Abstract

Machine faults and systematic failures are resulted from manufacturing process deterioration. With early recognition of patterns closely related to process deterioration, e.g., trends, preventative maintenance can be conducted to avoid severe loss of productivity. Change-point detection identifies the time when abnormal patterns occur, thus it is ideal for this purpose. However, trend detection is not extensively explored in existing studies about change-point detection – the widely adopted approaches mainly target abrupt mean shifts and offline monitoring. Practical considerations in manufacturing cast additional challenges to the methodology development: data complexity and real-time detection. Data complexity in manufacturing restricts the utilization of parametric statistical modeling; the industrial demand for online decision-making requires real-time detection. In this article, we develop an innovative change-point detection method based on Parsimonious Smoothing that targets trend detection in nonparametric, online settings. The proposed method is demonstrated to outperform benchmark approaches in capturing trends within complex data. A case study validates the feasibility and performance of the proposed method on real data from automotive manufacturing.

摘要

机器故障和系统性故障是由制造过程恶化引起的。尽早发现与流程恶化密切相关的模式，例如趋势，可以进行预防性维护，以避免严重损失生产力。更改点检测可识别异常模式发生的时间，因此非常适合此目的。但是，在有关变更点检测的现有研究中，并未广泛探索趋势检测-广泛采用的方法主要针对突然的均值漂移和离线监视。制造过程中的实际考虑因素给方法开发带来了额外的挑战：数据复杂性和实时检测。制造中的数据复杂性限制了参数统计建模的使用；在线决策的工业需求需要实时检测。在本文中，我们开发了一种基于简约平滑的创新变化点检测方法，该方法针对非参数趋势检测，在线设置。在捕获复杂数据中的趋势时，所提出的方法被证明优于基准方法。案例研究验证了该方法对汽车制造业真实数据的可行性和性能。