Bayesian NVH metamodels to assess interior cabin noise using measurement databases

V. Prakash,O. Sauvage,J. Antoni,L. Gagliardini

DOI: https://doi.org/10.48550/arXiv.2207.02120

2022-06-13

Abstract:In recent years, a great emphasis has been put on engineering the acoustic signature of vehicles that represents the overall comfort level for passengers. Due to highly uncertain behavior of production cars, probabilistic metamodels or surrogates can be useful to estimate the NVH dispersion and assess different NVH risks. These metamodels follow physical behaviors and shall aid as a design space exploration tool during the early stage design process to support the NVH optimization. The measurement databases constitute different noise paths such as aerodynamic noise (wind-tunnel test), tire-pavement interaction noise (rolling noise), and noise due to electric motors (whining noise). This research work proposes a global NVH metamodeling technique for broadband noises such as aerodynamic and rolling noises exploiting the Bayesian framework that takes into account the prior (domain-expert) knowledge about complex physical mechanisms. Generalized additive models (GAMs) with polynomials and Gaussian basis functions are used to model the dependency of sound pressure level (SPL) on predictor variables. Moreover, parametric bootstrap algorithm based on data-generating mechanism using the point estimates is used to estimate the dispersion in unknown parameters. Probabilistic modelling is carried out using an open-source library PyMC3 that utilizes No-U-Turn sampler (NUTS) and the developed models are validated using Cross-Validation technique.

Applications,Machine Learning

What problem does this paper attempt to address?

The problems that this paper attempts to solve are: How to use the measurement database and Generalized Additive Models (GAMs) in the Bayesian framework to construct a global NVH (Noise, Vibration and Harshness) metamodel in the early stage of automobile development, in order to evaluate the in - vehicle noise and quantify the uncertainty. Specifically, the paper aims to solve the following problems:

1. **In - vehicle noise evaluation**: By establishing a metamodel based on measurement data, quickly and accurately predict the in - vehicle noise levels under different working conditions (such as vehicle speed, wheel torque, etc.).
2. **Uncertainty quantification**: Consider the uncertainties brought by factors such as manufacturing tolerances, natural variations in material properties, and physical test conditions, and use the Bayesian method to model and quantify these uncertainties.
3. **Design space exploration**: Provide an effective tool for designers to evaluate different NVH risks in the early design stage and support NVH optimization.

### Main contributions of the paper

- **Application of the Bayesian framework**: Introduce the Bayesian method to handle uncertainties and prior knowledge, so that the model can not only give point estimates but also provide probability distributions, thereby better evaluating the risks of design choices.
- **Generalized Additive Models (GAMs)**: Use GAMs to model the dependence relationship between sound pressure level (SPL) and predictor variables, and combine polynomial and Gaussian basis functions to improve the flexibility and accuracy of the model.
- **Parametric bootstrap algorithm**: The parametric bootstrap algorithm based on the data generation mechanism is used to estimate the dispersion of unknown parameters, further enhancing the robustness of the model.
- **Use of open - source libraries**: Use open - source libraries such as PyMC3 to implement Bayesian modeling, and use the No - U - Turn Sampler (NUTS) for efficient parameter estimation.
- **Cross - validation technique**: Verify the effectiveness of the developed model through the K - fold cross - validation technique to ensure that the model has good generalization ability.

### Formula display

1. **General form of Generalized Linear Model (GLM)**:
\[
p(y_i|x_{1i},x_{2i},\ldots,x_{npi})=\exp\left(\frac{y_i\theta_i - b(\theta_i)}{\gamma}+c(y_i,\gamma_i)\right)
\]

2. **Expression of Generalized Additive Model (GAM)**:
\[
E[y_i]=\mu_i=\alpha + f_1(x_{1i})+f_2(x_{2i})+f_3(x_{3i},x_{4i})+\cdots
\]
where \( f_j \) represents the smoothing function of the predictor variable.

3. **Polynomial basis function model**:
\[
f_2(x_2)=\sum_{k = 0}^n\beta_{2k}x_2^k
\]

4. **Gaussian basis function model**:
\[
f_2(x_2)=\sum_{k = 0}^n a_k\exp\left(-\frac{(x_2 - b_k)^2}{c_k^2}\right)
\]

5. **Bayesian formula**:
\[
p(\theta|y)=\frac{f(y|\theta)p(\theta)}{p(y)}
\]
where \( f(y|\theta)=L(\theta|y) \) is the likelihood function, and \( p(y)=\int_\theta f(y|\theta)p(\theta)d\theta \) is the marginal distribution.

6. **R² calculation formula**:
\[
R^2 = 1-\frac{\sum_{i = 1}^n(y_i-\hat{y}_i)^2}{\sum_{i = 1}^n(y_i-\bar{y})^2}
\]

Through the solutions to these problems, the paper provides a more efficient and reliable NVH evaluation method for the automotive industry, which helps to improve the overall comfort of vehicles.