Search Results

Two methods have been developed for real time identification and classification of the roughness pattern of road surfaces using the neural network. These methods are directly available both for semi-active and active vibration controls of cars. Accelerations of the rear wheel axis under the suspension are used as the input data for real time identification. The neural network which has acquired the informations of the seven typical roughness patterns is used for real time classification of actual road surfaces during driving. Validity and usefulness of these methods are verified by simulation.

Two kinds of active control systems, using neural networks (NN), are presented for realizing optimal driving motion of four wheel steer (4WS) cars. The first system is based on the assumption that the car is simplified as a linear two wheel bycycle model, and that the friction force between tire and road surface is represented by Fiala's nonlinear model. The nonlinear relation between the slip angle of tire and the cornering force is expressed with NN. A model-following type control strategy is adopted in the first system, with both the feedforward and feedback gains for the control of the rear wheel steering angle adaptively determined with NN according to change of front wheel steering angle. The second system is based on the assumption that both the dynamical characteristics of the car and the tire friction force are nonlinear. The nonlinear dynamical characteristics of the car and the friction force are identified with NN, using the measured data of an actual car.

For commercial-vehicle Original Equipment Manufacturers (OEMs), predictive maintenance has drawn attention for the benefits of money saving and increased road safety. Data-driven models have been widely explored and implemented as predictive maintenance solutions. However, the working scenarios for different commercial-vehicles vary a lot, which makes it difficult to build a universal model suitable for all the cases. In this paper, we propose a Recurrent Neural Adaptive Processes (RNAPs) network to adapt to different scenarios by modeling the uncertain at the same time. The ensemble network combines the traits of neural processes, recurrent neural network and meta learning together. Neural processes consider the context information to calculate the uncertainty and improve the prediction results. Meta-learning works well when dealing with few-shot multi-tasks learning, and recurrent networks are utilized as the encoder of the proposed model to process time-series data.