Subjective evaluation method for sound quality of in-vehicle noise under all working conditions

Abstract

The invention discloses a subjective evaluation method for the sound quality of in-vehicle noise under all working conditions, and the method comprises the steps: constructing an in-vehicle noise sample database under all working conditions, and the database comprises a steady-state noise sample and an unsteady-state noise sample; evaluating the steady-state noise sample to obtain a first evaluation result, and mapping the first evaluation result to a continuous evaluation interval to generate a first evaluation result curve; evaluating the unsteady noise sample to obtain a time-varying second evaluation result which takes the first evaluation result as an initial reference to generate a second evaluation result curve; and performing dynamic correlation analysis on the first evaluation result curve and the second evaluation result curve, and establishing a corresponding relation between the subjective evaluation results of the steady-state noise and the non-steady-state noise by calculating an optimal matching path of the two curves. According to the method, the integrity and systematic coverage of an in-vehicle noise sound quality evaluation system are realized, and a unified data basis is provided for sound quality analysis under all working conditions.

Description

Technical Field

[0001] This invention belongs to the field of in-vehicle noise assessment, and in particular relates to a subjective evaluation method for in-vehicle noise quality under all operating conditions. Background Technology

[0002] In in-vehicle noise quality research, subjective evaluation methods are widely used to reflect the true auditory perception of noise by the human ear. Among them, semantic segmentation is a commonly used subjective evaluation method due to its ease of operation and clear semantic level evaluation scale. This method requires evaluators to score noise samples according to a preset level scale, thereby quantifying the auditory perception results. However, this method has certain limitations in practical applications: firstly, the evaluation results lack discrimination and are difficult to effectively reflect subtle differences between noise samples; secondly, the method simultaneously possesses both relative and absolute scale attributes, which can easily lead to conflicts, resulting in a decrease in the stability and consistency of the evaluation results.

[0003] To address the aforementioned shortcomings, existing technologies have proposed pairwise comparison methods. This method compares noisy samples pairwise, replacing absolute scoring with relative judgments, thus improving the reliability of the evaluation to some extent. However, when the sample size is large, the number of comparisons required by the pairwise comparison method increases exponentially, leading to excessive experimental costs and time consumption, limiting its application value. To balance evaluation accuracy and efficiency, some studies have proposed grouped pairwise comparison methods. This involves first grouping samples according to certain criteria, and then performing pairwise comparisons within each group, thereby reducing the number of comparisons and improving efficiency. However, the core challenge of grouped pairwise comparison methods lies in determining the grouping criteria; improper grouping may still affect the scientific validity and accuracy of the evaluation results.

[0004] On the other hand, vehicle noise during operation can be categorized into steady-state and non-steady-state noise. Steady-state noise typically occurs at constant speeds or under constant operating conditions, exhibiting relatively stable acoustic characteristics; non-steady-state noise, however, fluctuates dynamically with speed or time, displaying significant time-varying characteristics. Existing research often analyzes steady-state and non-steady-state noise within a unified evaluation framework. However, due to significant differences in their acoustic characteristics and auditory perception patterns, this approach struggles to reveal the dynamic perceptual characteristics of non-steady-state noise. Furthermore, establishing an effective correlation between steady-state and non-steady-state noise based on their evaluation remains a challenging problem. This deficiency significantly limits the completeness and explanatory power of subjective noise evaluation systems.

[0005] In summary, existing subjective noise evaluation methods have shortcomings in terms of discriminative power, stability, and efficiency. Furthermore, the relationship between steady-state and non-steady-state noise has not been effectively characterized, and a unified evaluation system that balances scientific rigor and systematic approach is lacking. Therefore, establishing a correlation analysis between the subjective evaluation results of steady-state and non-steady-state noise while addressing the inherent problems of semantic segmentation methods has become a critical technical challenge that urgently needs to be overcome. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a subjective evaluation method for in-vehicle noise quality under all operating conditions, aiming to simultaneously improve the accuracy and consistency of steady-state evaluation and construct a unified interpretation system for steady-state and non-steady-state noise. This includes: A database of in-vehicle noise samples under all operating conditions is constructed, the database containing steady-state noise samples and non-steady-state noise samples; The steady-state noise sample is evaluated to obtain a first evaluation result, and the first evaluation result is mapped to a continuous evaluation interval to generate a first evaluation result curve; The non-steady-state noise sample is evaluated to obtain a time-varying second evaluation result, wherein the second evaluation result uses the first evaluation result as an initial benchmark to generate a second evaluation result curve. Dynamic correlation analysis is performed on the first evaluation result curve and the second evaluation result curve. By calculating the optimal matching path between the two curves, the correspondence between the subjective evaluation results of steady-state noise and non-steady-state noise is established.

[0007] Preferably, the process of evaluating the steady-state noise sample to obtain a first evaluation result includes: The steady-state noise samples are initially scored using semantic segmentation to obtain preliminary scoring results. Based on the preliminary scoring results, the samples are grouped to minimize the difference between the mean of each group and the overall mean. Paired comparisons were performed within each group to obtain data on the relative preference relationships between samples. Calculate the capability value of each sample based on the relative preference relationship data; Based on the associated samples, the capability values ​​of cross-group samples are inverted and reconstructed to obtain the global capability values ​​of all samples; The global capability value is mapped to the continuous evaluation interval to obtain the first evaluation result.

[0008] Preferably, the process of initially scoring the steady-state noise samples according to the semantic segmentation method includes: The steady-state noise samples are scored according to a preset rating scale to obtain scoring data; Calculate the intra-rater correlation coefficient and the inter-rater correlation coefficient, and remove data with correlation coefficients below a set threshold to obtain valid rating data; The effective scoring data is averaged to obtain the preliminary scoring results.

[0009] Preferably, the process of grouping samples based on the preliminary scoring results includes: Sort the samples in ascending order of the initial score results and calculate the population mean of all samples; The sorted sample sequence is hierarchically divided, and samples are uniformly drawn from each level to form evaluation groups. The variance of the mean of each group and the mean of the population is calculated. The grouping is adjusted according to the variance so that the difference between the mean of each group and the mean of the population is less than a target preset threshold.

[0010] Preferably, the process of performing pairwise comparison evaluation within each group includes: Within each group, samples are compared pairwise, and a score matrix is ​​generated based on the comparison results; Based on the score matrix, a Bradley-Terry model is constructed, defining the relationship between sample ability values ​​and comparison probabilities; The Bradley-Terry model is solved iteratively by maximum likelihood estimation to obtain the capability value of each sample.

[0011] Preferably, the process of reconstructing the capability values ​​of cross-group samples based on associated samples includes: Two related samples are selected between adjacent groups as anchor points, and the anchor points satisfy the condition that the capability values ​​are distributed on both sides of the group mean and the differences are within a preset range. Calculate the proportional coefficient and translation adjustment amount based on the capacity values ​​of the anchor points in the preceding and following groups; Based on the aforementioned scaling factor and translation adjustment, the capability values ​​of subsequent groups of samples are linearly transformed to obtain global capability values ​​under a unified scale.

[0012] Preferably, the process of mapping the global capability value to the continuous evaluation interval includes: A linear mapping relationship between the global capability value and the preliminary scoring results is established using the least squares method, and the mapping coefficients are obtained. Based on the mapping coefficients, the global capability value is converted to a continuous evaluation interval to obtain a continuous first evaluation result.

[0013] Preferably, the process of evaluating the non-steady-state noise samples includes: Construct an interactive slider evaluation system, where the slider position range corresponds to a continuous evaluation interval; During the sample playback, slider position data is continuously recorded to obtain a time-varying evaluation curve; The time axis of the time-varying evaluation curve is mapped to the vehicle speed axis to generate a vehicle speed-evaluation relationship curve as the second evaluation result curve.

[0014] Preferably, the process of using the first evaluation result as an initial benchmark for the second evaluation result includes: Obtain the initial vehicle speed of the non-steady-state noise sample; Based on the initial vehicle speed, query the corresponding first evaluation result of steady-state noise; Set the initial position of the slider to the first evaluation result value found in the query, so that the initial state of the non-steady-state evaluation is consistent with the steady-state evaluation in terms of scale.

[0015] Preferably, the process of performing dynamic correlation analysis on the first evaluation result curve and the second evaluation result curve includes: Construct a distance matrix between the first evaluation result curve and the second evaluation result curve, and calculate the difference in evaluation values ​​at each vehicle speed point; The cumulative cost function of the distance matrix is ​​calculated using a dynamic programming algorithm to obtain the optimal matching path with the minimum total cost. The overall consistency level is quantified based on the cumulative distance value of the optimal matching path; Based on the local deviation of the optimal matching path from the reference diagonal, identify the lag or advance effect of the non-steady-state evaluation relative to the steady-state evaluation.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention effectively addresses the issues of insufficient accuracy and scale conflicts in evaluation results by combining semantic segmentation and component-based pairwise comparison methods, thereby improving the accuracy and consistency of subjective evaluation of steady-state noise. By constructing an interactive slider system, it achieves continuous recording of the non-steady-state noise perception process, overcoming the limitations of traditional discrete evaluation. Furthermore, by introducing DTW analysis, it establishes a systematic correspondence between steady-state and non-steady-state evaluation results, perfecting a unified interpretation system for both. Therefore, this invention forms a scientific, systematic, and operable subjective evaluation method for in-vehicle noise under all operating conditions, providing solid theoretical support and practical guidance for the optimization and improvement of in-vehicle noise sound quality evaluation. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the implementation of the grouping pairwise comparison method for associating samples in an embodiment of the present invention; Figure 3 This is a schematic diagram of the subjective evaluation results of steady-state noise in an embodiment of the present invention; Figure 4 This is a schematic diagram of a subjective evaluation system for unsteady-state noise according to an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the subjective evaluation of noise under steady-state and unsteady-state operating conditions in an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] like Figure 1 As shown in this embodiment, semantic segmentation is widely used in existing subjective evaluation methods for in-vehicle noise due to its intuitiveness and clear semantic level evaluation scale. However, this method suffers from insufficient differentiation of evaluation results and conflicts between relative and absolute scales, affecting the stability and consistency of the evaluation results. On the other hand, non-steady-state noise has time-varying characteristics, and traditional discrete evaluation methods cannot effectively record its dynamic perception process. Furthermore, existing methods lack a mechanism to effectively correlate steady-state and non-steady-state noise evaluation results, making it difficult to establish an in-vehicle noise sound quality evaluation system under all operating conditions. Therefore, a new method is urgently needed that can simultaneously improve the accuracy of steady-state noise evaluation and establish a unified interpretation system for steady-state and non-steady-state noise results.

[0021] This embodiment provides a subjective evaluation method for in-vehicle noise quality under all operating conditions, including: A database of in-vehicle noise samples under all operating conditions was constructed, which includes steady-state noise samples and non-steady-state noise samples. The steady-state noise sample is evaluated to obtain the first evaluation result, and the first evaluation result is mapped to the continuous evaluation interval to generate the first evaluation result curve; The non-steady-state noise sample is evaluated to obtain a time-varying second evaluation result, wherein the second evaluation result uses the first evaluation result as the initial benchmark to generate the second evaluation result curve; A dynamic correlation analysis was performed on the first and second evaluation result curves. By calculating the optimal matching path between the two curves, the correspondence between the subjective evaluation results of steady-state noise and non-steady-state noise was established.

[0022] Specifically, this embodiment collects noise samples from different vehicles under steady-state and non-steady-state conditions. By filtering and cropping the collected signals, invalid or redundant segments are removed, and effective audio data that can truly reflect the characteristics of in-vehicle noise are retained, thereby constructing a noise sample database under all operating conditions. For steady-state noise samples, a subjective evaluation method is proposed that combines the advantages of semantic segmentation's rating scale with the relative preference relationship analysis of component pairwise comparison. This method transforms discrete absolute scores into continuously distributed ability values, maps them to a subjective evaluation range of 1–7 levels, and presents them as continuous numerical values. This approach reveals subtle differences between samples while overcoming the limitations of traditional methods that rely on integer ratings. Furthermore, for subjective evaluation scores at different vehicle speeds, spline interpolation is used to fit them into continuous curves, which more intuitively reflect the trend of steady-state subjective evaluation values ​​as a function of vehicle speed. In the non-steady-state noise evaluation stage, an interactive slider evaluation system was designed. By using the steady-state evaluation results as the initial value of the slider, the comparability of the subjective evaluation results of steady-state and non-steady-state noise under the same standard is ensured. During sample playback, evaluators continuously drag the slider to adjust the intensity of their auditory perception of non-steady-state noise in real time, thereby forming a time-varying subjective evaluation curve of non-steady-state noise. This method not only accurately captures the dynamic changes of non-steady-state noise but also effectively improves the continuity of the non-steady-state noise evaluation results. Based on this, the consistency of the subjective evaluation curve of non-steady-state noise is checked to ensure the reliability and stability of the evaluation data. To establish a correlation analysis between steady-state and non-steady-state evaluation results, this embodiment also introduces the Dynamic Time Warping (DTW) method. By comparing the matching degree of steady-state and non-steady-state time-varying curves, the overall consistency is quantified. Based on this, the DTW method can reveal the lag or advance effect of non-steady-state auditory perception relative to the steady-state benchmark, thereby effectively capturing the dynamic differences between the two under different operating conditions, and thus forming a unified in-vehicle noise quality evaluation system under all operating conditions.

[0023] Furthermore, the process of evaluating the steady-state noise samples and obtaining the first evaluation result includes: The steady-state noise samples are initially scored using the semantic segmentation method to obtain preliminary scoring results. Based on the preliminary scoring results, the samples are grouped to minimize the difference between the mean of each group and the overall mean. Paired comparisons were performed within each group to obtain data on the relative preference relationships between samples. Calculate the capability value for each sample based on relative preference relationship data; Based on the associated samples, the capability values ​​of cross-group samples are inverted and reconstructed to obtain the global capability values ​​of all samples; The global capability value is mapped to a continuous evaluation interval to obtain the first evaluation result.

[0024] Furthermore, in this embodiment, a semantic segmentation method is first used to initially score the samples. Evaluators subjectively score the steady-state noise samples according to an agitation level scale of 1-7. To ensure the consistency and reliability of the evaluation results, the reliability of the scoring results is tested by calculating the intra-evaluator Pearson correlation coefficient and the inter-evaluator ICC correlation coefficient. Evaluation data with correlation coefficients below 0.6 are removed, and the average score of valid evaluators is retained as the evaluation result of the semantic segmentation method. Secondly, based on the evaluation results obtained by semantic segmentation, the samples are grouped according to the principle of minimizing mean difference, so that the samples within the group are homogeneous and the differences between the groups are reasonable. Within each group, the pairwise comparison method is used to obtain the relative preference relationship between the samples, and the evaluation data with logical inconsistency are eliminated by AA, AB and ABC tests. Then, the pairwise comparison results within the group are used to output the ability value within the group using the Bradley-Terry model. Combined with the inversion reconstruction of related samples, the global ability value of all samples on the continuous scale is calculated. Finally, by using least squares scaling to unify the global capability values ​​to a semantic scale of 1-7, this method improves the accuracy of evaluation results while maintaining the intuitiveness of semantic segmentation. This overcomes the limitations of traditional single methods in terms of evaluation granularity and scale consistency, making the evaluation results of different samples more comparable.

[0025] Furthermore, the process of initially scoring steady-state noise samples based on semantic segmentation includes: The steady-state noise samples are scored according to a preset rating scale to obtain scoring data; Calculate the intra-rater correlation coefficient and the inter-rater correlation coefficient, and remove data with correlation coefficients below a set threshold to obtain valid rating data; The valid scoring data are averaged to obtain preliminary scoring results.

[0026] Furthermore, the process of grouping samples based on the preliminary scoring results includes: Sort the samples in ascending order of the initial score results and calculate the population mean of all samples; The sorted sample sequence is hierarchically divided, and samples are evenly drawn from each level to form evaluation groups. The variance of the mean of each group to the overall mean is calculated. The variance is adjusted to group the population so that the difference between the mean of each group and the population mean is less than the target preset threshold.

[0027] Furthermore, in this embodiment, the steady-state noise samples are sorted in ascending order based on the subjective evaluation scores using a semantic segmentation method, and the overall average score of all samples is calculated. Then, the ordered sample sequence is divided into multiple levels, and the overall samples are allocated to each group according to a preset number of groups. Each group selects samples evenly from different scoring levels, ensuring that the average score of each group is close to the overall average, thereby guaranteeing the rationality and balance of the grouping results.

[0028] Furthermore, the process of performing paired comparison evaluations within each group includes: Within each group, samples are compared pairwise, and a score matrix is ​​generated based on the comparison results; The Bradley-Terry model is constructed based on the score matrix, and the relationship between sample ability values ​​and comparison probabilities is defined. The Bradley-Terry model is solved iteratively by maximum likelihood estimation to obtain the capability value of each sample.

[0029] Furthermore, the process of reconstructing the capability values ​​of cross-group samples based on associated samples includes: Two related samples are selected between adjacent groups as anchor points. The anchor points satisfy the condition that the capability values ​​are distributed on both sides of the group mean and the differences are within a preset interval. Calculate the proportional coefficient and translation adjustment amount based on the capacity values ​​of the anchor points in the two groups before and after; A linear transformation is performed on the capability values ​​of subsequent groups of samples based on the scaling factor and translation adjustment to obtain a global capability value under a unified scale.

[0030] Furthermore, the process of mapping global capability values ​​to continuous evaluation intervals includes: A linear mapping relationship between the global capability value and the preliminary scoring results is established using the least squares method, and the mapping coefficients are obtained. The global capability value is converted to a continuous evaluation interval based on the mapping coefficient to obtain the continuous first evaluation result.

[0031] In the above implementation, this embodiment combines the advantages of semantic segmentation's explicit semantic level rating scale with the relative preference relationship analysis of the pairwise comparison method for steady-state noise evaluation during the steady-state noise evaluation stage. The initial evaluation results of semantic segmentation are used as the basis for grouping, and scientific grouping is achieved by minimizing mean differences, ensuring the homogeneity of samples within a group and the rationality of differences between groups. Based on this, the pairwise comparison method is used to obtain the relative preference relationship between samples, and logically inconsistent data is eliminated through AA, AB, and ABC tests. Then, the Bradley-Terry model outputs the intra-group capability values ​​and cross-group inversion reconstruction of related samples. Finally, the capability values ​​of the noise samples are mapped to the 1-7 level evaluation range of semantic segmentation and presented in continuous numerical form, enabling the evaluation results to reveal subtle differences between samples. This breaks through the limitations of traditional integer levels, making the evaluation results between different samples more comparable.

[0032] Furthermore, the process of evaluating non-steady-state noise samples includes: Construct an interactive slider evaluation system, where the slider position range corresponds to a continuous evaluation interval; During the sample playback, slider position data is continuously recorded to obtain a time-varying evaluation curve; The time axis of the time-varying evaluation curve is mapped to the vehicle speed axis, and the vehicle speed-evaluation relationship curve is generated as the second evaluation result curve.

[0033] Furthermore, this embodiment evaluates non-steady-state noise samples using a non-steady-state in-vehicle noise evaluation system. The non-steady-state in-vehicle noise evaluation system includes: The slider interaction module provides a continuously adjustable evaluation slider, with the slider position range corresponding to the 1-7 levels of annoyance level and their semantic descriptions in the semantic subdivision method. The audio playback and switching module is used to enable sequential playback and free switching of non-steady-state noise samples to ensure the integrity and continuity of the evaluation process. The time synchronization module is used to ensure that the position of the evaluation slider corresponds one-to-one with the sample playback time; The data visualization module is used to display the dynamic feedback curves of evaluators in real time; The data storage module is used to record in real time the changes in the slider position of the evaluators throughout the entire playback process, which facilitates subsequent analysis and export.

[0034] In the above implementation, this embodiment designs an interactive slider evaluation system for the non-steady-state noise evaluation stage, taking into account its dynamic time-varying characteristics. The system determines the initial value of the slider based on the evaluation value of steady-state noise, ensuring the comparability of steady-state and non-steady-state noise evaluation results on a unified scale. During sample playback, evaluators continuously drag the slider to adjust their perceived noise intensity in real time, forming a time-varying subjective evaluation curve. The obtained curve undergoes a consistency check, yielding a non-steady-state noise time-varying evaluation result that reflects the group's perception patterns. Based on the velocity function v(t), the obtained non-steady-state noise time-varying evaluation curve is mapped, uniformly representing the time-varying perception data of non-steady-state noise as a vehicle speed-subjective evaluation relationship curve.

[0035] As an additional implementation, this embodiment also includes verification of subjective evaluation data for non-steady-state noise, the specific process of which includes: The Pearson correlation coefficient between evaluators and the group average curve was calculated. Evaluation data with a correlation coefficient lower than 0.6 were removed, and the evaluation results of valid evaluators were retained. Based on the evaluation results of valid evaluators, mean processing was performed to obtain continuous time-varying subjective evaluation results of non-steady-state noise. The obtained time-varying evaluation curve of non-steady-state noise was mapped based on the velocity function v(t), and the time-varying perception data of non-steady-state noise was uniformly represented as the vehicle speed-subjective evaluation relationship curve, so as to conduct more intuitive comparative analysis in the vehicle speed dimension.

[0036] Furthermore, the process of using the first evaluation result as the initial benchmark for the second evaluation result includes: Obtain the initial vehicle speed of the non-steady-state noise sample; Based on the initial vehicle speed, query the corresponding first evaluation result of steady-state noise; Set the initial position of the slider to the first evaluation result value found in the query, so that the initial state of the non-steady-state evaluation is consistent with the steady-state evaluation in terms of scale.

[0037] Furthermore, the process of performing dynamic correlation analysis on the first evaluation result curve and the second evaluation result curve includes: Construct a distance matrix between the first evaluation result curve and the second evaluation result curve, and calculate the difference in evaluation values ​​at each vehicle speed point; The cumulative cost function of the distance matrix is ​​calculated using a dynamic programming algorithm to obtain the optimal matching path with the minimum total cost. The overall consistency is quantified based on the cumulative distance value of the optimal matching path; Based on the local deviation of the optimal matching path from the baseline diagonal, identify the lag or advance effect of the non-steady-state evaluation relative to the steady-state evaluation.

[0038] Furthermore, the time dynamic warping analysis in this embodiment includes: Using steady-state and non-steady-state evaluation curves as a benchmark, a distance matrix is ​​constructed between the two. This matrix provides an accurate comparison of the subjective evaluation results of steady-state and non-steady-state noise in the time dimension by quantifying the differences between the data at each moment. A dynamic programming approach is employed to process the distance matrix and calculate the optimal matching path, ensuring monotonicity and continuity during the matching process and avoiding the discontinuous matching problem encountered in traditional methods. This process, by optimizing path selection, maximizes the preservation of the dynamic characteristics in both steady-state and non-steady-state noise evaluation curves, enabling precise alignment of subjective evaluation results for both types of noise within the same system. Based on the results of the optimal path, the cumulative distance value is calculated to quantify the overall consistency of the two curves. This value provides a quantitative basis for the similarity between the steady-state and non-steady-state noise evaluation results, thus scientifically verifying the stability of the evaluation results. Furthermore, based on the local deviation of the optimal path, difference information within specific speed ranges is extracted, and the lag effect or advance effect is analyzed and revealed. This analysis not only captures the dynamic differences between non-steady-state and steady-state noise under different operating conditions but also improves the consistency and comparability of the subjective evaluation results of steady-state and non-steady-state noise. In the above implementation, to achieve correlation analysis between steady-state and non-steady-state noise evaluation results, this embodiment breaks through the limitations of traditional isolated analysis and proposes a complete correlation analysis system. Spline interpolation is performed on the subjective evaluation data obtained from steady-state evaluation to generate continuous vehicle speed-subjective evaluation curves. Based on this, the Dynamic Time Warping (DTW) method is introduced to solve the problem of difficulty in directly comparing two curves due to their different change characteristics. This method effectively overcomes the alignment difficulties caused by inconsistent vehicle speed change rates by constructing a distance matrix between the two curves and using dynamic programming to search for the optimal matching path. Based on this optimal path, not only is a normalized global cumulative distance value output to quantify the overall consistency between steady-state and non-steady-state perception, but also the local deviation features of the path relative to the diagonal are extracted, thereby accurately explaining the perception lag or advance effect within a specific vehicle speed range.

[0039] As a preferred implementation method, the subjective evaluation method for in-vehicle noise quality under all operating conditions proposed in this embodiment includes the overall process of collecting in-vehicle noise samples and constructing a database, subjective evaluation of steady-state noise, subjective evaluation of non-steady-state noise, and correlation analysis of steady-state and non-steady-state noise evaluation results. The method includes: In the noise sample construction stage, this embodiment collects in-vehicle noise signals of different vehicles under full working conditions through microphones, and filters and trims the collected signals, removing invalid or redundant segments and retaining effective audio that can accurately characterize acoustic features, thereby establishing an in-vehicle noise sample dataset under full working conditions. In the steady-state noise evaluation stage, this embodiment combines semantic segmentation and group-based pairwise comparison to conduct subjective evaluation. Semantic segmentation is used to obtain preliminary scores for samples and eliminates inconsistent data through data verification to obtain valid subjective evaluation values. Group-based pairwise comparison groups samples according to the principle of minimizing mean differences, conducts pairwise comparisons within each group, and introduces a logical consistency check. Subsequently, it uses the Bradley-Terry model for modeling and cross-group inversion reconstruction of associated samples to ultimately achieve subjective evaluation results of steady-state noise at a uniform scale, revealing subtle differences between samples, and generating continuous steady-state noise subjective evaluation curves through spline interpolation. In the non-steady-state noise evaluation stage, this embodiment designs an interactive slider evaluation system. This system uses the steady-state noise subjective evaluation value as the initial position of the slider, ensuring consistency between steady-state and non-steady-state noise evaluation results under a unified scale. During sample playback, evaluators continuously drag the slider to adjust their perceived noise intensity in real time, forming a time-varying subjective evaluation curve. The curve is then subjected to consistency checks, resulting in a non-steady-state noise time-varying evaluation result that reflects the group's auditory perception patterns. Based on the velocity function v(t), the obtained non-steady-state noise time-varying evaluation curve is mapped, uniformly representing the time-varying perception data of non-steady-state noise as a vehicle speed-subjective evaluation relationship curve. In the correlation analysis phase, this embodiment introduces the Dynamic Time Warping (DTW) method, using the steady-state evaluation curve and the non-steady-state time-varying curve as inputs to construct a distance matrix and calculate the optimal matching path using dynamic programming. Based on this, the global cumulative distance value is obtained to characterize the overall consistency between the two, while local path features are extracted to reveal the lag effect or advance effect in specific intervals, ultimately establishing a systematic correspondence between the subjective evaluation values ​​of steady-state and non-steady-state noise.

[0040] The specific implementation process is as follows: The first step involves collecting in-vehicle noise signals at different measurement points under all operating conditions, including steady-state constant speed, non-steady-state acceleration, and non-steady-state deceleration. The collected raw signals are then selected, clipped, and preprocessed to retain representative audio that accurately characterizes the noise. This process ultimately forms an in-vehicle noise sample database under all operating conditions.

[0041] The second step involves organizing evaluators to conduct preliminary scoring of the steady-state noise samples using semantic segmentation. Annoyance levels are set on a scale of 1-7, with 1 representing almost no annoyance and 7 representing extreme annoyance. The noise samples from the first step are then randomly shuffled and played twice in different playback orders. The evaluation matrix is ​​represented as follows: in, This represents the score given by the j-th evaluator to the i-th sample during the k-th playback; N is the number of noise samples; and M is the number of evaluators.

[0042] 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: in, Give the Pearson correlation coefficient between the two evaluation results for the i-th evaluator; M is the number of evaluators; and These are the two evaluation scores given by the i-th evaluator for the j-th noise sample; and is the average score of the i-th evaluator for all noise samples in two evaluations.

[0043] Based on the obtained Pearson correlation coefficient, if the correlation coefficient between two subjective evaluations by an evaluator is less than 0.6, the evaluator is removed, and other evaluators are retained for further analysis.

[0044] In addition, to ensure consistency of evaluation standards among different evaluators, the ICC correlation coefficient is introduced to test inter-evaluator reliability. The formula is as follows: in, Mean square between groups; The mean square within the group; M is the number of evaluators.

[0045] Based on the obtained ICC correlation coefficient, evaluators with a correlation below 0.6 were removed.

[0046] After passing the consistency test described above, qualified evaluators were selected. Based on this, the valid evaluation results for each noise sample were statistically averaged to obtain the final subjective score for that sample.

[0047] The third step is to sort the initial scores of the steady-state noise samples obtained by the semantic segmentation method in ascending order: in This represents the average score of the Nth sample after sorting.

[0048] The formula for calculating the population mean for all samples is: in This represents the population mean. The number of samples; Let be the evaluation score of the i-th noise sample.

[0049] To ensure that each group draws samples from different levels and avoids extreme differences in the mean, the sorted samples are divided into several levels according to the score intervals, thus: Where L is the number of layers; For the first Number of samples in the layer The sample is divided into G groups, and samples are drawn uniformly from different layers for each group, such that the group mean tends to the global mean. The mean of each group is defined as: Where, μ g The average subjective score for group g; This represents the number of samples within the group. Let be the evaluation score of the i-th noise sample.

[0050] Furthermore, to quantitatively measure the effect of grouping, the mean and variance are used: Where G is the total number of groups; It characterizes the degree of dispersion of each group's mean from the population mean; if A larger value indicates a greater difference in the mean between groups, requiring further grouping adjustments; if... The smaller the value, the more concentrated the group means are. The grouping effect is ideal in the vicinity.

[0051] The fourth step, after completing the uniform grouping of the samples, is to evaluate the sound quality within each group using a pairwise comparison method based on the Bradley-Terry model. The evaluation is first performed on the first group of samples: In the pairwise comparison method, for any two samples within the same group... , Let's set some scoring criteria: If sample i is better than j, then i gets 1 point; If sample i is less than sample j, then i gets 0 points; That is, define the pairwise comparison score function p(i,j): in, This indicates that the evaluators believe the sample The level of irritability was higher than that of the sample. For the first group of samples, each sample has an ability value. The Bradley-Terry model is then defined as follows: Where p(i>j) is the sample In relation to the sample The probability is obtained from the comparison; , For the sample and samples Ability stats.

[0052] The maximum likelihood function is: The capability values ​​of the first set of samples are obtained through iterative optimization.

[0053] After the first set of scores is completed, two related samples are selected based on the correlation screening strategy and added to the second set as anchor points: Among them, S1 1 and S1 2 V represents the ability values ​​of two reference samples; V represents the average ability values ​​of all samples in group 1; V i and V j These are the capability values ​​of sample i and sample j in group 1, respectively; ε is the adjustment coefficient, which is usually 2 to 5 units.

[0054] To ensure the continuity and comparability of evaluation results across different groups on a uniform scale, a related sample mechanism is introduced during the grouping calculation process. Specifically, two samples that meet the related criteria are introduced into adjacent groups, and this operation is maintained in all subsequent groups, thereby establishing a stable reference relationship between groups and achieving consistency in the cross-group scoring system. Figure 2 As shown.

[0055] Fifth, after the evaluation, to ensure data consistency, a misjudgment analysis is performed on the evaluation results: Analysis of AA Misjudgments: Whether the same evaluator's comparison results for two identical samples are consistent.

[0056] AB misjudgment analysis: Whether the same evaluator's comparison results of two samples in different orders are consistent.

[0057] ABC Misjudgment Analysis: When an evaluator is evaluating a set of noisy samples, the scores given to three samples are contradictory. Specifically, evaluation result A is greater than B, B is greater than C, but C is greater than A.

[0058] To ensure the scientific validity and reliability of the evaluation results using the grouping pairwise comparison method, a weighted consistency coefficient is adopted to comprehensively test for various types of misjudgments. Where Ei is the possible number of misclassifications of type i, and Ci is the actual misclassification rate of type i.

[0059] The weighted consistency coefficient is: When the consistency weight factor of the evaluation results is greater than 0.7, the evaluation results are considered valid, and the portion with the consistency weight factor in the last 10% is removed.

[0060] Step 6: Perform inversion reconstruction based on the scores of all groups except the first group to ensure that the scores of all groups are compared on a uniform scale. The inversion reconstruction formula is as follows: in, V represents the reconstructed capability value of the i-th sample in the j-th group; 1j and V 2j V represents the normalized value of the original evaluation value of two reference samples in group j; ij is the normalized value of the original evaluation value of sample i in group j; k is the proportional coefficient; β is the translation adjustment amount.

[0061] The proportionality coefficient k and the translation adjustment β are obtained using the formula: Where V and V' are the capability values ​​of the associated sample in the preceding and following groups.

[0062] Step 7: After completing the grouping pairwise comparison method and inversion reconstruction, obtain the capability value of each noise sample. To achieve consistency with the semantic segmentation method's evaluation levels and improve the accuracy of subjective evaluation, the least squares method is used to map the capability values ​​to the 1-7 level evaluation range of the semantic segmentation method. The formula is as follows: in, For the evaluation results; Here, a represents the capability value of the noise sample; a and b are the coefficients of the fitting function obtained by the least squares method, used to minimize: Where Li is the semantic average subjective score of each noise sample, and the subjective evaluation score of the steady-state noise sample is as follows: Figure 3 As shown.

[0063] The eighth step involves continuously recording time-varying non-steady-state noise signals to achieve this. An interactive subjective evaluation system for non-steady-state noise was developed, such as... Figure 4 As shown, the system interface includes an audio playback and switching module, a slider interaction module, a time synchronization module, a data curve visualization module, and a data storage module, among which: The slider interaction module provides a continuously adjustable evaluation slider, with the slider position range corresponding to the 1-7 levels of annoyance level and their semantic descriptions in the semantic subdivision method. The audio playback and switching module is used to enable sequential playback and free switching of non-steady-state noise samples to ensure the integrity and continuity of the evaluation process. The time synchronization module is used to ensure that the position of the evaluation slider corresponds one-to-one with the sample playback time; The data visualization module is used to display the dynamic feedback curves of evaluators in real time; The data storage module is used to record in real time the changes in the slider position of the evaluators throughout the entire playback process, which facilitates subsequent analysis and export.

[0064] Step 9: To ensure scale consistency between the subjective evaluation values ​​of steady-state and non-steady-state noise, the subjective evaluation value obtained from the subjective evaluation of steady-state noise is used as the initial value of the slider. Specifically, if the initial speed of a certain non-steady-state noise sample is V0, the initial position of the slider is set to the subjective score S(V0) of the steady-state evaluation at that speed when the interface starts playing. This setting ensures that the subjective evaluation of non-steady-state noise is consistent with the subjective evaluation result of steady-state noise at the initial moment, and avoids the delay phenomenon in the evaluation of non-steady-state noise, thereby enhancing the comparability between the two.

[0065] In the evaluation of non-steady-state noise samples, the samples are randomly shuffled and played in different sequences. Evaluators provide real-time feedback on their subjective perception levels by dragging a slider. The system uses a fixed sampling rate of 50ms for data recording. This sampling rate is designed to take into account both the temporal resolution of human noise perception and the system's response delay, ensuring accurate capture of the continuous changes in the evaluator's subjective perception. The result is a time-varying subjective evaluation curve.

[0066] Step 10: In the subjective evaluation processing of non-steady-state noise, correlation analysis is performed among evaluators to eliminate evaluation data with inconsistent evaluation standards or deviations from the group trend. Specifically, for the same non-steady-state noise sample, time-varying subjective evaluation curves for all evaluators are obtained. Using the group average curve of this sample as a benchmark, the Pearson correlation coefficient between each evaluator's curve and the group average curve is calculated. The formula is as follows: Where, x jt This represents the evaluation value of the j-th evaluator at time t; This represents the average rating from the evaluators; This represents the evaluation value of the population curve at time t; The mean of the population average curve is represented by T, which represents the number of sampling points corresponding to the non-steady-state sample. This represents the consistency coefficient between the j-th evaluator and the group average curve.

[0067] If the consistency coefficient is less than 0.6, the evaluator is removed. This method can ensure the group consistency and reliability of the subjective evaluation curve of non-steady-state noise.

[0068] Based on this, the subjective evaluation results of effective evaluators are averaged to form a continuous non-steady-state noise subjective evaluation curve, which provides a reliable data basis for subsequent correlation analysis of steady-state and non-steady-state noise subjective evaluation values.

[0069] Step 11: Perform spline interpolation fitting on the subjective evaluation values ​​of the steady-state noise samples at different vehicle speeds to obtain a smooth, continuous steady-state evaluation curve. Where v is the vehicle speed; y i For the vehicle speed point v i Discrete evaluation values ​​under N; i (v) represents the corresponding spline basis function; n represents the number of discrete vehicle speed points; and S(v) represents the steady-state noise evaluation curve.

[0070] The time-varying subjective evaluation curve of non-steady-state noise is obtained by the slider system and is defined as follows: Where t is time, and N(t) is the subjective evaluation value of the non-steady-state noise at time t.

[0071] To compare with the subjective evaluation curve of steady-state noise, the time axis t is mapped to vehicle speed v. Based on the speed variation law under unsteady conditions, a vehicle speed-time function v(t) can be established, such as... Figure 5 As shown, the subjective evaluation curve of non-steady-state noise is re-characterized as follows: in, It is the inverse function of v(t), representing the time point corresponding to reaching a certain vehicle speed v. Therefore, the subjective evaluation result of non-steady-state noise is also expressed as a function of vehicle speed N(v), which facilitates comparative analysis with the subjective evaluation curve of steady-state noise S(v) on the same dimension.

[0072] Step 12: After obtaining the fitting curves for the steady-state noise subjective evaluation values ​​and the curves for the non-steady-state noise group subjective evaluation values, the Dynamic Time Warping (DTW) algorithm is used to establish the relationship between the steady-state and non-steady-state noise subjective evaluation values. The steps are as follows: Construct the distance matrix between the steady-state curve S(v) and the non-steady-state curve N(v): Wherein, S(v i ) and N(v j ) represent the steady-state and unsteady-state noise at vehicle speed v, respectively. i and v j The subjective evaluation value, D(i,j), reflects the difference between the two at the corresponding point.

[0073] The cumulative cost function is calculated using dynamic programming. Where C(i,j) represents the minimum cumulative cost from the starting point to position (i,j), and the monotonicity and continuity of the path are guaranteed during the recursion process.

[0074] The final optimal matching path P* is the path with the minimum total cost, and its cumulative cost is: Where n and m are the number of sampling points for the steady-state noise subjective evaluation curve and the non-steady-state noise subjective evaluation curve, respectively.

[0075] To further quantify the consistency between subjective evaluation results of steady-state and non-steady-state noise, the cumulative distance values ​​are normalized to make them comparable. The normalization formula is as follows: Wherein, K is the number of matching point pairs contained in the optimal planning path P*.

[0076] The normalized cumulative distance DTWnormal represents the average difference in subjective evaluation at each point, and the smaller the value, the higher the consistency between the two.

[0077] Furthermore, by analyzing the local deviation of the optimal path P* relative to the diagonal, the discrepancy between unsteady-state perception and steady-state system can be revealed, namely the perception difference effect. This effect is specifically manifested as follows: The path is close to the diagonal: This indicates that in this speed range, the subjective evaluation trends of unsteady and steady-state noise are consistent, and there is no significant dynamic effect in perception. The path is below the diagonal (i.e., j < i): This indicates that in this interval, the unsteady subjective evaluation values are systematically low. To match the steady-state subjective score S(vi), the algorithm associates the score N(vj) of a lower vehicle speed vj in the unsteady curve, which reveals the hysteresis effect, that is, the subjective auditory perception changes slower than the actual change; The path is below the diagonal (i.e., j > i): This indicates that in this interval, the unsteady subjective evaluation values are systematically high. To match the steady-state subjective score S(vi), the algorithm associates the score N(vj) of a higher vehicle speed vj in the unsteady curve, which reveals the anticipation effect, that is, the subjective auditory perception is faster than the actual change.

[0078] Through the above systematic steps, the subjective evaluation grade and the time-varying curve of the in-vehicle noise samples under all working conditions are finally determined on a unified scale. For steady-state noise, through the combination of semantic segmentation method and grouped paired comparison method, grouping with minimized mean difference, Bradley–Terry model and associated sample inversion reconstruction, the accuracy of the subjective evaluation of steady-state noise is ensured; for unsteady noise, further by building a subjective evaluation system for unsteady noise, introducing a slider real-time feedback mechanism and aligning with the dynamic time warping (DTW) method, the systematic correlation between the subjective evaluation results of steady-state and unsteady noises is achieved. The finally obtained subjective evaluation results can more accurately reflect the subjective auditory perception characteristics of the vehicle's noise under all working conditions, providing a scientific basis and technical support for the subsequent optimization of in-vehicle sound quality.

[0079] Above, it is only a preferred specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in this application should be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A method for subjective evaluation of sound quality of in-vehicle noise in all operating conditions, characterized by, The method comprises the following steps: constructing a full-condition vehicle interior noise sample database, which comprises steady-state noise samples and non-steady-state noise samples; evaluating the steady-state noise samples to obtain a first evaluation result, and mapping the first evaluation result to a continuous evaluation interval to generate a first evaluation result curve; evaluating the non-steady-state noise samples to obtain a time-varying second evaluation result, wherein the second evaluation result takes the first evaluation result as an initial reference to generate a second evaluation result curve; performing dynamic correlation analysis on the first evaluation result curve and the second evaluation result curve, and establishing a corresponding relationship between the subjective evaluation results of the steady-state noise and the non-steady-state noise by calculating the optimal matching path of the two curves.

2. The method of claim 1, wherein the process of evaluating the steady-state noise samples to obtain a first evaluation result comprises: performing preliminary scoring on the steady-state noise samples according to a semantic segmentation method to obtain a preliminary scoring result; grouping the samples according to the preliminary scoring result to minimize the difference between the mean value of each group and the overall mean value; performing pairwise comparison evaluation within each group to obtain relative preference relationship data between the samples; calculating the ability value of each sample according to the relative preference relationship data; reconstructing the ability value of the cross-group samples by inversion according to the associated samples to obtain the global ability value of all samples; mapping the global ability value to the continuous evaluation interval to obtain the first evaluation result.

3. The method of claim 2, wherein the process of performing preliminary scoring on the steady-state noise samples according to a semantic segmentation method comprises: scoring the steady-state noise samples according to a pre-set rating scale to obtain scoring data; calculating the intra-evaluator correlation coefficient and the inter-evaluator correlation coefficient, and removing data with a correlation coefficient lower than a set threshold to obtain valid scoring data; performing mean value processing on the valid scoring data to obtain the preliminary scoring result.

4. The method of claim 2, wherein the process of grouping the samples according to the preliminary scoring result comprises: arranging the samples in ascending order according to the preliminary scoring result, and calculating the overall mean value of all samples; hierarchically dividing the sorted sample sequence, and uniformly extracting samples from each hierarchy to form evaluation groups, and calculating the variance of the mean value of each group and the overall mean value; adjusting the grouping according to the variance to make the difference between the mean value of each group and the overall mean value less than a target pre-set threshold.

5. The method of claim 2, wherein the process of performing pairwise comparison evaluation within each group comprises: comparing the samples in pairs within each group, and generating a score matrix according to the comparison results; constructing a Bradley-Terry model according to the score matrix to define the relationship between the sample ability value and the comparison probability; iteratively solving the Bradley-Terry model by maximum likelihood estimation to obtain the ability value of each sample.

6. The method of claim 2, wherein the process of reconstructing the ability value of the cross-group samples by inversion according to the associated samples comprises: ​ ​ ​ ​ ​ Two associated samples between adjacent groups are selected as anchor points, which satisfy that the ability values are distributed on both sides of the group mean and the difference is located in a preset interval; A proportion coefficient and a translation adjustment amount are calculated according to the ability values of the anchor points in the two groups before and after; The ability values of subsequent groups of samples are linearly transformed according to the proportion coefficient and the translation adjustment amount to obtain global ability values under a unified scale.

7. The method of claim 2, wherein, The process of mapping the global ability values to the continuous evaluation interval includes: A linear mapping relationship between the global ability values and the preliminary scoring results is established according to the least squares method to obtain a mapping coefficient; The global ability values are converted to the continuous evaluation interval according to the mapping coefficient to obtain a continuous first evaluation result.

8. The method of claim 1, wherein, The process of evaluating the non-steady-state noise samples includes: An interactive slider evaluation system is constructed, wherein the slider position range corresponds to the continuous evaluation interval; Slider position data is continuously recorded during sample playback to obtain a time-varying evaluation curve; The time axis of the time-varying evaluation curve is mapped to a vehicle speed axis to generate a vehicle speed-evaluation relationship curve as the second evaluation result curve.

9. The method of claim 1, wherein, The process of taking the first evaluation result as the initial reference for the second evaluation result includes: The starting vehicle speed of the non-steady-state noise sample is obtained; The corresponding steady-state noise first evaluation result is queried according to the starting vehicle speed; The initial position of the slider is set to the queried first evaluation result value, so that the initial state of the non-steady-state evaluation is consistent in scale with the steady-state evaluation.

10. The method of claim 1, wherein, The process of dynamically correlating the first evaluation result curve and the second evaluation result curve includes: A distance matrix between the first evaluation result curve and the second evaluation result curve is constructed to calculate the evaluation value difference at each vehicle speed point; An accumulated cost function of the distance matrix is calculated according to the dynamic programming algorithm to obtain an optimal matching path with the minimum total cost; The overall consistency degree is quantified according to the accumulated distance value of the optimal matching path; The lag effect or the advance effect of the non-steady-state evaluation relative to the steady-state evaluation is identified according to the local deviation amount of the optimal matching path relative to the reference diagonal line.

Images