Full vehicle modal identification is a major challenge for both experimental and simulated modal results. A global modal is usually masked by nearby local modes, so that even well-experienced engineers have difficulty to identify vehicle modes efficiently. Besides different vehicle configurations e.g. SUV, MPV and hatchback can make the challenge even greater, since the same modal for them will present different characteristics.
This paper proposes a deep neural network for vehicle modal identification. This method takes advantage of the deep learning method which has achieved outstanding performance in language translation, computer vision and image processing. It also has the potentiality to improve modal identification efficiency.
In general commercial neural network applications, a large number of data is available for training to achieve a robust output. However, the training data for a vehicle modal identification is always limited, as a result, training the network becomes difficult. This paper proposes an architecture of a multi-layer neural network. Each layer of the neural network has been optimized to gain a robust output with limited training data.
In the end this paper focuses on modal identification from FEA analysis. Several full vehicle FEA modes have been established, and the modals from 0~50Hz have been calculated using finite element method. The modal identification shows encouraging results in which the proposed neural network is able to identify the global mode at certain degree of confidence.