Deep Learning Vehicle NVH Prediction Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for predicting vehicle noise, vibration, and harshness (NVH) performance, such as the finite element method and transfer path analysis, are inaccurate in complex systems and fail to reflect non-linear characteristics, especially in varying driving conditions, leading to unreliable performance evaluation.
Innovation Solution
A deep learning-based method using an artificial neural network to establish a vehicle NVH system model, which preprocesses and learns driving data to form correlation functions between noise and vibration sources, allowing for efficient prediction and visualization of NVH performance without requiring prior characterization of transfer paths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods such as finite element method or transfer path analysis are used to predict vehicle NVH performance, then the prediction is accurate in simple systems, but the prediction accuracy deteriorates in complex systems with various bushes, welding, and assembling processes
Solution Approach 1:
The patent replaces conventional mechanical analysis methods (FEM, TPA) with a data-driven deep learning model. The neural network learns nonlinear relationships between noise sources and vehicle responses directly from measured data, eliminating the need for complex mechanical modeling of bushes, welds, and assemblies. This substitution enables accurate prediction in complex systems where traditional mechanical methods fail.
Solution Approach 2:
The patent transforms the NVH prediction problem from a mechanical parameter-based approach to a data-based parameter approach. Instead of using physical parameters like stiffness, damping, and geometry in FEM models, the invention uses measured input-output data pairs to train a neural network that captures system behavior through learned parameters, achieving better accuracy in complex assembled systems.
2Ease of operation
If conventional methods are used for NVH prediction, then the analysis is straightforward in simple systems, but it becomes difficult to accurately reflect non-linear characteristics in varying driving situations
Solution Approach 1:
The patent implements a dynamic prediction approach where the deep learning model is trained on data from multiple driving situations and conditions. The model adapts to varying operating conditions by learning nonlinear relationships from diverse training data, enabling reliable performance evaluation across different driving scenarios rather than being limited to single-condition analysis.
Solution Approach 2:
The neural network model automatically captures nonlinear characteristics through self-learning from measured data without requiring manual intervention to identify or model non-linear behaviors. The system performs self-characterization of transfer paths and nonlinear relationships, eliminating the need for complex manual analysis while maintaining reliability across varying conditions.
3Reliability
If deep learning-based method is used to predict vehicle NVH performance, then the model can reflect non-linear driving characteristics and work well in complex systems, but it requires significant data processing and computational resources
Solution Approach 1:
The patent performs preliminary data processing and feature extraction during the offline training phase. The neural network model is pre-trained on comprehensive driving data to capture nonlinear characteristics and system behavior. Once trained, the model can perform rapid predictions with minimal computational resources during actual NVH evaluation, as the heavy computational work has already been completed during model training.
Data Source
AI summary
A method for predicting performance of a vehicle NVH system based on deep learning is provided. The method includes preprocessing learning data collected for each channel associated with noise and vibration while driving, learning a model forming a correlation function between multiple inputs and multiple outputs corresponding to the preprocessed learning data using an artificial neural network, and predicting performance using a vehicle NVH system model formed through the learned model.