Hybrid Data-driven Physics-based Modeling for Machine Fault Detection, Diagnosis, and Prediction

优化操作和维护活动将极大地提高加拿大大多数工商业部门的效率和生产率。然而，这需要收集和适当使用与系统性能、降级和故障相关的有意义的参数。现有的大多数机械和结构部件和系统的监测和维护决策支持策略仍然需要人工监督和决策，特别是当所考虑的系统是复杂的、移动的、远程的和/或以非稳态模式运行时。这项活动的重要部分迫切需要实现自动化。当有大量历史数据可用时，故障检测和诊断成为可能。数据驱动的方法在这里展示了巨大的潜力，因为它们能够对数据进行分类并识别代表错误条件的模式。然而，当只有表示故障和/或降级的有限数据可用时，这些方法受到严重限制。将探索新的递归数据处理策略(特别适用于从旋转机械收集的动态信号)，以提高这些方法在数据稀缺时的稳健性。此外，预测在使用基于数据的方法时更具挑战性，因为它们只代表过去的经验。将开发能够整合在线收集的新数据的新技术，从而允许快速更新模型以改善预后。基于物理的模型是很好的预测工具。这些模型在大小和复杂性上可能有很大的差异，但进行修改以允许合并组件故障或系统降级相对容易。这有助于系统或组件性能预测。将开发用于齿轮齿、行星齿轮系统和电机/发电机系统的新型号。将来自多个来源的信息组合在一起可以显著提高置信度。数据驱动和基于物理的混合协议将允许增强两者的优点，并将缺点降至最低。这种混合方法将有助于优化系统运行和维护。这个风向标的初步工作已经表明，在精确度方面有可能大幅提高。进一步的发展可能会在系统退化检测、故障诊断和故障预测方面带来巨大的改进。迫切需要在混合动力战略设计及其在更广泛的工业和商业应用中的应用方面取得突破，以满足自动驾驶系统(汽车、轻轨列车、风力涡轮机)的快速采用。

Optimization of operation and maintenance activities would result in huge efficiency and productivity improvements across most industrial and commercial sectors in Canada. However, this requires the collection and appropriate use of meaningful parameters that correlate with system performance, degradation, and failure. Existing monitoring and maintenance decision support strategies for most mechanical and structural components and systems still require human supervision and decision making, especially when the system being considered is complex, mobile, remote and/or operates in non-steady state modes. Automation of significant parts of this activity is urgently needed. When large amounts of historical data are available, fault detection and diagnosis is possible. Data-driven methods demonstrate huge potential here because of their ability to sort data and recognize patterns representing faulty conditions. However, when only limited data is available that represents failure and/or degradation, these methods are severely constrained. New recursive data processing strategies (particularly appropriate for dynamic signals collected from rotating machinery) will be explored to improve the robustness of these methods when data is scarce. Additionally, prediction is more challenging when using data-based methods because they only represent past experience. New techniques will be developed that can integrate new data collected on-line allowing for rapidly updated models for improved prognostics. Physics-based models are excellent tools for prediction. These models may range dramatically in size and complexity, but modification to allow incorporation of component faults or system degradation is relatively easy. This facilitates system or component performance prediction. New models will be developed for gear teeth, planetary gear systems, and motor/generator systems. Combining information from multiple sources significantly improves the confidence level. Hybrid data-driven and physics-based protocols will allow the advantages of both to be enhanced and the disadvantages to be minimized. Such hybrid approaches will facilitate the optimization of system operation and maintenance. Preliminary work in this vane has already shown that dramatic improvements in accuracy are possible. Further development could result in huge improvements in system degradation detection, fault diagnosis and failure prediction. A breakthrough in hybrid strategy designs and their application across a wider array of industries and commercial applications is critically needed to service the rapidly expanding adoption of autonomous systems (cars, light rail trains, wind turbine generators).