Steering Feedback Torque Prediction Based on Sequence-to-Sequence Network With Switcher-Assisted Training Algorithm
Steering Feedback Torque Prediction Based on Sequence-to-Sequence Network With Switcher-Assisted Training Algorithm
复制标题
DOI：
10.1109/tii.2023.3329650
复制
发表时间：
2024-03
影响因子：
12.3
通讯作者：
Yicai Liu;Guowang Zhang;Jian Li;Changyao Huang;Xiang-yu Wang;Liang Li;Xun Zhao
中科院分区：
文献类型：
--
作者：
Yicai Liu;Guowang Zhang;Jian Li;Changyao Huang;Xiang-yu Wang;Liang Li;Xun Zhao
线控转向(SBW)系统以其安全性、简单性和灵活性等特点被公认为智能汽车的未来发展方向。然而，消除机械联动需要提供人工转向反馈力矩(SFT)，这对潜在驾驶员操纵车辆至关重要。为了提高操纵感，本文扩展了SFT的设计范围，提出了一种基于S2S(Sequence-to-Sequence)网络和开关辅助(SA)训练算法的SFT预测方案。首先建立了电动助力转向系统(EPS)和电动助力转向系统(SBW)的模型，分析了系统的输入特性。然后提出了具有门控递归单元(GRU)的S2S网络，其中编码方案结合了压缩和激励(SE)操作来实现自适应特征重新校准。随后，基于多任务学习(MTL)原理，提出了以在线训练为主、离线训练为辅的SA算法来缓解暴露偏差。任务之间的权重系数使用名为Switcher的辅助网络进行优化，便于从MTL逐步过渡到单个主任务，从而避免了复杂的调整过程。验证结果表明，该方法在估计和预测方面优于现有的方法。通过消融实验进一步验证了SE块和SA算法的有效性。最后进行了SFT结构仿真，包括目标预测和力矩跟踪，验证了变长预测能够适应不同的条件，提高了跟踪性能。
The steer-by-wire (SbW) system has gained recognition as the future of intelligent vehicles due to its attributes, such as safety, simplification, and flexibility. However, the elimination of mechanical linkage necessitates the provision of artificial steering feedback torque (SFT), which is crucial for the potential driver to manipulate the vehicle. To enhance the steering feel, this article extends the SFT design range and proposes an SFT prediction scheme based on the sequence-to-sequence (S2S) network with the switcher-assisted (SA) training algorithm. The models of the electric power steering (EPS) and SbW are first established to analyze the input features. The S2S network with gated recurrent units (GRU) is then presented, where the encoder scheme incorporates the squeeze-and-excitation (SE) operation to achieve adaptive feature recalibration. Subsequently, the SA algorithm is proposed to alleviate the exposure bias based on the principle of multitask learning (MTL), wherein online training is regarded as the main task and offline training is treated as the auxiliary task. The weighting coefficients between tasks are optimized using an assisted network named switcher, facilitating a gradual transition from MTL to the single main task, thereby avoiding complex tuning processes. The validation results indicate the proposed scheme outperforms existing methods regarding estimation and prediction. The ablation experiments are further conducted to illustrate the effectiveness of SE blocks and the SA algorithm. Finally, the SFT construction simulation, involving target prediction and torque tracking, is conducted, validating that variable-length prediction can adapt to various conditions and improve tracking performance.