Automobile sound effect intelligent evaluation system and method based on deep learning

Abstract

The invention discloses an automobile sound effect intelligent evaluation system and method based on deep learning, and relates to the field of automobile engineering and artificial intelligence, and the system comprises an audio data collection module, a psychoacoustic feature extraction module, a deep learning evaluation module and an optimization guidance module. The system obtains an original audio signal of a vehicle, extracts psychoacoustic parameters such as loudness and sharpness, and constructs a time sequence feature matrix; performing convolution operation and nonlinear mapping on the feature matrix by using a pre-trained one-dimensional convolutional neural network model, and outputting predicted values of subjective evaluation indexes such as fluency, comfort and luxury; and key parameters influencing the sound quality are identified by calculating the gradient of the predicted value relative to the input matrix. According to the method, automatic objective evaluation of the automobile sound effect is realized, a mapping relation between physical parameters and subjective hearing is established, optimization guidance based on attribution analysis can be provided, and the adaptability of an evaluation model to a dynamic driving scene is ensured.

Description

A Deep Learning-Based Intelligent Evaluation System and Method for Car Audio Effects Technical Field

[0001] This application relates to the fields of automotive engineering and artificial intelligence, and in particular to a deep learning-based intelligent evaluation system and method for automotive sound effects. Background Technology

[0002] With the rapid development of the automotive industry and the increasing demands of consumers for driving quality, in-vehicle sound quality has become an important indicator for measuring vehicle comfort, luxury, and brand characteristics. Unlike traditional noise, vibration, and harshness (NVH) control, which only focuses on reducing sound pressure levels or eliminating noise, modern sound quality evaluation places more emphasis on studying the subjective perception of sound by the human ear from a psychoacoustic perspective, aiming to create a pleasant acoustic environment or one that matches a specific driving atmosphere.

[0003] Currently, the evaluation of automotive sound quality mainly relies on subjective human assessment, which involves organizing professionally trained evaluators or a panel of ordinary users to score collected samples under specific listening conditions. While this method directly reflects subjective auditory perception, it has significant limitations. First, the subjective evaluation process is extremely time-consuming and labor-intensive, and organizing large-scale reviews is costly, severely hindering the efficiency of vehicle development and component selection. Second, the evaluation results are highly susceptible to the influence of subjective factors such as the evaluator's physiological state, psychological state, personal preferences, and auditory fatigue, resulting in poor consistency, stability, and reproducibility of the evaluation data, making it difficult to establish unified quantitative standards.

[0004] To overcome the drawbacks of subjective human evaluation, the industry has attempted to adopt objective evaluation methods based on physical acoustic parameters, such as measuring A-weighted sound pressure level and one-third octave bands. However, there is no linear correspondence between simple physical energy indicators and complex human auditory perception; a simple reduction in sound pressure level does not equate to an improvement in listening experience. For example, some low-sound-pressure-level high-frequency howling or modulation noise at specific frequencies is often more irritating than broadband background noise. Although some existing technologies incorporate psychoacoustic parameters (such as loudness and sharpness) for auxiliary analysis, they often use static averages or simple linear regression models, neglecting the non-stationary characteristics of sound signals evolving over time under dynamic conditions such as acceleration and braking, as well as the nonlinear coupling relationships between multiple parameters. This results in a significant discrepancy between objective predictions and actual subjective perception.

[0005] Furthermore, with the development of artificial intelligence technology, establishing sound quality prediction models using neural networks has become a research hotspot. However, most existing deep learning-based evaluation methods treat the model as a black box. Although they can output an evaluation score based on the input signal, they lack model interpretability. When the model determines that the sound effect of a certain car model is unqualified or the comfort level is low, the system cannot tell the engineer which specific acoustic feature (such as roughness or fluctuation) became abnormal at what time, nor can it pinpoint the physical cause of the score decline. This black-box model, lacking attribution analysis capabilities, makes it difficult for acoustic engineers to make targeted rectifications and optimizations based on the evaluation results, limiting the guiding value and application breadth of intelligent evaluation systems in actual engineering design. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based intelligent evaluation system and method for automotive sound effects, which at least solves one of the technical problems of traditional automotive sound effect evaluation relying on subjective human listening, resulting in high cost, low efficiency and poor consistency, and existing objective evaluation methods having difficulty in establishing an accurate nonlinear mapping between physical acoustic parameters and subjective listening experience and lacking guidance for attribution analysis of sound quality problems.

[0007] This invention provides the following solution:

[0008] The first aspect of this invention provides a deep learning-based intelligent evaluation system for automotive sound effects.

[0009] This system is based on a data-driven supervised learning strategy, establishing a nonlinear mapping mechanism from objective acoustic physical quantities to subjective sound quality perception. The system's operation begins with the digital capture of the physical sound field, acquiring the raw audio signal of the vehicle under test under specific operating conditions (such as full-speed acceleration) through an audio data acquisition module. Unlike traditional methods that directly process sound pressure level or spectral data, this system introduces a psychoacoustic feature extraction module. This module, based on the human auditory perception model, performs time-frequency domain transformation and feature calculation on the raw audio signal.

[0010] Specifically, the system constructs a two-dimensional temporal feature matrix containing both time and feature dimensions as its core data structure. In this matrix, the time dimension preserves the dynamic trend of sound changes with operating conditions (such as increasing engine speed), while the feature dimension characterizes the auditory attributes of the sound through a specific set of psychoacoustic parameters. Preferably, the feature dimension consists of five types of parameters: loudness, sharpness, roughness, fluctuation, and annoyance. This feature combination method transforms the physical signal into an intermediate feature representation that more closely resembles the human brain's perception level, providing high signal-to-noise ratio input data for subsequent deep learning models.

[0011] During the feature processing and evaluation inference phase, the deep learning evaluation module loads and runs a pre-trained deep neural network model. This model is configured as an end-to-end mapper, receiving the aforementioned temporal feature matrix as input and directly generating quantified subjective evaluation indicators of sound quality (such as smoothness, comfort, and luxury) as output.

[0012] To effectively capture the evolution of sound signals over time (e.g., the fluctuations in sound waves during acceleration), this invention preferably employs a one-dimensional convolutional neural network (1DCNN) architecture. Its working principle is as follows: A one-dimensional convolutional kernel performs sliding operations on the time axis of the temporal feature matrix. The first convolutional layer extracts the local correlation features of each psychoacoustic parameter within adjacent time steps; subsequently, pooling layers reduce the data dimensionality while preserving salient features; a deeper second convolutional layer further extracts complex collaborative change patterns between parameters; finally, a global pooling layer compresses the variable-length temporal features into a fixed-length feature vector, and a fully connected layer performs linear transformation to complete regression prediction. This architecture design enables the model to understand the dynamic texture of sound, rather than merely processing static acoustic indicators.

[0013] Furthermore, to ensure the interpretability and engineering guidance value of the evaluation results, the preferred implementation includes an optimization guidance module. This module, based on the gradient backpropagation principle, calculates the gradient distribution of the output index relative to the input matrix. The system quantifies the contribution weight of each psychoacoustic parameter to the final score by statistically analyzing the gradient magnitude. When the score falls below a preset standard, the system automatically identifies the key physical parameters causing the score decrease (e.g., excessive roughness at a certain moment) according to their weights, thereby transforming the black-box model prediction into actionable acoustic optimization directions.

[0014] At the model training level, the system employs a supervised learning framework, using manually labeled real data to constrain network parameter updates through a mean squared error loss function. Simultaneously, a data augmentation mechanism is preferentially introduced, superimposing random perturbation noise onto the feature matrix to force the model to learn robust features insensitive to measurement errors.

[0015] A second aspect of this invention provides a deep learning-based intelligent evaluation method for automotive sound effects.

[0016] This method achieves sound effect evaluation by performing a series of ordered data processing steps: First, the original audio signal of the vehicle under test is acquired under preset operating conditions; then, the signal is analyzed to extract psychoacoustic parameters such as loudness, sharpness, roughness, fluctuation, and annoyance, and a temporal feature matrix is ​​constructed according to the sampling time sequence; then, the matrix is ​​input into a one-dimensional convolutional neural network model, and the temporal features and parameter co-features are extracted by using convolution operations; finally, a predicted value representing smoothness, comfort, or luxury is output.

[0017] The technical solution provided by this invention achieves automated and objective evaluation of subjective perception of automotive sound effects by constructing a processing link that combines psychoacoustic parameter serialization with one-dimensional convolutional feature extraction. This solution utilizes five-dimensional psychoacoustic parameters to replace single physical quantities, covering the comprehensive perception of sound intensity, frequency components, modulation characteristics, and periodic fluctuations by the human ear. By using a one-dimensional convolutional neural network to process the time-series matrix, it effectively captures the dynamic changes in sound effects during vehicle operation, improving the consistency between the evaluation results and human auditory perception. Simultaneously, the gradient-based attribution analysis mechanism provides a clear path for optimizing physical parameters in vehicle acoustic design.

[0018] The above solution achieves the following beneficial technical effects:

[0019] This application constructs a temporal feature matrix by extracting five-dimensional psychoacoustic parameters—loudness, sharpness, roughness, fluctuation, and annoyance—as input, replacing the traditional raw sound wave or single spectral feature input. It transforms the physical acoustic signal into an intermediate feature representation that conforms to the human ear's auditory perception model, covering the comprehensive perception of sound intensity, frequency components, and modulation characteristics by the human ear. This provides deep neural networks with input data that has a high signal-to-noise ratio and is strongly correlated with subjective feelings, enabling the objective quantification of abstract subjective indicators such as luxury and comfort.

[0020] This application employs a one-dimensional convolutional neural network (1DCNN) architecture to process the temporal feature matrix. It utilizes the sliding operation of the convolutional kernel in the time dimension to extract features, which aligns with the dynamic evolution of car sound effects with operating conditions (such as changes in vehicle speed). This effectively captures the fluctuation patterns of psychoacoustic parameters on the time axis and the local collaborative patterns between different parameters, avoiding the loss of temporal information by traditional static statistical methods.

[0021] This application introduces a gradient-based feature attribution analysis mechanism, which quantifies the contribution weight of each parameter by calculating the gradient of the predicted value relative to the input matrix. This solves the problem of poor interpretability of deep learning models and can reverse the model prediction results to locate key psychoacoustic parameters (such as roughness anomalies at specific times) that lead to a drop in scores. It transforms the black box model that can only output scores into an analysis tool that can guide the optimization of acoustic engineering. Attached Figure Description

[0022] Figure 1 is a schematic diagram of the system structure of the present invention.

[0023] Figure 2 is a schematic diagram of the audio data acquisition and truth label construction process of the present invention.

[0024] Figure 3 is a schematic diagram of the operation flow of the optimization guidance mechanism of the present invention.

[0025] Figure 4 is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0027] Referring to Figure 1, the present invention provides a deep learning-based intelligent evaluation system for automotive sound effects. The system runs on a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to realize the functions of the system.

[0028] The deep learning-based intelligent evaluation system for automotive audio effects includes: an audio data acquisition module, a psychoacoustic feature extraction module, a deep learning evaluation module, a data augmentation unit, and an optimization guidance module. These modules are connected via a data bus or communication interface to enable one-way or two-way data transmission.

[0029] The audio data acquisition module is configured to acquire the raw audio signal of the vehicle under test under preset operating conditions. The module connects to a high-fidelity microphone array or an in-vehicle audio interface to record acoustic waveform data as the vehicle's speed changes. The module also includes a time synchronization unit to record timestamp information while acquiring the raw audio signal, establishing a temporal correspondence between the audio signal and the vehicle's operating conditions.

[0030] The psychoacoustic feature extraction module, connected to the audio data acquisition module, is configured to receive the raw audio signal and perform time-frequency analysis calculations on it. The module incorporates acoustic calculation algorithms to extract psychoacoustic parameters, including loudness, sharpness, roughness, fluctuation, and annoyance. Further configured, the module arranges the extracted psychoacoustic parameters into a time-series feature matrix based on a discretized sampling time step.

[0031] For each audio sample, the temporal feature matrix constructed by the psychoacoustic feature extraction module is represented as follows: Its mathematical expression is as follows:

[0032] ;

[0033] in, This represents the length of the sampling sequence used to discretize the original audio signal, i.e., the total number of time steps. This indicates the number of types of psychoacoustic parameters extracted; Indicates the first The sampling time, the first The numerical values ​​of psychoacoustic parameters.

[0034] The deep learning evaluation module, connected to the psychoacoustic feature extraction module, is configured to store and run pre-trained deep neural network models. The deep learning evaluation module receives a temporal feature matrix. As input data, the system utilizes one-dimensional convolutional layers, pooling layers, and fully connected layers of a deep neural network model to perform calculations and output predicted values ​​for subjective sound quality evaluation indicators.

[0035] A deep neural network model is defined as a nonlinear mapping function. Its output predicted value vector The calculation is as follows:

[0036] ;

[0037] in, This represents the set of all learnable weight and bias parameters in a deep neural network model. This is a vector that includes scores for smoothness, comfort, and luxury.

[0038] The data augmentation unit, connected to the deep learning evaluation module, is configured to process the input temporal feature matrix during the model training phase. The data augmentation unit generates an augmented feature matrix by superimposing a preset proportion of random perturbation noise onto the values ​​in the temporal feature matrix, and then provides this augmented feature matrix to the deep learning evaluation module for parameter updates.

[0039] The optimization guidance module, connected to the deep learning evaluation module, is configured to perform backpropagation analysis on the deep neural network model. The optimization guidance module calculates predicted values ​​for subjective sound quality evaluation indicators. Temporal feature matrix relative to the input The gradient matrix is ​​calculated, and the contribution weight of each psychoacoustic parameter to the predicted value is calculated based on the gradient matrix. The key parameter types that cause the predicted value to be lower than the preset threshold are identified and output.

[0040] The gradient matrix calculated by the optimization guidance module It is expressed as follows:

[0041] ;

[0042] in, The partial derivative of the output vector with respect to the input matrix is ​​represented by the sign […]. This indicates the absolute value operation.

[0043] Referring to Figure 2, this section details the specific implementation process of the audio data acquisition module in obtaining raw data and constructing subjective evaluation truth labels for supervised learning.

[0044] The audio data acquisition module is equipped with a high-precision acoustic sensor array. In this embodiment, a binaural microphone or an artificial head is used as the sound pickup device, positioned near the driver's ear or the rear passenger's seat in the vehicle under test. The audio data acquisition module sets the sampling frequency to 44.1kHz or 48kHz and the quantization bit depth to 24bit to ensure that the acquired raw audio signal can cover the entire audible frequency range of 20Hz to 20kHz while preserving high dynamic range acoustic details.

[0045] During data acquisition, the audio data acquisition module executes a standardized operating condition acquisition procedure. The vehicle under test runs on a test road or a semi-anechoic chamber rotary drum test bench that meets acoustic testing standards. The acquisition conditions cover the entire process of the vehicle accelerating from a standstill to 100 km / h, including full-throttle rapid acceleration and partial-load slow acceleration. The types of signals recorded by the audio data acquisition module include in-vehicle powertrain noise, road noise, wind noise, and alert sounds emitted by the vehicle's acoustic alarm system.

[0046] To establish a precise correspondence between physical acoustic signals and vehicle operating status, the audio data acquisition module reads real-time vehicle operating parameters via the controller area network bus or on-board diagnostic system interface. These operating parameters include engine speed, vehicle speed, accelerator pedal opening, and motor torque. The audio data acquisition module uses a unified clock source to synchronously tag the raw audio signals and operating parameters, generating a timestamped multi-dimensional data stream to ensure that each frame of audio data corresponds to a unique vehicle operating condition.

[0047] After acquiring the raw audio signal, subjective evaluation ground truth labels need to be constructed for model training. This process is carried out by an evaluation team in a controlled listening room environment. The background noise in the listening room is controlled below 25 dB(A). The evaluation team consists of a predetermined number of evaluators (e.g., 30 people), including acoustic engineers and ordinary drivers, covering different age groups and gender ratios.

[0048] The evaluation employs a rating scale. For each audio sample, evaluators score it based on predefined subjective sound quality evaluation indicators. These indicators are quantified as continuous values ​​from 1 to 10, and are defined as follows:

[0049] Smoothness: Characterizes the linearity and lack of abrupt changes in sound as it changes with engine speed or vehicle speed. A higher score indicates a smoother sound transition.

[0050] Comfort level: This represents the degree of auditory pleasure of a sound. The higher the score, the less oppressive, sharp, or rough the sound is.

[0051] Luxury: Represents the richness and precision of the sound. The higher the score, the fuller the low-frequency components and the clearer the hierarchical structure.

[0052] Regarding the first One audio sample, subjective evaluation of the true value label. The calculation process is as follows: Collect all rating data from all evaluators for this sample, and remove outliers that deviate from the mean by more than the preset standard deviation (e.g., 2 times the standard deviation). Then, calculate the arithmetic mean of the remaining valid ratings as the final true value label for this sample.

[0053] No. The audio sample at the ... True values ​​on each evaluation metric The calculation formula is:

[0054] ;

[0055] in, Indicates the first The number of valid evaluators on the first The sample at the th Scoring on each indicator This represents the total number of valid evaluators after outlier removal. This calculation process transforms subjective auditory perception into an objective numerical vector, which is then compared with the temporal feature matrix defined in the first part above. One-to-one correspondence, forming a complete training dataset. .

[0056] Referring to Figure 4, this section details the specific implementation process of the psychoacoustic feature extraction module in converting physical sound wave signals into a multi-dimensional data structure that conforms to the characteristics of human auditory perception.

[0057] The psychoacoustic feature extraction module first preprocesses the raw audio signal received from the audio data acquisition module. The preprocessing includes removing infrasound noise below 20Hz and ultrasonic noise above 20kHz using an infinite impulse response (IER) or finite impulse response (FIR) filter to eliminate interference from non-auditory frequency signals in the calculation of psychoacoustic parameters. Subsequently, the psychoacoustic feature extraction module uses a time-varying loudness analysis model to perform frame-based processing on the preprocessed signal. In this embodiment, the analysis window length is set to 2ms to 10ms, and the overlap rate is 50%, thereby generating a discrete-time series with high temporal resolution.

[0058] For each discrete time window, the psychoacoustic feature extraction module calculates the instantaneous loudness according to the ISO 532-1:2014 standard. This calculation process first converts the time-domain signal into a frequency-domain signal using a Fast Fourier Transform, then maps the frequency-domain signal to a critical frequency band. Considering the differences in human ear sensitivity to different frequencies and the frequency masking effect, a loudness value in solenoids is calculated. This loudness value characterizes the subjective perception of sound intensity by the human ear, rather than simply the sound pressure level.

[0059] Building upon loudness calculations, the psychoacoustic feature extraction module further calculates sharpness. This calculation, based on the DIN45692 standard or the Aures model, determines the location of the spectral centroid by calculating the first moment of the loudness proportion. The sharpness value is directly proportional to the proportion of high-frequency components in the spectrum, and the unit is acum. This parameter quantifies the harshness or brightness of the high-frequency components in a sound, and has a significant characterizing effect on evaluating car wind noise and high-frequency motor whine.

[0060] The psychoacoustic feature extraction module simultaneously calculates roughness and volatility, both parameters characterizing the amplitude modulation properties of a sound signal. The module distinguishes between the two by extracting the signal envelope and analyzing its modulation spectrum. For amplitude variations with modulation frequencies ranging from 20Hz to 300Hz, roughness is calculated in asper; for slow amplitude variations with modulation frequencies below 20Hz (typically around 4Hz), volatility is calculated in vacil. During vehicle acceleration, the combustion order of the engine or the particle texture of the road surface primarily excites roughness features, while low-speed vehicle body undulations or unstable idling primarily excite volatility features.

[0061] The psychoacoustic feature extraction module calculates the annoyance level at the end. Unlike the four basic parameters mentioned above, the annoyance level is a composite derived parameter. The psychoacoustic feature extraction module is equipped with a weighted calculation unit that, based on the Zwicker psychoacoustic model, combines the loudness, sharpness, roughness, and volatility calculated at the current time step according to a nonlinear weighted formula. Specifically, the annoyance level is proportional to the loudness and is affected by correction terms for sharpness, roughness, and volatility. This parameter is a dimensionless value used to comprehensively characterize the degree of subjective discomfort caused by the sound.

[0062] After completing the frame-by-frame calculation of the above five parameters, the psychoacoustic feature extraction module performs a matrix construction operation. Based on the matrix structure defined in the first part, the psychoacoustic feature extraction module uses the loudness, sharpness, roughness, fluctuation, and annoyance values ​​calculated at the same time as a feature vector. Then, in chronological order, The feature vectors at each time step are stacked along the time dimension to form a shape of dimension . The temporal feature matrix.

[0063] To eliminate differences in the dimensions and numerical ranges of various psychoacoustic parameters and accelerate the convergence of subsequent neural networks, the psychoacoustic feature extraction module is also equipped with a normalization unit. This unit performs Z-score normalization on the generated temporal feature matrix, that is, for each column of the matrix (corresponding to each parameter), subtracts the mean of that parameter in the entire sample set and divides by the standard deviation, so that the data distribution of each parameter in the processed feature matrix approximately satisfies a standard normal distribution with a mean of 0 and a variance of 1. The normalized temporal feature matrix is ​​then transmitted to the deep learning evaluation module as the input tensor of the model.

[0064] Referring to Figure 4, this section elaborates on the specific hierarchical structure, parameter configuration, and computational logic of the neural network inside the deep learning evaluation module. This part is the core computational process for realizing the nonlinear mapping from objective physical parameters to subjective evaluation indicators.

[0065] The deep neural network model built inside the deep learning evaluation module is configured as a serial cascaded architecture, with the data flow passing sequentially through the input interface, the first convolutional layer, the pooling layer, the second convolutional layer, the global pooling layer, the fully connected layer, and the output layer.

[0066] The deep learning evaluation module receives the temporal feature matrix from the psychoacoustic feature extraction module through an input interface. In this embodiment, the input data is set to a three-dimensional tensor, with the three dimensions corresponding to the batch size, time step, and number of feature channels, respectively. The number of feature channels is fixed at 5, corresponding to the five psychoacoustic parameters extracted earlier.

[0067] The first convolutional layer is configured as the front-end network for feature extraction. This first convolutional layer contains a first number of one-dimensional convolutional kernels, and the time window size of the convolutional kernels is set to a preset odd number. During the operation, each one-dimensional convolutional kernel performs a sliding operation on the time dimension of the temporal feature matrix. For a specific convolutional kernel, the calculation process involves performing a dot product operation between the weight vector of the convolutional kernel and the local data of the temporal feature matrix within the current time window, and then adding a bias term to the result.

[0068] To introduce nonlinear feature transformation capabilities, the first convolutional layer is configured with a linear rectified unit as the activation function. For the output value of the convolution operation, the linear rectified unit performs a truncation operation, setting negative values ​​to zero and retaining positive values. Through this operation, the model can filter background noise responses and preserve significant psychoacoustic feature variation patterns, such as abrupt sharp peaks or continuous roughness fluctuations. After processing by the first convolutional layer, the first feature map is generated.

[0069] The pooling layer, connected after the first convolutional layer, is configured to perform downsampling. In this embodiment, max pooling is employed, with a preset pooling window size and stride. This pooling layer selects the maximum value from every two or more adjacent time steps of the first feature map as its output. This operation reduces the length of the time dimension, thereby reducing the amount of data required for subsequent computations and endowing the model with invariance to small time shifts in the input signal; that is, small shifts in the specific timing of acoustic feature occurrences do not affect the overall evaluation result. After processing by the pooling layer, a second feature map is generated.

[0070] The second convolutional layer is configured to extract deep semantic features. This second convolutional layer contains more one-dimensional convolutional kernels than the first convolutional layer, for example, twice the number of kernels in the first convolutional layer. This second convolutional layer receives a second feature map as input and uses the convolutional kernels to capture the cooperative variations between different psychoacoustic parameters and longer-span temporal dependencies. The second convolutional layer also uses linear rectified units as activation functions, generating a third feature map after computation.

[0071] The global pooling layer is configured to handle variable-length inputs. Since the duration of the test vehicle's operation may vary, the length of the third feature map in the time dimension is not fixed. The global pooling layer employs a global average pooling strategy, calculating the arithmetic mean of each feature channel of the third feature map over the entire time axis. Regardless of the input time step size, this operation ultimately generates a fixed-length feature vector, the length of which is equal to the number of convolutional kernels in the second convolutional layer.

[0072] The fully connected layer is configured as a feature fusion and mapping unit. This fully connected layer contains a predetermined number of neurons that perform fully connected operations with the feature vectors. To improve the model's generalization ability, this fully connected layer is configured with a random deactivation mechanism during the training phase, which randomly discards the activation values ​​of some neurons with a predetermined probability to prevent the model from overfitting to specific training samples.

[0073] The output layer, as the end of the network, contains neurons corresponding to the number of subjective evaluation indicators of sound quality. In this embodiment, three neurons are set to correspond to smoothness, comfort, and luxury, respectively. The output layer uses a linear activation function to directly output the weighted summation result, thereby achieving regression prediction of continuous numerical ratings. Through the above hierarchical structure, the deep neural network model gradually abstracts the low-level physical acoustic parameters into high-level acoustic semantic features, completing the mapping from objective parameters to subjective indicators.

[0074] Referring to Figure 4, this section elaborates on the specific operating mechanism of the deep learning evaluation module and data augmentation unit during the model training phase, including the preprocessing and augmentation of training samples, the calculation logic of the loss function, and the iterative optimization process of network parameters.

[0075] Before the training phase begins, the data augmentation unit is configured to perform noise injection on the input temporal feature matrix. Due to the complex driving environment, the acquired audio signals are often accompanied by random physical disturbances or sensor measurement errors. To improve the robustness of the deep neural network model to such minor disturbances, the data augmentation unit does not directly input the original temporal feature matrix into the network, but instead applies random additive noise that follows a Gaussian distribution. The mathematical expression of the data augmentation operation for each data point in the temporal feature matrix is ​​as follows:

[0076] ;

[0077] in, The enhanced temporal feature matrix represents the matrix at the 1st... The time step, the first The values ​​on each feature channel; This represents the numerical value corresponding to the original time-series feature matrix; Indicates the first Standard deviation of psychoacoustic parameters on the training dataset; This represents the preset noise intensity coefficient, which is set to a value between 0.01 and 0.05 in this embodiment; Indicates the distribution from the standard normal distribution Random variables sampled from the middle.

[0078] The augmented samples, processed by the data augmentation unit, are input into the deep learning evaluation module. The deep learning evaluation module performs forward propagation calculations, generating predicted values ​​for subjective sound quality evaluation metrics using the current network parameters. To measure the difference between these predicted values ​​and the manually labeled true subjective evaluation values, the deep learning evaluation module is configured to calculate the mean squared error (MSE) as a loss function. This loss function quantifies the average squared deviation in the regression prediction task.

[0079] For a training batch containing multiple samples, the loss function is calculated using the following formula:

[0080] ;

[0081] in, This represents the total loss value under the current network parameter configuration; Indicates the size of the training batch; This indicates the number of dimensions for the output evaluation metrics, which is 3 in this embodiment; Indicates the first The sample at the th Predicted output values ​​for each evaluation indicator; Indicates the first The sample at the th The corresponding true label value for each evaluation indicator.

[0082] The deep learning evaluation module is equipped with a backpropagation algorithm and a parameter optimizer to minimize the loss function. In this embodiment, the Adaptive Moment Estimation (Adam) optimizer is used to update the network parameters. This optimizer combines the momentum of gradients and the root mean square propagation characteristics, enabling it to adaptively adjust the learning rate for different parameters. In each training iteration, the deep learning evaluation module first calculates the gradient of the loss function with respect to the weight and bias parameters of each layer in the network, and then updates the parameters according to the gradient descent rule.

[0083] The update rules for network parameters are as follows:

[0084] ;

[0085] in, Indicates the first Network parameters at the next iteration; This indicates the updated network parameters; This represents the global learning rate, which is initially set to 0.001 in this embodiment. The main first moment estimate of the gradient (momentum term); The main second-order moment estimate of the gradient (adaptive term); This represents a small constant used to maintain numerical stability.

[0086] To prevent overfitting, the deep learning evaluation module is equipped with an early stopping mechanism. During training, the system divides the dataset into a training set and a validation set. The deep learning evaluation module monitors the changes in the loss value on the validation set. When the validation set loss value does not decrease within a preset number of iterations, the system automatically terminates the training process and saves the network parameters corresponding to the minimum validation set loss value as the final model parameters.

[0087] Referring to Figure 3, this section details the specific implementation process of the optimization guidance module in locating and attributing the dominant physical factors affecting sound quality by utilizing the interpretability of deep neural networks.

[0088] The optimization guidance module is configured to intervene after the deep learning evaluation module completes its forward inference output. First, the optimization guidance module performs a threshold discrimination operation, comparing the predicted values ​​of the subjective sound quality evaluation indicators output by the deep neural network model with preset quality thresholds. When a predicted value for a certain evaluation indicator (such as comfort) is detected to be lower than the quality threshold, the system automatically triggers an attribution analysis process to identify the acoustic causes leading to the low value of that indicator.

[0089] To quantify the impact of input features on the output, the optimization guidance module employs a saliency map calculation method. Unlike the model training phase, which updates network weight parameters, this phase keeps all weight and bias parameters of the deep neural network model frozen and uses the chain rule to calculate the gradient of the predicted value of the target evaluation index relative to the input temporal feature matrix. This gradient reflects the sensitivity of the final evaluation score to small numerical changes at a specific time step and in a specific frequency feature channel.

[0090] For a specific output evaluation metric, the optimization guidance module calculates the original gradient contribution value of each data point in the time-series feature matrix, and the calculation formula is as follows:

[0091] ;

[0092] in, Indicates the first The time step, the first The significance values ​​corresponding to psychoacoustic parameters; This represents the predicted value of the subjective evaluation index of sound quality currently of concern; symbol Represents partial differential operations; symbol This indicates taking the absolute value, which preserves the magnitude information of the gradient while ignoring the direction information. This is because, whether it is positive promotion or negative suppression, a larger gradient magnitude means that the feature has a significant impact on the result.

[0093] After obtaining the saliency matrix containing complete temporal information, the optimization guidance module further calculates the global contribution weight of each psychoacoustic parameter. Since the perception of sound quality is often a cumulative effect throughout the entire operating process, the optimization guidance module aggregates the saliency matrix along the time dimension. Specifically, for each type of psychoacoustic parameter, its average saliency over the entire sampling sequence length is calculated and used as the contribution weight of that parameter.

[0094] The formulas for calculating and normalizing contribution weights are as follows:

[0095] ;

[0096] ;

[0097] in, Indicates the first The absolute contribution of psychoacoustic parameters; Indicates the total number of time steps; This represents the normalized relative contribution weight, which ranges from 0 to 1, and the sum of the relative weights of all parameters is 1.

[0098] The optimization guidance module sorts all psychoacoustic parameters in descending order based on their calculated relative contribution weights. The system selects the parameter with the largest weight value, or the parameter whose sum of weights exceeds a preset proportion (e.g., 80%). These parameters are identified as key factors affecting the current sound quality.

[0099] The optimization guidance module is equipped with an output interface to feed back identified key factors to the user terminal. For example, if the calculation results show that the roughness parameter has the highest contribution weight, the optimization guidance module outputs a prompt message, instructing acoustic engineers to focus on investigating the balance of engine combustion order or the road surface excitation transmission path; if the sharpness parameter has the highest weight, it instructs them to investigate high-frequency whistling of the motor or wind noise sealing defects. Through this mechanism, the system transforms the black-box prediction results of the deep learning model into optimization directions with physical engineering significance.

[0100] Referring to Figure 4, the deep learning-based intelligent evaluation method for car sound effects provided by the present invention relies on the aforementioned computer system and hardware environment for execution, and includes a series of data processing steps executed in a time-sequential manner.

[0101] Step S100: Perform raw data acquisition and operating condition synchronization. The system first detects the vehicle's operating status signal. When it receives a signal indicating zero vehicle speed and an on / off accelerator pedal position sensor via the vehicle controller local area network bus, it triggers an audio acquisition command. The audio data acquisition module activates the high-fidelity microphone array to continuously record the in-vehicle sound field environment at a preset sampling frequency (e.g., 48kHz). During recording, the system synchronously reads vehicle speed and engine speed data from the vehicle bus. Acquisition stops when the vehicle speed reaches 100km / h or a test stop signal is detected. The system aligns the acquired time-domain sound wave data with the vehicle operating condition data along the time axis and extracts the effective audio segment from the start to the target vehicle speed range as the raw audio signal.

[0102] Step S200: Perform time-frequency analysis and feature matrix construction. The psychoacoustic feature extraction module receives the raw audio signal and first filters out background noise using a bandpass filter. Then, the system performs frame-by-frame processing on the filtered signal to generate a continuous sequence of time windows. For each time window, the system calculates physical parameter values ​​in parallel across five dimensions: loudness, sharpness, roughness, volatility, and annoyance, based on the psychoacoustic model. The system stacks the calculated parameter values ​​along the feature dimensions according to the order of the time windows to construct an initial temporal feature matrix. To adapt to the input distribution requirements of the neural network, the system further performs Z-Score normalization on the initial temporal feature matrix, i.e., subtracting the mean of the feature on the statistical benchmark from the feature value in each row of the matrix and dividing by the standard deviation to generate a normalized temporal feature matrix.

[0103] Step S300: Perform deep neural network inference and output evaluation metrics. The system inputs the normalized temporal feature matrix into a pre-trained one-dimensional convolutional neural network model. The data stream first passes through the first convolutional layer, using a sliding convolutional kernel to extract local temporal features; then it passes through a pooling layer for temporal downsampling; subsequently, it passes through the second convolutional layer to extract deep semantic features and collaborative patterns between parameters. After the variable-length temporal features are transformed into fixed-length feature vectors by a global average pooling layer, the data enters a fully connected layer for linear transformation. Finally, the output layer calculates the predicted values ​​of subjective sound quality evaluation metrics, including smoothness, comfort, and luxury, based on the aforementioned nonlinear mapping function. This step only involves forward propagation calculations and does not update parameters.

[0104] Step S400: Perform evaluation result verification and optimization attribution analysis. The system compares the predicted values ​​of the subjective sound quality evaluation indicators output in step S300 with the preset pass thresholds. If the predicted values ​​of all indicators are higher than or equal to the pass thresholds, the system determines that the sound effect of the vehicle under test is qualified and generates a pass report. If the predicted value of any indicator is lower than the pass threshold, the system automatically triggers the optimization guidance process. In this process, the system keeps the network parameters fixed and calculates the gradient matrix of the non-compliant indicators relative to the input temporal feature matrix. Based on the aforementioned contribution weight calculation logic, the system statistically analyzes the gradient magnitude of each psychoacoustic parameter, identifies the parameter type that contributes the most to the low score (e.g., excessively high roughness in a specific frequency band), and outputs the parameter type and its corresponding time segment as optimization suggestions to the user interface, completing the single evaluation loop.

[0105] Referring to Figures 1-4, the deep learning-based intelligent evaluation system for automotive sound effects of the present invention is not limited to a single local computing architecture or a specific network topology. This section elaborates on alternative implementation methods and functional expansion schemes that can be adopted in practical engineering applications.

[0106] In the first alternative implementation, the system adopts a distributed deployment architecture with edge-cloud collaboration. The audio data acquisition module and psychoacoustic feature extraction module are deployed on the in-vehicle edge computing terminal, responsible for real-time acquisition of raw audio signals and performing primary processing from sound wave data to a temporal feature matrix. The in-vehicle edge computing terminal establishes an encrypted connection with a remote cloud server via a 4G / 5G wireless communication module, uploading the temporal feature matrix (after removing privacy-sensitive information) to the cloud. The deep learning evaluation module and optimization guidance module are deployed on a cloud server cluster. The cloud server utilizes a high-performance graphics processing unit (GPU) cluster to run a deep neural network model, performing inference calculations on the received feature matrix and sending the generated subjective sound quality evaluation indicators and optimization suggestions back to the in-vehicle display terminal or the engineering database of the R&D center. This implementation reduces the computing power requirements of the in-vehicle terminal and facilitates centralized iterative updates of the model version.

[0107] In the second alternative implementation, the data dimensions of the temporal feature matrix are expanded to construct a multimodal fusion input. In addition to the five psychoacoustic parameter channels mentioned above (loudness, sharpness, roughness, volatility, and annoyance), the system further acquires vehicle state physical parameters via the vehicle's CAN bus. These physical parameters include real-time engine / motor speed, throttle opening percentage, and longitudinal acceleration. The system resamples and normalizes these physical parameters at the same time step, incorporating them as additional feature channels into the temporal feature matrix. In this case, the feature channel dimension of the input tensor is expanded from 5 to 5+K (K being the number of newly added physical parameters). This implementation allows the deep neural network model to learn the dynamic correlation between sudden changes in vehicle operating conditions (such as sudden acceleration) and acoustic response lag, further improving the prediction accuracy of subjective indicators related to dynamics or responsiveness.

[0108] In a third alternative implementation, a variant design is used for the network architecture of the deep learning evaluation module. While the foregoing embodiments preferably employ a one-dimensional convolutional neural network, recurrent neural network structures or attention mechanisms can be introduced when processing extremely long temporal signals (e.g., continuous acceleration exceeding 30 seconds). Specifically, Long Short-Term Memory (LSTM) network units or gated recurrent units can be used to replace or concatenate the one-dimensional convolutional layers. LSTM units, through their internal forget gate, input gate, and output gate structures, can effectively preserve long-distance temporal dependencies and prevent gradient vanishing. Alternatively, a self-attention layer can be introduced before the global pooling layer to calculate the weight distribution of features at different times along the time axis, enabling the model to automatically focus on critical time segments where sound quality deteriorates (e.g., transient noise during gear shifting), rather than averaging across the entire time period.

[0109] In the fourth alternative implementation, this system is applied to the evaluation of high-frequency howling in electric vehicles. The configuration parameters of the psychoacoustic feature extraction module are adaptively adjusted to suit the characteristics of the electric vehicle powertrain. Specifically, the upper limit of the focus frequency band for spectrum analysis is increased to 20kHz or higher to cover the switching frequency of the motor controller and the gear meshing frequency of the reducer. Simultaneously, when calculating the sharpness parameter, the Aures model, which is more sensitive to high-frequency pure tone components, is used instead of the traditional DIN45692 model. For the output evaluation indicators, the luxury index is replaced with technological sophistication or purity, and the corresponding truth label acquisition standards are redefined to adapt to the unique acoustic design goals of electric vehicles.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based intelligent evaluation system for automotive sound effects, characterized in that, include: The audio data acquisition module is used to acquire the original audio signal of the vehicle under test under preset working conditions; the psychoacoustic feature extraction module is used to perform time-frequency analysis on the original audio signal, extract psychoacoustic parameters, and construct a time-series feature matrix according to the time dimension. The deep learning evaluation module is used to store and run a pre-trained deep neural network model. The deep neural network model is used to receive the temporal feature matrix as input and output the predicted value of the subjective evaluation index of sound quality. The deep neural network model is obtained by supervised learning training based on sample audio signals and their corresponding subjective evaluation real labels that are manually labeled.

2. The intelligent evaluation system for car sound effects based on deep learning according to claim 1, characterized in that, The audio data acquisition module specifically collects in-vehicle sound effects, vehicle acoustic alarm system sounds, or engine sounds as the vehicle speed changes from 0 km / h to km / h, and records the timestamps simultaneously.

3. The intelligent evaluation system for car sound effects based on deep learning according to claim 2, characterized in that, The psychoacoustic parameters extracted by the psychoacoustic feature extraction module include at least the following five parameters: loudness, sharpness, roughness, fluctuation, and annoyance.

4. The intelligent evaluation system for car sound effects based on deep learning according to claim 3, characterized in that, The temporal feature matrix is ​​constructed as a two-dimensional data structure containing a time dimension and a feature dimension; wherein, the length of the time dimension corresponds to the length of the sampling sequence for discretizing the original audio signal, and the length of the feature dimension corresponds to the number of types of extracted psychoacoustic parameters; the data units in the temporal feature matrix represent the values ​​of the corresponding category of psychoacoustic parameters at the sampling time.

5. The intelligent evaluation system for car sound effects based on deep learning according to claim 4, characterized in that, The deep neural network model adopts a one-dimensional convolutional neural network architecture, which sequentially includes: a first convolutional layer, used to receive the temporal feature matrix, and perform sliding convolution operations on the time dimension of the temporal feature matrix using multiple one-dimensional convolutional kernels to output a first feature map; a pooling layer, used to downsample the first feature map to output a second feature map with a compressed time dimension; a second convolutional layer, used to perform convolution operations on the second feature map to output a third feature map with an increased number of channels; a global pooling layer, used to perform averaging or maximization operations on the third feature map in the time dimension to generate a one-dimensional feature vector of fixed length; and a fully connected layer and an output layer, used to receive the one-dimensional feature vector and output the predicted value of the subjective evaluation index of sound quality through linear transformation.

6. The intelligent evaluation system for car sound effects based on deep learning according to claim 5, characterized in that, The deep learning evaluation module uses mean squared error as the loss function during the training phase, and updates the network parameters by minimizing the difference between the predicted value and the true label of the subjective evaluation.

7. The intelligent evaluation system for car sound effects based on deep learning according to claim 1, characterized in that, It also includes a data augmentation unit, which is used to apply a preset proportion of random perturbation noise to the values ​​in the time-series feature matrix during the training process to generate augmented training samples.

8. The intelligent evaluation system for car sound effects based on deep learning according to claim 5, characterized in that, It also includes an optimization guidance module, which is used to perform the following operations: calculate the gradient of the predicted value of the subjective evaluation index of sound quality relative to the temporal feature matrix of the input end; calculate the contribution weight of each psychoacoustic parameter to the predicted value based on the gradient; when the predicted value is lower than a preset threshold, identify the psychoacoustic parameter with the highest contribution weight as the key factor affecting sound quality and output it.

9. The intelligent evaluation system for car sound effects based on deep learning according to claim 5, characterized in that, The first and second convolutional layers are configured to use linear rectified units as activation functions to introduce nonlinear feature transformations; the output layer is configured to use a linear activation function to predict the subjective sound quality evaluation index in continuous numerical form based on the feature vector regression.

10. A deep learning-based intelligent evaluation method for car sound effects, characterized in that, The application of the deep learning-based intelligent evaluation system for automotive sound effects as described in any one of claims 1-9 includes: acquiring the original audio signal of the vehicle under test under preset operating conditions; performing time-frequency analysis on the original audio signal, extracting psychoacoustic parameters, and constructing a temporal feature matrix based on the time dimension; inputting the temporal feature matrix into a pre-trained deep neural network model, and outputting a predicted value of the subjective evaluation index of the sound quality of the vehicle under test through the deep neural network model; wherein, the deep neural network model is obtained through supervised learning training based on sample audio signals and their corresponding, manually labeled, true subjective evaluation labels.

Images