acoustic feature vectors, and based on the attention weights, fusing the N acoustic feature vectors to obtain a target acoustic vector; decoding the target acoustic vector to obtain a predicted sound pressure time domain signal at a target position in the vehicle.

2. The method of claim 1, wherein, The acquisition of the target working condition parameter in vehicle driving and the determination of the N target sound pressure time domain signals comprise: performing first preprocessing on N initial sound pressure time domain signals collected by the vehicle-mounted microphones to obtain the N target sound pressure time domain signals, wherein the first preprocessing at least includes time alignment, first resampling and filtering processing; performing second preprocessing on the acquired initial working condition parameter to obtain the target working condition parameter, wherein the second preprocessing at least includes time alignment with the target sound pressure time domain signal, normalization and second resampling processing.

3. The method of claim 1, wherein, Before the feature extraction of the N target sound pressure time domain signals, it further comprises: According to the acoustic transfer function from the target position in the vehicle to the N vehicle-mounted microphones, pre-compensate the N target sound pressure time domain signals to obtain N target sound pressure time domain signals after compensation.

4. The method of claim 1, wherein, The feature extraction of the N target sound pressure time domain signals comprises: parallel inputting the N target sound pressure time domain signals into a feature extraction network with shared weights; extracting local time domain structure features and frequency domain spectrum features of the N target sound pressure time domain signals through the feature extraction network, and outputting N acoustic feature vectors; wherein the local time domain structure features are used to represent impact and periodic vibration events in the target sound pressure time domain signal; and the frequency domain spectrum features are used to represent frequency spectrum components related to engine, tire and wind noise in the target sound pressure time domain signal; The differences between the N acoustic feature vectors represent the spatial difference features of the sound field perceived by different vehicle-mounted microphone spatial positions.

5. The method of claim 1, wherein, Based on the working condition vector, the determination of the attention weights corresponding to the N acoustic feature vectors, and the fusion of the N acoustic feature vectors based on the attention weights to obtain a target acoustic vector, comprises: linearly projecting the working condition vector into a query vector, and linearly projecting a matrix composed of the N acoustic feature vectors into a key matrix and a value matrix, respectively; based on the working condition vector, calculating the correlation between the query vector and N key vectors in the key matrix to obtain N attention weights matched with the current working condition represented by the working condition vector; and According to the N-way attention weight, corresponding N-way value vectors in the value matrix are weighted and fused to obtain the target acoustic vector.

6. The method of claim 5, wherein, The correlation between the query vector and N-way key vectors in the key matrix is calculated based on the working condition vector to obtain N-way attention weights matched with the current working condition represented by the working condition vector, including: Based on the working condition vector, a dominant noise source in the current working condition is identified, and spatial correlation coefficients of the N-way acoustic feature vectors are determined according to the strength of the spatial coupling relationship between the N-way vehicle-mounted microphones and the dominant noise source. The strength of the spatial coupling relationship is represented by the acoustic transmission path between the dominant noise source and the N-way vehicle-mounted microphones, and the stronger the strength of the spatial coupling relationship, the larger the corresponding spatial correlation coefficient. On the frequency domain representation of the N-way acoustic feature vectors, a target frequency band region corresponding to the noise component of the dominant noise source is determined to generate a frequency domain correlation coefficient. On the time domain representation of the N-way acoustic feature vectors, a target time segment is determined based on a preset acoustic event window in the current working condition to generate a time domain correlation coefficient. Based on the spatial correlation coefficient, the frequency domain correlation coefficient and the time domain correlation coefficient, the N-way attention weight is fused and generated.

7. The method of claim 1, wherein, The target working condition parameter includes at least one of noise source information and vehicle driving state information.

8. An in-vehicle sound pressure signal prediction device characterized by comprising: The device includes: A determination module is configured to obtain a target working condition parameter in vehicle driving and determine N-way target sound pressure time domain signals, where N is an integer greater than 1, and N-way represents the number of transmission paths of data collected by vehicle-mounted microphones; A feature extraction module is configured to encode the target working condition parameter to obtain a working condition vector, and extract features from the N-way target sound pressure time domain signals to obtain N-way acoustic feature vectors; wherein the working condition vector represents a current working condition type; A fusion module is configured to fuse the N-way acoustic feature vectors based on the working condition vector to obtain a target acoustic vector; A prediction module is configured to decode the target acoustic vector to obtain a predicted sound pressure time domain signal at a target position in the vehicle.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-7.

Images