Modern control synthesis assumes that observability of a dynamic system is satisfied. However, in particular, observability may not be met due to manufacturing cost. To cope with this challenging dilemma, we consider training a neural network with a large dataset that contains prior information of a given nonlinear vehicle dynamics system. In this paper, we proposed and designed a long short-term memory (LSTM)-based neural network to estimate the velocity and position states of a full car’s suspension system using only online data streams from cheap inertial sensory measurements. In the training stage, we collect the input and output data that we want to reconstruct using a nonlinear full car model simulation. An LSTM-based neural network is subsequently trained with the collected data, and it reconstructs the velocity and position states only from the acceleration information inputs in the production stage. Finally, to further enhance the performance via Bayesian filtering, the neural network’s outputs are projected into the vehicle dynamics using the extended Kalman filter. To demonstrate the effectiveness of our approach, we compare the performance of the proposed method against a conventional kinematic Kalman filter. The results show that the proposed method recorded 148.8 times smaller value in terms of the mean squared error (MSE) than that of the benchmark. Furthermore, we analyzed the sensitivity of the proposed neural network with or without gyroscope sensors in the full car model. Then, an illustrative control example for implementing the proposed method is presented. Our approach shows that with only noisy accelerometer and gyroscope sensors, we can successfully reconstruct nearly perfect states in a full car’s nonlinear dynamic system using a well-trained LSTM-based neural network combined with the extended Kalman filter.

现代控制综合假设满足动态系统的可观察性。然而，特别是，由于制造成本，可能无法满足可观察性。为了应对这一具有挑战性的困境，我们考虑使用包含给定非线性车辆动力学系统的先验信息的大型数据集来训练神经网络。在本文中，我们提出并设计了一种基于长短期记忆 (LSTM) 的神经网络，以仅使用来自廉价惯性感官测量的在线数据流来估计整车悬架系统的速度和位置状态。在训练阶段，我们收集要使用非线性完整汽车模型模拟重建的输入和输出数据。随后使用收集的数据训练基于 LSTM 的神经网络，它仅根据生产阶段的加速度信息输入来重建速度和位置状态。最后，为了通过贝叶斯滤波进一步提高性能，使用扩展卡尔曼滤波器将神经网络的输出投影到车辆动力学中。为了证明我们的方法的有效性，我们将所提出的方法的性能与传统的运动卡尔曼滤波器进行了比较。结果表明，所提出的方法记录的均方误差 (MSE) 比基准值小 148.8 倍。此外，我们分析了在完整汽车模型中使用或不使用陀螺仪传感器的拟议神经网络的灵敏度。然后，介绍了用于实现所提出方法的说明性控制示例。