The high functional performance exhibited by modern applications is very often established by an aggregation of various intricate mechanical mechanisms, providing the required motion dynamics to the overall system. Above all, the mechanism’s behavior should be reliable for a wide range of operating conditions to assure at all times appropriate functioning of the entire application. In particular, cam-follower mechanisms, which translate a rotational movement into a linear displacement, are plagued by the high dynamics induced by the reciprocating motions. For specific operating conditions, the follower tends to detach from the cam perimeter, resulting in harmful bouncing behavior. This paper presents the use of recurrent neural networks to estimate the follower jump trajectory, based on cam rotation measurements, for a wide range of operating conditions and system modifications. Although these data-driven models are typically known to learn intricate patterns directly from raw data, enhanced prediction performances are observed when providing physics-inspired features to the model. The effect is especially more pronounced when learning from a small amount of data or from datasets for which the data are not uniformly distributed along the parameter space. In addition, this paper presents the use of an additive feature attribution method to quantify the contribution of features in multivariate timeseries on the prediction output of recurrent neural network models. Hence, we show that, by means of the Shapley additive explanation (SHAP) values, the model prioritizes the incorporation of physics-inspired features, explaining the improved generalization capabilities of the prediction model. In general, these presented results indicate the potential to incorporate physics-inspired expert knowledge into various other prediction models, enabling advanced methodologies to monitor inconvenient phenomena in mechanical systems.

现代应用所表现出的高功能性能通常是通过各种复杂机械机构的聚合来建立的，从而为整个系统提供所需的运动动力学。最重要的是，该机构的行为在广泛的操作条件下应该是可靠的，以确保整个应用程序在任何时候都能正常运行。特别是，将旋转运动转化为线性位移的凸轮从动机构受到往复运动引起的高动态性的困扰。对于特定的操作条件，从动件往往会从凸轮周边脱离，从而导致有害的弹跳行为。本文介绍了使用递归神经网络来估计跟随者跳跃轨迹，基于凸轮旋转测量，适用于各种操作条件和系统修改。虽然这些数据驱动的模型通常会直接从原始数据中学习复杂的模式，但在为模型提供受物理启发的特征时，可以观察到增强的预测性能。当从少量数据或从数据沿参数空间不均匀分布的数据集中学习时，效果尤其明显。此外，本文提出了使用加性特征归因方法来量化多元时间序列中特征对循环神经网络模型预测输出的贡献。因此，我们表明，通过 Shapley 加性解释 (SHAP) 值，该模型优先考虑物理启发的特征，解释了预测模型改进的泛化能力。总的来说，这些呈现的结果表明将受物理学启发的专家知识融入各种其他预测模型的潜力，使先进的方法能够监测机械系统中的不便现象。