بهبود قابلیت اطمینان شاخص کیفیت صدا برای سیستم تهویه مطبوع خودرو با استفاده از مدل رگرسیون و شبکه عصبی

The reduction of vehicle interior noise has long been the main interest of noise and vibration harshness (NVH) engineers. A driver’s perception of vehicle noise is largely affected by psychoacoustic noise characteristics and SPL. Among the various types of vehicle interior noise, the sound of the heating, ventilation, and air conditioning (HVAC) systems is a source of distraction for drivers. HVAC noise is not as loud as the overall noise level; however, it affects a driver’s subjective perception and may lead to feelings of nervousness or annoyance. Therefore, vehicle engineers work not only to reduce noise, but also to improve sound quality. In this paper, HVAC noise samples were taken from many types of vehicles. Objective and subjective sound quality (SQ) evaluations were obtained, simple and multiple regression models were generated, and these were used with the Semantic Differential Method (SDM) to determine what characteristics trigger a “pleasant” response from listeners. The regression analysis produced diagnostic statistics and regression estimates. In addition, neural network (NN) models were created using three objective numerical inputs (loudness, sharpness, and roughness) of the SQ metrics and one subjective output (“pleasant”). The NN model was used primarily because human perceptions are very complex and often hard to estimate. The estimation models were compared via correlations between SQ output indices and hearing test results. Results demonstrated that the NN model is most highly correlated with SQ indices, which led to determination of suggested methods for SQ metrics prediction.

In the automobile industry, the end product has long been conceptualized as a simple machine that is evaluated on its performance alone. However, customers prefer comfortable surroundings in their vehicles, especially since the automobile has become a cultural object in which people spend much of their time [1] and [2]. For this reason, many previous studies have dealt with noise level reduction. Many ongoing studies aim to meet the needs of customers in ensuring that the noise is pleasant and as soft as possible. It has become common for drivers to listen to music or make telephone calls using hands-free systems or their vehicle’s speakers. Some people also upgrade their vehicle’s audio equipment, which may increase interior noise levels. Therefore, a vehicle’s interior noise level must be considered in vehicle development. Noise transmitted from exterior sources such as the engine, road, tires, and wind, has been decreased dramatically in recent years. Noise that comes from the interior has become more significant to driver comfort. The major contributor to this interior noise is the HVAC system, which is the focus of this report [3]. People hear sound in an emotional and subjective way and thus it is difficult to express it numerically. It is clear that a new measure is needed, in addition to or in place of current objective measures of noise, such as dB(A). HVAC noise is not loud compared to the general interior noise level, but it has been found to have subjective effects on the driver’s emotions. In a previous study, HVAC noise was shown to be a combination of the radiated noise caused by driving, noise from the blower motor, and vibrations of the vehicle body [4]. Thus, noise caused by driving is becoming more important among the factors that influence the SQ of the HVAC system. Improvement of interior SQ sources has also become necessary to realize “Brand Sound,” which expresses the particular characteristics and images of an automotive company. In one study, a calculation model for hearing elements was established to improve SQ, and many other studies have striven to construct a model that is more attuned to human hearing by comparing hearing models [5] and [6]. Hearing models have also been created using regression and neural network models, but studies that apply both models simultaneously have so far been inadequate. A regression model has the advantage that its estimations are easy to explain, so it is used in a wide variety of applications. However, it also has several disadvantages, especially its assumptions of homoscedasticity, independence, and normality, and its problems with multicollinearity and sensitivity to specific values. Additionally, there is also uncertainty in the selection of the optimum regression model, and this can create difficulties for the qualitative evaluation of HVAC vehicle noise characteristics. Neural network models can be used for analysis even if the input variables are imperfect or the range of fluctuation is wide. Additionally, limited or irregular data can be used because the model reduces error through recursive learning. The ability to estimate subjective SQ evaluations accurately and effectively using the neural network model with imported HVAC SQ characteristics and then comparing them to the regression model would be a landmark in the construction of a robust SQ index. Therefore, in this study, we perform SQ evaluations of the HVAC noise of various vehicles using both objective and subjective methods, and we select the SQ metrics with highly correlated objective and subjective evaluations. Then we focus on prediction and comparison of subjective SQ characteristics using regression and neural network models.

This study was performed based on the theory that objective SQ metrics and subjective evaluations of HVAC noise would be correlated. Regression analysis and neural network models were compared as methods for predicting positive sound evaluations. The SQ indices for a “pleasant” result were expressed in linear Eqs. (1) and (2), obtained via regression analysis. The neural network index used the input variables of the objective SQ metrics and the output result “pleasant” to “learn” the relationships between them. By examining the correlations between the objective SQ metrics and the subjective evaluations through the neural network, a positive reaction prediction index was constructed that had a higher correlation (98.9%) than either regression model. If these three models were used in a complementary manner, the HVAC system SQ value might be predicted even more effectively. This should be helpful in computing a useful and meaningful SQ grade. The following study limitations should be noted in considering these results. The SQ indices suggested in this study cannot be generalized widely because the number and types of vehicles were limited, and the number of participants in the hearing test was small. Thus, in order to achieve more reliable and generalizable results, this experiment should be replicated using larger samples of vehicles and human participants.