A method for evaluating vehicle transient noise annoyance based on human ear dynamic characteristics

Abstract

The present application relates to a kind of vehicle transient noise annoyance evaluation methods based on human ear dynamics characteristics, comprising: collecting the vehicle in transient condition under the vehicle noise signal, constructs the vehicle transient noise sample database;Subjective annoyance evaluation is carried out to the vehicle transient noise signal, based on the vehicle transient noise signal and subjective annoyance evaluation result, constructs the vehicle transient noise annoyance prediction model;Vehicle transient noise sample corresponding Jerk-PCD fusion feature vector is obtained, and Jerk-PCD fusion feature vector is input into the vehicle transient noise annoyance prediction model after training, and the output vehicle transient noise annoyance prediction value.This method can improve the objectivity and accuracy of transient noise annoyance evaluation from the perspective of auditory physiological response.

Description

Technical Field

[0001] This invention relates to the field of vehicle noise quality evaluation and NVH performance analysis technology, and in particular to a method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics characteristics. Background Technology

[0002] When a vehicle travels over uneven road surfaces such as speed bumps, transient impact noise is generated inside the vehicle. This type of noise is characterized by its suddenness, significant time-varying nature, unstable frequency spectrum, and prominent subjective annoyance, which can easily cause auditory irritation for drivers and passengers, seriously affecting the comfort of the in-vehicle acoustic environment and the overall quality of the vehicle.

[0003] Existing methods for evaluating in-vehicle noise mainly include sound pressure level evaluation, spectrum analysis, and evaluation methods based on psychoacoustic parameters such as loudness, sharpness, roughness, and fluctuation intensity. While these methods can characterize the physical intensity of noise and some auditory perception characteristics to a certain extent, most are designed for steady-state or quasi-steady-state noise. For strong transient signals like vehicle impact noise, they often struggle to depict the dynamic response process within the auditory system and reflect changes in agitation caused by transient stimuli. In particular, existing indicators largely focus on the external physical representation of sound signals, lacking consideration for the human ear's response mechanisms. Therefore, they remain insufficient in describing the subjective agitation caused by transient noise.

[0004] On the other hand, while subjective evaluations can directly reflect occupants' feelings of irritation, they suffer from high evaluation costs, large fluctuations in subject states, and poor repeatability and consistency, making them difficult to directly apply to large-scale, high-efficiency engineering applications. Therefore, how to extract key features from in-vehicle transient noise that reflect the transient response of the human ear and further establish a mapping relationship between these features and irritation levels is a crucial issue in current vehicle transient noise evaluation.

[0005] Furthermore, when a vehicle passes over a speed bump, the transient impact noise inside the vehicle typically includes multiple stages, such as front wheel impact, vehicle body structural response, and rear wheel impact. The time interval between the front and rear impacts varies significantly at different vehicle speeds. Traditional fixed windowing, hard truncation, or ordinary windowing methods struggle to simultaneously preserve the impact center and ensure a smooth transition of the background segment, easily leading to loss of transient impact information, abrupt changes in window boundaries, or spectral leakage. This, in turn, affects the stability of subsequent human ear model response calculations and agitation feature extraction. Therefore, developing a preprocessing method for in-vehicle transient noise that can adaptively identify the impact center while maintaining signal continuity is of great significance.

[0006] In summary, existing methods for evaluating in-vehicle transient noise mainly rely on traditional acoustic parameters or subjective evaluation results, and generally suffer from insufficient characterization of transient impact response, inadequate consideration of the nonlinear auditory mechanisms of the human ear, and low stability and efficiency in engineering applications. Furthermore, existing methods have not effectively established a mapping relationship between human ear dynamics and subjective agitation levels, lacking a unified evaluation framework that balances physiological interpretability, predictive accuracy, and engineering applicability. Therefore, extracting key features that characterize the transient auditory load and compression distortion effect of vehicle transient noise, and establishing a correlation model between these features and subjective agitation levels, has become a critical technical challenge that urgently needs to be addressed. Summary of the Invention

[0007] In existing methods for evaluating transient noise inside vehicles, traditional acoustic parameters and psychoacoustic parameters are widely used due to their ease of calculation and broad application. However, these methods primarily describe noise characteristics from physical or psychoacoustic perspectives such as sound pressure level, loudness, sharpness, and roughness, making it difficult to fully reflect the dynamic response of the human auditory system under transient impact noise such as speed bumps. Furthermore, transient noise inside vehicles is characterized by short duration, rapid energy changes, strong impact, and significant abrupt changes in auditory perception, making it difficult for traditional steady-state or quasi-steady-state sound quality evaluation methods to accurately characterize the resulting agitation. Simultaneously, most existing objective evaluation methods lack the involvement of the nonlinear transduction mechanisms of the outer ear, middle ear, cochlea, and outer hair cells, leading to discrepancies between evaluation results and actual subjective agitation. Therefore, a new method is urgently needed that can incorporate the physiological auditory response mechanisms of the human ear and establish a mapping relationship between the objective characteristics of transient noise inside vehicles and subjective agitation levels.

[0008] To achieve the above objectives, the present invention provides the following solution: A method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics characteristics, comprising: Collect in-vehicle noise signals under transient operating conditions and construct an in-vehicle transient noise sample database; Subjective annoyance assessment of in-vehicle transient noise signals is conducted, and a vehicle transient noise annoyance prediction model is constructed based on the in-vehicle transient noise signals and the subjective annoyance assessment results. Obtain the Jerk-PCD fusion feature vector corresponding to the vehicle transient noise sample, input the Jerk-PCD fusion feature vector into the trained vehicle transient noise annoyance prediction model, and output the vehicle transient noise annoyance prediction value.

[0009] Optionally, the transient operating condition includes the speed bump operating condition, which is used to characterize the short-term impact noise generated when a vehicle passes over a locally raised obstacle; Constructing the in-vehicle transient noise sample database includes: The collected raw signals are numbered, labeled with operating conditions, and archived as samples. The vehicle speed, sampling frequency, measurement point location, and operating condition type corresponding to each sample are recorded, ultimately forming an in-vehicle transient noise sample database.

[0010] Optionally, a subjective annoyance assessment of in-vehicle transient noise signals includes: The organization's evaluators conducted a subjective annoyance assessment of the transient noise samples inside the vehicle and set an evaluation scale level. Perform consistency checks on the duplicate evaluation results of the evaluators and remove the evaluator data with consistency below the preset threshold; The valid evaluation results that pass the consistency test are averaged to obtain the subjective annoyance label corresponding to each vehicle transient noise sample.

[0011] Optionally, based on in-vehicle transient noise signals and subjective annoyance evaluation results, a vehicle transient noise annoyance prediction model is constructed, including: The in-vehicle transient noise signal is filtered and trimmed to retain effective noise segments containing the main transient impact events; Transient impact detection is performed on the clipped in-vehicle transient noise signal, the impact center is selected, and an adaptive bimodal smoothing window function is constructed based on the impact center to obtain the adaptively windowed in-vehicle transient noise signal; The adaptive windowed in-vehicle transient noise signal is input into a preset human ear dynamics model, and the nonlinear cochlear response and the linear reference cochlear response are solved respectively. The physiological jerk response curve is calculated based on the nonlinear cochlear response, and the Jerk features are extracted. The physiological compression distortion (PCD) response curve is calculated based on the nonlinear cochlear response and the linear reference cochlear response, and PCD features are extracted. The Jerk features and PCD features are fused to construct a Jerk-PCD fused feature vector; Using the Jerk-PCD fusion feature vector as input and the subjective annoyance evaluation result as label, the vehicle transient noise annoyance prediction model is established.

[0012] Optionally, transient impact detection is performed on the clipped in-vehicle transient noise signal, and the impact center is selected from: The Teager-Kaiser energy operator was used to perform transient enhancement processing on the clipped in-vehicle transient noise signal to obtain the transient energy sequence; The transient energy sequence is truncated for negative values, smoothed by moving average, and normalized to obtain the normalized transient energy curve; Peak detection is performed based on the normalized transient energy curve, and the main impact center is selected according to the magnitude of the peak value.

[0013] Optionally, an adaptive bimodal smoothing window function is constructed based on the impact center to obtain the in-vehicle transient noise signal after adaptive windowing, including: Construct a local gain window function for each impact center; Maintain full gain within the impact retention range near the impact center; Retain a preset percentage of background gain in non-impact background areas; Cosine fade-in and cosine fade-out segments are set before and after the impact retention interval, respectively. For the local gain windows corresponding to multiple impact centers, a global adaptive gain window is obtained by the maximum value synthesis method, and the global adaptive gain window is then subjected to secondary smoothing. Applying the double-smoothed window function to the original signal yields the adaptively windowed in-vehicle transient noise signal.

[0014] Optionally, the impact retention interval is determined based on the impact center position, the retention length before impact, the retention length after impact, and the fade-in / fade-out length; For the first impact event in the bimodal transient impact noise, the post-impact retention length of the first impact event is adaptively adjusted according to the time interval between the two main impact centers to avoid excessive overlap between the front wheel impact window and the rear wheel impact window.

[0015] Optionally, the human ear dynamics model includes an outer ear transmission part, a middle ear transmission part, and a cochlear dynamics response part; The external ear transmission section is used to simulate the acoustic transmission process of transient noise inside the vehicle through the external auditory canal to the tympanic membrane, converting the transient noise signal inside the vehicle into a sound pressure signal at the tympanic membrane. The middle ear transmission section is used to simulate the mechanical transmission process between the tympanic membrane, ossicular chain and stapes, convert the sound pressure signal at the tympanic membrane into a stapes motion response, and obtain the cochlear input signal based on the stapes motion response; The cochlear dynamics response section includes a nonlinear cochlear model and a linear reference cochlear model. The same cochlear input signal obtained through the outer ear transmission section and the middle ear transmission section is input into the nonlinear cochlear model and the linear reference cochlear model respectively to obtain the nonlinear cochlear response and the linear reference cochlear response. The nonlinear cochlear model includes a nonlinear electromechanical transduction mechanism of the outer hair cells. Based on the displacement and velocity of the reticular plate, the receptor current of the outer hair cells is obtained through the nonlinear electromechanical transduction relationship. The receptor current of the outer hair cells causes changes in the membrane potential and contraction response of the outer hair cells. The active force generated by the outer hair cells is fed back to the cochlear dynamics equation, thereby obtaining a nonlinear cochlear response. The linear reference cochlea model is obtained by turning off or linearizing the nonlinear feedback term of the outer hair cells, and is used to obtain the linear reference cochlea response under the same input conditions and compare it with the nonlinear cochlea response.

[0016] Optionally, the Jerk feature is obtained by calculating the higher-order rate of change of the nonlinear cochlear response along the time direction and performing comprehensive processing at the spatial location of the cochlea, and is used to characterize the intensity of the abrupt change in the cochlear response caused by transient noise. The Jerk features include at least one of the following: Jerk maximum value, Jerk root mean square value, Jerk 95th percentile value, Jerk area under the curve, Jerk maximum occurrence time, and Jerk high response duration. The PCD feature is obtained by comparing the difference between the nonlinear cochlear response and the linear reference cochlear response, and is used to characterize the degree of nonlinear compression distortion of the cochlea; The PCD features include at least one of the following: PCD maximum value, PCD root mean square value, PCD 95th percentile value, PCD curve area, PCD maximum occurrence time, and PCD high response duration. The Jerk-PCD fusion feature vector includes Jerk features, PCD features, and interaction features between Jerk features and PCD features; the interaction features include at least one of the following: the ratio between Jerk cumulative response and PCD cumulative response, the product between Jerk maximum value and PCD maximum value, the product between Jerk 95th quantile value and PCD 95th quantile value, and the product between Jerk root mean square value and PCD root mean square value.

[0017] Optionally, the vehicle transient noise annoyance prediction model adopts a machine learning regression model to establish a nonlinear mapping relationship between the Jerk-PCD fused feature vector and the subjective annoyance label; The machine learning regression model includes the XGBoost regression model, which optimizes model parameters by partitioning the training and test sets and using K-fold cross-validation. The machine learning regression model adopts a sample weighted training method, assigning weights to different training samples based on the distribution of Jerk-PCD fusion features, the distribution of subjective irritability labels, or the values ​​of key interaction features. The predicted values ​​of vehicle transient noise agitation are output in the form of continuous numerical values ​​at different levels and are used for vehicle transient noise sound quality analysis and NVH performance optimization.

[0018] The beneficial effects of this invention are as follows: By introducing a human ear dynamics model, this invention elevates the analysis of in-vehicle transient noise from traditional sound pressure signal analysis to the level of auditory physiological response analysis, making the evaluation process closer to the actual human auditory perception mechanism. By incorporating the MET nonlinear transduction of outer hair cells and the cochlear active compression mechanism, it can characterize the nonlinear dynamic response of transient noise generated within the cochlea. By constructing Jerk features, it can effectively describe the intensity of abrupt changes in auditory response caused by transient impact noises such as speed bumps and road seams. By constructing PCD features, it can characterize the impact of cochlear nonlinear compression distortion on the perception of irritability. By fusing Jerk features, PCD features, and their interaction features, and combining them with a machine learning model to establish a mapping relationship with subjective irritability, the accuracy and stability of in-vehicle transient noise irritability prediction can be improved. Therefore, this invention forms an in-vehicle transient noise irritability evaluation method with auditory physiological basis, objective feature support, and engineering application value, providing a scientific basis and technical support for vehicle NVH performance optimization, in-vehicle sound quality improvement, and transient impact noise control. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of a method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the adaptive bi-peak smoothing window processing flow for in-vehicle transient noise according to an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the actual and predicted values ​​of vehicle transient noise annoyance according to an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the importance analysis of Jerk-PCD fusion features in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, this embodiment proposes a method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics, including: Collect in-vehicle noise signals under transient operating conditions and construct an in-vehicle transient noise sample database; Subjective annoyance assessment of transient noise signals inside the vehicle; The transient noise signals inside the vehicle are filtered and trimmed to retain effective noise segments containing the main transient impact events. Transient impact detection is performed on the clipped in-vehicle transient noise signal, the impact center is selected, and an adaptive bimodal smoothing window function is constructed based on the impact center to obtain the adaptively windowed in-vehicle transient noise signal; The adaptive windowed in-vehicle transient noise signal is input into a preset human ear dynamics model, and the nonlinear cochlear response and the linear reference cochlear response are solved respectively. The physiological jerk response curve is calculated based on the nonlinear cochlear response, and the Jerk features are extracted. The physiological compression distortion (PCD) response curve is calculated based on the nonlinear cochlear response and the linear reference cochlear response, and PCD features are extracted. The Jerk features and PCD features are fused to construct a Jerk-PCD fused feature vector; Using the Jerk-PCD fusion feature vector as input and the subjective annoyance evaluation result as label, the vehicle transient noise annoyance prediction model is established.

[0024] Obtain the Jerk-PCD fusion feature vector corresponding to the vehicle transient noise sample, input the Jerk-PCD fusion feature vector into the trained vehicle transient noise annoyance prediction model, and output the vehicle transient noise annoyance prediction value.

[0025] Furthermore, the transient operating condition includes the speed bump condition, which is used to characterize the short-term impact noise generated when a vehicle passes over a locally raised obstacle. Constructing the in-vehicle transient noise sample database includes: The collected raw signals are numbered, labeled with operating conditions, and archived as samples. The vehicle speed, sampling frequency, measurement point location, and operating condition type corresponding to each sample are recorded, ultimately forming an in-vehicle transient noise sample database.

[0026] Furthermore, the subjective annoyance assessment of in-vehicle transient noise signals includes: The organization's evaluators conducted a subjective annoyance assessment of the transient noise samples inside the vehicle and set an evaluation scale level. Perform consistency checks on the duplicate evaluation results of the evaluators and remove the evaluator data with consistency below the preset threshold; The valid evaluation results that pass the consistency test are averaged to obtain the subjective annoyance label corresponding to each vehicle transient noise sample.

[0027] Furthermore, transient impact detection is performed on the trimmed in-vehicle transient noise signal, and the impact center is selected as follows: The Teager-Kaiser energy operator was used to perform transient enhancement processing on the clipped in-vehicle transient noise signal to obtain the transient energy sequence; The transient energy sequence is truncated for negative values, smoothed by moving average, and normalized to obtain the normalized transient energy curve; Peak detection is performed based on the normalized transient energy curve, and the main impact center is selected according to the magnitude of the peak value.

[0028] Furthermore, based on the impact center, an adaptive bimodal smoothing window function is constructed to obtain the in-vehicle transient noise signal after adaptive windowing, including: Construct a local gain window function for each impact center; Maintain full gain within the impact retention range near the impact center; Retain a preset percentage of background gain in non-impact background areas; Cosine fade-in and cosine fade-out segments are set before and after the impact retention interval, respectively. For the local gain windows corresponding to multiple impact centers, a global adaptive gain window is obtained by the maximum value synthesis method, and the global adaptive gain window is then subjected to secondary smoothing. Applying the double-smoothed window function to the original signal yields the adaptively windowed in-vehicle transient noise signal.

[0029] Furthermore, the impact retention interval is determined based on the impact center position, the retention length before impact, the retention length after impact, and the fade-in / fade-out length; For the first impact event in the bimodal transient impact noise, the post-impact retention length of the first impact event is adaptively adjusted according to the time interval between the two main impact centers to avoid excessive overlap between the front wheel impact window and the rear wheel impact window.

[0030] Furthermore, the human ear dynamics model includes an outer ear transmission part, a middle ear transmission part, and a cochlear dynamics response part; The external ear transmission section is used to simulate the acoustic transmission process of transient noise inside the vehicle through the external auditory canal to the tympanic membrane, converting the transient noise signal inside the vehicle into a sound pressure signal at the tympanic membrane. The middle ear transmission section is used to simulate the mechanical transmission process between the tympanic membrane, ossicular chain and stapes, convert the sound pressure signal at the tympanic membrane into a stapes motion response, and obtain the cochlear input signal based on the stapes motion response; The cochlear dynamics response section includes a nonlinear cochlear model and a linear reference cochlear model. The same cochlear input signal obtained through the outer ear transmission section and the middle ear transmission section is input into the nonlinear cochlear model and the linear reference cochlear model respectively to obtain the nonlinear cochlear response and the linear reference cochlear response. The nonlinear cochlear model includes a nonlinear electromechanical transduction mechanism of the outer hair cells. Based on the displacement and velocity of the reticular plate, the receptor current of the outer hair cells is obtained through the nonlinear electromechanical transduction relationship. The receptor current of the outer hair cells causes changes in the membrane potential and contraction response of the outer hair cells. The active force generated by the outer hair cells is fed back to the cochlear dynamics equation, thereby obtaining a nonlinear cochlear response. The linear reference cochlea model is obtained by turning off or linearizing the nonlinear feedback term of the outer hair cells, and is used to obtain the linear reference cochlea response under the same input conditions and compare it with the nonlinear cochlea response.

[0031] The outer ear transmission section and the middle ear transmission section are shared front-end transmission links in both cochlear models, and their output cochlear input signals serve as the same inputs for both models. The main difference between the two cochlear models lies in the electromechanical transduction relationship of the outer hair cells, which uses a nonlinear form and a linear reference form, respectively.

[0032] Furthermore, the Jerk feature is obtained by calculating the higher-order rate of change of the nonlinear cochlear response along the time direction and performing comprehensive processing at the spatial location of the cochlea, and is used to characterize the intensity of the abrupt change in the cochlear response caused by transient noise. The Jerk features include at least one of the following: Jerk maximum value, Jerk root mean square value, Jerk 95th percentile value, Jerk area under the curve, Jerk maximum occurrence time, and Jerk high response duration. The PCD feature is obtained by comparing the difference between the nonlinear cochlear response and the linear reference cochlear response, and is used to characterize the degree of nonlinear compression distortion of the cochlea; The PCD features include at least one of the following: PCD maximum value, PCD root mean square value, PCD 95th percentile value, PCD curve area, PCD maximum occurrence time, and PCD high response duration. The Jerk-PCD fusion feature vector includes Jerk features, PCD features, and interaction features between Jerk features and PCD features; the interaction features include at least one of the following: the ratio between Jerk cumulative response and PCD cumulative response, the product between Jerk maximum value and PCD maximum value, the product between Jerk 95th quantile value and PCD 95th quantile value, and the product between Jerk root mean square value and PCD root mean square value.

[0033] Furthermore, the vehicle transient noise annoyance prediction model employs a machine learning regression model to establish a nonlinear mapping relationship between the Jerk-PCD fusion feature vector and the subjective annoyance label. The machine learning regression model includes the XGBoost regression model, which optimizes model parameters by partitioning the training and test sets and using K-fold cross-validation. The machine learning regression model adopts a sample weighted training method, assigning weights to different training samples based on the distribution of Jerk-PCD fusion features, the distribution of subjective irritability labels, or the values ​​of key interaction features. The predicted values ​​of vehicle transient noise agitation are output in the form of continuous numerical values ​​at different levels and are used for vehicle transient noise sound quality analysis and NVH performance optimization.

[0034] like Figure 1 As shown in this embodiment, a method for evaluating the annoyance level of in-vehicle transient noise based on the fusion of human ear dynamics model and Jerk-PCD features is proposed. The overall process includes the acquisition and preprocessing of in-vehicle transient noise samples, solving the human ear dynamics model response, extracting Jerk features, extracting PCD features, constructing the Jerk-PCD fused feature vector, training the annoyance level prediction model, and outputting the evaluation results. The method in this embodiment includes: The first step involves collecting in-vehicle transient noise signals at different measurement points under transient conditions, including speed bump conditions. Speed ​​bump conditions characterize the short-term impact noise generated when a vehicle passes over a locally protruding obstacle. The collected raw signals are numbered, labeled with the operating condition, and archived as samples. The vehicle speed, sampling frequency, measurement point location, and operating condition type for each sample are recorded, ultimately forming an in-vehicle transient noise sample database.

[0035] The second step involves organizing evaluators to conduct subjective annoyance assessments of the transient noise samples inside the vehicle. An evaluation scale of 1-7 is used, where 1 indicates perceptible but not disruptive, and 7 indicates intolerable. To minimize the impact of the playback order on the evaluation results, the transient noise samples inside the vehicle are randomly shuffled and evaluated according to different playback orders. The evaluation matrix is ​​represented as follows: (1) in, Indicates the first j The evaluator commented on the first i Annoyance score of a transient noise sample inside the vehicle during the kth playback; N This represents the number of transient noise samples inside the vehicle. M The number of people being evaluated; k=1,2 indicates that the same noise sample is played twice.

[0036] Furthermore, to ensure that each evaluator follows the same standard in two evaluations of the same noise sample, the Pearson correlation coefficient is introduced for internal reliability testing, with the formula as follows: (2) in, is the Pearson correlation coefficient between the two evaluation results of the j-th evaluator; N is the number of transient noise samples inside the vehicle. and These are the two evaluation scores given by the j-th evaluator for the i-th noise sample; and is the average score of the j-th evaluator's two evaluations of all noise samples.

[0037] Based on the obtained Pearson correlation coefficient, if the correlation coefficient between two subjective evaluations by an evaluator is less than 0.7, the evaluator is removed, and other evaluators are retained for further analysis.

[0038] After passing the above consistency test, the scoring results of the valid evaluators were averaged to obtain the subjective annoyance label for each in-vehicle transient noise sample: (3) in, Indicates the first i Subjective annoyance level label for each transient noise sample inside the vehicle; This indicates the number of valid evaluators after passing the reliability test.

[0039] The third step involves filtering and trimming the collected in-vehicle transient noise signals, removing invalid and redundant segments, and retaining valid noise segments that reflect the transient impact characteristics of the vehicle when passing over speed bumps. For speed bump conditions, it is preferable to retain complete segments that include the entire process of front wheel impact, vehicle body response, and rear wheel impact.

[0040] The fourth step involves adaptive windowing processing of the cropped in-vehicle transient noise signal. First, transient impact detection and energy feature extraction are performed. To achieve accurate localization of impact events at speed bumps and road seams, the Teager-Kaiser energy operator is used for transient enhancement processing of the signal, and the transient energy sequence of the signal is calculated. (4) in, Indicates the first i The sample at the th n Teager–Kaiser transient energy at each sampling point; Indicates the first step after the cropping is completed. A discrete sound pressure level signal of transient noise inside the vehicle; n is the sampling point number. This operator can simultaneously reflect the amplitude and local variation characteristics of the transient signal, and is used to enhance the location of the impact event. To eliminate negative interference and enhance the effective impact, negative energy values ​​are truncated to zero, and a short-time moving average is applied to the energy sequence: (5) in, This represents the transient energy sequence after truncating negative values ​​and undergoing short-time smoothing. This represents the length of the moving average window, preferably... ; The sampling frequency is used. Negative truncation avoids interference from non-physical negative energy on peak detection, while moving averages suppress local spikes and enhance the stability of the impact energy envelope. The energy sequence is further normalized: (6) in, This represents the normalized transient energy curve; To prevent extremely small positive numbers with a denominator of zero, normalization is used to eliminate the influence of amplitude differences between different samples on the subsequent peak detection threshold. Peak detection is performed based on the normalized energy curve, and a minimum peak spacing is set. With minimum peak threshold Extract the set of impact center locations: (7) in, Indicates the first The set of candidate impact centers detected in each sample; Indicates the first The sampling points where the candidate impact centers are located; Indicates the peak detection threshold, preferably taken as... ; This represents the minimum distance between two adjacent impact centers, preferably taken as... Further sorting by peak amplitude, the two most significant peaks were selected as the double impact centers: (8) in, This represents the final set of major impact centers selected. Indicates the set of candidate impact centers The two peak values ​​of the normalized transient energy are selected from highest to lowest. For the double-impact sample from the speed bump, and The main transient events corresponding to the front wheel impact and rear wheel impact are arranged in chronological order.

[0041] The fifth step involves constructing an adaptive bimodal smoothing window function based on the detected impact center location. For the k-th impact center in the i-th sample... The starting point of the impact retention interval is determined based on the location of the impact center, the retention length before impact, and the retention length after impact. End of the impact retention zone The impact retention interval for the k-th impact event is defined by the start and end points. A local gain window function is constructed for each impact event: (9) in, Indicates the first In the nth sample Local gain window corresponding to each impact event; This represents the retention gain in the non-impact background region, preferably taken as... ; and They represent the first The window defines the start and end points of each impact retention interval. The gain within the impact retention interval is set to 1, and the gain outside the interval is set to γ. To avoid abrupt gain changes at the boundaries of the retention intervals, fade-in and fade-out transition zones are set before the start point and after the end point, respectively. Compared to hard truncation by directly setting the gain to zero, setting the background retention gain can preserve a certain proportion of non-impact background signals, preventing the complete loss of ambient sound information. The window boundaries are defined as follows: (10) in, Indicates the first The starting point of the impact retention range; Indicates the fade-in / fade-out length; Indicates the first The location of the impact center; Indicates the length retained before impact, preferably taken as the preferred length. By introducing The constraint ensures that the front side of the impact retention interval has sufficient length to construct the cosine fade-in segment.

[0042] (11) in, Indicates the first The endpoint of the impact retention zone; This indicates the total number of sampling points for the noise signal; Indicates the first The length of the material retained after an impact. By introducing... Constraints ensure sufficient length behind the impact retention zone for constructing a cosine fade-out segment. In the case of a double impact, the distance between the two impact centers is considered. Adaptive adjustment of the rear window length: (12) in, This represents the post-impact retention length of the first impact event in a double-impact sample. This represents the sampling point interval between the two main impact centers. The formula adaptively adjusts the rear window length of the first impact based on the front and rear wheel impact distance to avoid excessive overlap between the two impact retention intervals. For the second impact or a single impact sample, Preferred selection .

[0043] To avoid abrupt changes at the window function boundaries, a cosine fade-in / fade-out function is introduced: (13) in, This represents the cosine fade-in gain vector; ; Indicates the fade-in / fade-out length, preferably selected. This function makes the gain retain the background value. A smooth transition to step 1 is achieved, while the fade-out phase uses the reverse sequence of this vector to avoid abrupt changes at the window function boundaries. For multiple impact windows, a global window function is constructed using a maximum value synthesis method. (14) in, Indicates the first Global adaptive gain window for each sample; This indicates the number of major impacts detected. Using a maximum synthesis method avoids the gain exceeding 1 caused by the direct superposition of multiple local windows, while ensuring a smooth transition in overlapping areas. To further eliminate discontinuities at window intersections, a secondary smoothing process is applied to the synthesized window function: (15) in, This represents the adaptive window after secondary smoothing. This represents the length of the quadratic smoothing window, preferably taken as... This processing further eliminates potential V-shaped sharp angles or local discontinuities at the intersection of multiple local windows. Finally, the smoothed window function is applied to the original signal to obtain the in-vehicle transient noise signal used as input for the subsequent human ear dynamics model. : (16) The in-vehicle transient noise adaptive bi-peak smoothing window processing flow is as follows: Figure 2 As shown.

[0044] The sixth step is to process the in-vehicle transient noise signal after adaptive windowing. The response is solved by inputting a human ear dynamics model. First, the discrete signal is converted into a continuous-time form. This sound pressure signal is then used as the input sound pressure signal to the outer ear, acting on the outer ear-middle ear system to obtain an equivalent excitation. (17) in, This represents the equivalent stimulus input to the human ear model; The equivalent area of ​​the tympanic membrane is given. The external ear-middle ear model is used to simulate the process of sound waves being transmitted through the external auditory canal, tympanic membrane, and ossicular chain to the oval window, and outputs the middle ear terminal response.

[0045] Step 7: Based on the output of the outer ear-middle ear model, construct a nonlinear cochlear model and solve for the cochlear response. The cochlea is discretized into multiple units along its length, each unit including the basilar membrane, reticular lamina, and outer hair cell structures. The nonlinear electromechanical transduction relationship of the outer hair cells is expressed as: (18) in, For acceptor current; For the displacement of the mesh plate; For the speed of the mesh plate; and These are the displacement and velocity sensitivity coefficients, respectively; It is a nonlinear function. The hyperbolic tangent function is preferred for representation. (19) in, ; This represents the maximum range of variation in the receptor current of the outer hair cells. The saturation compression characteristics during the electromechanical transduction process of the outer hair cells can be described using the hyperbolic tangent function. The introduction of nonlinearity into the electromechanical transduction channel of the outer hair cells is an important foundation for using nonlinear cochlear models to describe the cochlear compression response.

[0046] Therefore, a nonlinear dynamic equation for the cochlea was established, and the nonlinear cochlear response was obtained by numerical methods: (20) in, Indicates the first One sample was taken from the cochlear region. and time The nonlinear cochlear response is described. Preferably, the cochlear response is a response to basilar membrane displacement or reticular plate displacement. This indicates the position coordinates from the base of the cochlea to the apex.

[0047] Step 8: To characterize the nonlinear compression effect of the cochlea, a linear reference cochlea model is established for the same input signal, i.e., the nonlinear feedback term of the outer hair cells is removed to obtain the linear response: (twenty one) in, Indicates the first The cochlear responses of each sample were obtained from a linear reference cochlear model under the same input stimulus. The linear reference cochlear model was obtained by disabling or linearizing the nonlinear feedback term of the outer hair cells and was used for comparison with the nonlinear cochlear response.

[0048] Step 9: Calculate the Jerk response curve based on the nonlinear cochlear response. By taking the third time derivative of the cochlear response and averaging it in the spatial dimension, we obtain: (twenty two) in, Indicates the first Physiological jerk response curves for each sample; Indicates the number of discrete locations in the cochlea; Indicates the first Discrete locations of the cochlea; This represents the nonlinear cochlear response at that location. When When taking the displacement response, the third-order time derivative is used to characterize the degree of abrupt change in the cochlear response, thereby reflecting the intensity of auditory abrupt changes caused by transient impact noise.

[0049] Step 10: Extract Jerk feature parameters from the Jerk response curve, including... , , , , and .in, This represents the maximum value of the Jerk response, used to characterize the strongest transient impact intensity; This represents the root mean square value of the Jerk response, used to characterize the overall rapid energy level; This represents the 95th percentile of the Jerk response, used to characterize the typical intensity of high-level transient responses; The area under the Jerk response curve represents the cumulative level of transient auditory load. Indicates the time when Jerk's maximum value occurs; This indicates the duration of the Jerk high response.

[0050] Step 11: Calculate the compression distortion (PCD) curve based on nonlinear and linear cochlear responses: (twenty three) in, Indicates the first Compression distortion response curves for each sample; Indicates the first Nonlinear cochlear response at each cochlear location; This represents the linear reference cochlear response at the same location and time. To prevent extremely small positive numbers with a denominator of zero, this formula characterizes the degree of deviation of the nonlinear cochlear response from the linear reference response by normalizing the difference, and is used to reflect the influence of cochlear nonlinear compression and the active mechanism of outer hair cells on transient noise perception.

[0051] Step 12: Extract characteristic parameters from the PCD response curve, including... , , , , and .in, This represents the maximum value of the PCD response, used to characterize the strongest nonlinear compression distortion. This represents the root mean square value of the PCD response, used to characterize the overall compression distortion energy level; This represents the 95th percentile of the PCD response, used to characterize the typical intensity of high-level compression distortion; The area under the PCD response curve represents the cumulative effect of compression distortion. Indicates the time when the maximum value of PCD occurs; This indicates the duration of the high response of the PCD.

[0052] Step 13: Fuse Jerk features with PCD features to construct a feature vector: (twenty four) Where Xi represents the Jerk-PCD fusion feature vector corresponding to the i-th in-vehicle transient noise sample, which is used to simultaneously characterize the intensity of cochlear response mutation and the degree of cochlear nonlinear compression distortion.

[0053] Step fourteen: Construct the interaction features between Jerk and PCD to characterize the relative relationship and coupling between transient impact intensity and cochlear nonlinear compression distortion. (25) in, This represents the ratio characteristic between the Jerk cumulative response and the PCD cumulative response; This represents the area under the Jerk response curve; Indicates the area under the PCD response curve; To prevent extremely small positive numbers with a denominator of zero, this interactive feature is used to characterize the relative contribution relationship between transient impact intensity and cochlear nonlinear compression distortion.

[0054] Step 15: Construct the dataset by combining the feature vectors with the subjective annoyance labels: (26) in, This represents the dataset used for modeling irritability prediction; Indicates the first Jerk-PCD fusion feature vectors for each sample; Indicates the first The subjective annoyance level label corresponding to each sample; This represents the total number of samples.

[0055] Step sixteen: Build an XGBoost-based regression model to predict irritability, the output of which is: (27) in, Indicates the first Predicted annoyance level for each sample; Xi represents the trained XGBoost regression model; Xi represents the first... The model generates a Jerk-PCD fusion feature vector for each sample. It then learns a non-linear mapping between the fusion features and the subjective annoyance label to predict the annoyance level of unknown in-vehicle transient noise samples.

[0056] Furthermore, to enhance the model's learning ability on key transient samples and highly discriminative samples, the XGBoost regression model can employ a sample-weighted training method. Specifically, based on the Jerk-PCD fusion feature distribution, subjective annoyance label distribution, or key interaction feature values ​​of the training samples, different weights are assigned to different training samples. This ensures that samples with strong transient impact characteristics, significant cochlear nonlinear compression distortion characteristics, or high annoyance discrimination value contribute more to the model training process. These sample weights, as optional parameters in the XGBoost regression model training process, are used to adjust the influence of different training samples on the loss function, thereby improving the stability and generalization ability of in-vehicle transient noise annoyance prediction.

[0057] Comparison of actual and predicted values ​​of vehicle transient noise annoyance Figure 3 As shown, the importance analysis of Jerk-PCD fusion features is as follows: Figure 4 As shown.

[0058] Step seventeen involves evaluating model performance using Pearson correlation coefficient, root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) as evaluation metrics. The Pearson correlation coefficient characterizes the correlation between predicted agitation level and subjective agitation level; RMSE and MAE characterize model prediction error; and R² characterizes the model's ability to explain changes in subjective agitation level. In a preferred embodiment, an XGBoost model without sample weights is compared with an XGBoost model with sample weights to verify the effectiveness of the sample weight mechanism in improving prediction correlation, prediction error, and model stability.

[0059] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics, characterized in that, include: Collect in-vehicle noise signals under transient operating conditions and construct an in-vehicle transient noise sample database; Subjective annoyance assessment of in-vehicle transient noise signals is conducted, and a vehicle transient noise annoyance prediction model is constructed based on the in-vehicle transient noise signals and the subjective annoyance assessment results. Obtain the Jerk-PCD fusion feature vector corresponding to the vehicle transient noise sample, input the Jerk-PCD fusion feature vector into the trained vehicle transient noise annoyance prediction model, and output the vehicle transient noise annoyance prediction value.

2. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 1, characterized in that, The transient operating conditions include the speed bump operating conditions, which are used to characterize the short-term impact noise generated when a vehicle passes over a locally protruding obstacle. Constructing the in-vehicle transient noise sample database includes: The collected raw signals are numbered, labeled with operating conditions, and archived as samples. The vehicle speed, sampling frequency, measurement point location, and operating condition type corresponding to each sample are recorded, ultimately forming an in-vehicle transient noise sample database.

3. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 1, characterized in that, Subjective annoyance assessment of in-vehicle transient noise signals includes: The organization's evaluators conducted a subjective annoyance assessment of the transient noise samples inside the vehicle and set an evaluation scale level. Perform consistency checks on the duplicate evaluation results of the evaluators and remove the evaluator data with consistency below the preset threshold; The valid evaluation results that pass the consistency test are averaged to obtain the subjective annoyance label corresponding to each vehicle transient noise sample.

4. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 1, characterized in that, Based on in-vehicle transient noise signals and subjective annoyance evaluation results, a vehicle transient noise annoyance prediction model is constructed, including: The in-vehicle transient noise signal is filtered and trimmed to retain effective noise segments containing the main transient impact events; Transient impact detection is performed on the clipped in-vehicle transient noise signal, the impact center is selected, and an adaptive bimodal smoothing window function is constructed based on the impact center to obtain the adaptively windowed in-vehicle transient noise signal; The adaptive windowed in-vehicle transient noise signal is input into a preset human ear dynamics model, and the nonlinear cochlear response and the linear reference cochlear response are solved respectively. The physiological jerk response curve is calculated based on the nonlinear cochlear response, and the Jerk features are extracted. The physiological compression distortion (PCD) response curve is calculated based on the nonlinear cochlear response and the linear reference cochlear response, and PCD features are extracted. The Jerk features and PCD features are fused to construct a Jerk-PCD fused feature vector; Using the Jerk-PCD fusion feature vector as input and the subjective annoyance evaluation result as label, the vehicle transient noise annoyance prediction model is established.

5. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 4, characterized in that, Transient impact detection was performed on the trimmed in-vehicle transient noise signal, and the impact center was selected including: The Teager-Kaiser energy operator was used to perform transient enhancement processing on the clipped in-vehicle transient noise signal to obtain the transient energy sequence; The transient energy sequence is truncated for negative values, smoothed by moving average, and normalized to obtain the normalized transient energy curve; Peak detection is performed based on the normalized transient energy curve, and the main impact center is selected according to the magnitude of the peak value.

6. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 4, characterized in that, Based on the impact center, an adaptive bimodal smoothing window function is constructed to obtain the in-vehicle transient noise signal after adaptive windowing, including: Construct a local gain window function for each impact center; Maintain full gain within the impact retention range near the impact center; Retain a preset percentage of background gain in non-impact background areas; Cosine fade-in and cosine fade-out segments are set before and after the impact retention interval, respectively. For the local gain windows corresponding to multiple impact centers, a global adaptive gain window is obtained by the maximum value synthesis method, and the global adaptive gain window is then subjected to secondary smoothing. Applying the double-smoothed window function to the original signal yields the adaptively windowed in-vehicle transient noise signal.

7. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 6, characterized in that, The impact retention interval is determined based on the impact center position, the retention length before impact, the retention length after impact, and the fade-in / fade-out length.

8. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 4, characterized in that, The human ear dynamics model includes an outer ear transmission component, a middle ear transmission component, and a cochlear dynamics response component. The external ear transmission section is used to simulate the acoustic transmission process of transient noise inside the vehicle through the external auditory canal to the tympanic membrane, converting the transient noise signal inside the vehicle into a sound pressure signal at the tympanic membrane. The middle ear transmission section is used to simulate the mechanical transmission process between the tympanic membrane, ossicular chain and stapes, convert the sound pressure signal at the tympanic membrane into a stapes motion response, and obtain the cochlear input signal based on the stapes motion response; The cochlear dynamics response section includes a nonlinear cochlear model and a linear reference cochlear model. The same cochlear input signal obtained through the outer ear transmission section and the middle ear transmission section is input into the nonlinear cochlear model and the linear reference cochlear model respectively to obtain the nonlinear cochlear response and the linear reference cochlear response. The nonlinear cochlear model includes a nonlinear electromechanical transduction mechanism of the outer hair cells. Based on the displacement and velocity of the reticular plate, the receptor current of the outer hair cells is obtained through the nonlinear electromechanical transduction relationship. The receptor current of the outer hair cells causes changes in the membrane potential and contraction response of the outer hair cells. The active force generated by the outer hair cells is fed back to the cochlear dynamics equation, thereby obtaining a nonlinear cochlear response. The linear reference cochlea model is obtained by turning off or linearizing the nonlinear feedback term of the outer hair cells, and is used to obtain the linear reference cochlea response under the same input conditions and compare it with the nonlinear cochlea response.

9. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 4, characterized in that, The Jerk feature is obtained by calculating the higher-order rate of change of the nonlinear cochlear response along the time direction and performing comprehensive processing at the spatial location of the cochlea, and is used to characterize the intensity of the abrupt change in the cochlear response caused by transient noise. The Jerk features include at least one of the following: Jerk maximum value, Jerk root mean square value, Jerk 95th percentile value, Jerk area under the curve, Jerk maximum occurrence time, and Jerk high response duration. The PCD feature is obtained by comparing the difference between the nonlinear cochlear response and the linear reference cochlear response, and is used to characterize the degree of nonlinear compression distortion of the cochlea; The PCD features include at least one of the following: PCD maximum value, PCD root mean square value, PCD 95th percentile value, PCD curve area, PCD maximum occurrence time, and PCD high response duration. The Jerk-PCD fusion feature vector includes Jerk features, PCD features, and interaction features between Jerk features and PCD features; the interaction features include at least one of the following: the ratio between Jerk cumulative response and PCD cumulative response, the product between Jerk maximum value and PCD maximum value, the product between Jerk 95th quantile value and PCD 95th quantile value, and the product between Jerk root mean square value and PCD root mean square value.

10. The method for evaluating the annoyance level of vehicle transient noise based on human ear dynamics as described in claim 1, characterized in that, The vehicle transient noise annoyance prediction model uses a machine learning regression model to establish a nonlinear mapping relationship between the Jerk-PCD fusion feature vector and the subjective annoyance label. The machine learning regression model includes the XGBoost regression model, which optimizes model parameters by partitioning the training and test sets and using K-fold cross-validation. The machine learning regression model adopts a sample weighted training method, assigning weights to different training samples based on the distribution of Jerk-PCD fusion features, the distribution of subjective irritability labels, or the values ​​of key interaction features. The predicted values ​​of vehicle transient noise annoyance are output in the form of continuous numerical values ​​at different levels and are used for vehicle transient noise sound quality analysis and NVH performance optimization.

Images