STTR Phase I: Probabilistic and Explainable Deep Learning for the Intuitive Predictive Maintenance of Industrial and Agricultural Equipment

这个小企业技术转移（STTR）第一阶段项目的更广泛影响/商业潜力提出了一种深度学习方法，用于工业和农业设备的预测性维护。高生产率要求和准时制制造方法意味着设备停机时间可能非常昂贵；研究表明，制造商平均每年要处理800小时的停机时间。此外，越来越多的老化设备和大部分达到退休年龄的维修人员导致维修人员往往缺乏知识、培训和/或人力来处理日益增长的老化资产池。该项目旨在提出一套新颖的直观和可解释的技术，可以帮助减少或消除意外停机时间，并帮助以数字方式捕获和转移退休劳动力的专业知识。这个小型企业技术转移（STTR）第一阶段项目提出了一种新颖的机器预测深度学习方法。许多现有的深度学习方法专注于给定一组训练数据的最可能失败的场景。然而，被监测的设备可能不会表现出该训练集所涵盖的行为，从而导致低置信度的预测。这种方法不仅可以预测机器部件的剩余使用寿命，而且还可以通过模型集合和预测的时间融合来量化预测的不确定性。因此，可以从基于风险的角度做出维护决策，从而消除由于低置信度预测而产生的不必要的维护。此外，许多现有的深度学习方法缺乏直观地向人类用户解释其预测的能力。在错误预测会造成严重后果的关键应用中，维护人员必须理解并信任人工智能的预测性维护合作伙伴。提出的解决方案通过突出显示对预测贡献最大的原始数据信号片段并将其动画化，为模型的预测提供直观的视觉解释。这将允许训练有素的人员通过将数据驱动的见解与他们现有的领域专业知识融合在一起，快速做出最佳的维护决策。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact/commercial potential of this Small Business Technology Transfer (STTR) Phase I project proposes a deep learning approach to predictive maintenance of industrial and agricultural equipment. High productivity demands and just-in-time approaches to manufacturing mean that equipment downtime can be extremely expensive; Research shows that the average manufacturer deals with 800 hours of downtime per year. Further, an increasing quantity of aging equipment and a maintenance workforce that is largely reaching retirement age have led to a situation where maintenance staff often lack the knowledge, training, and/or manpower to address a growing pool of aging assets. This project intends to bring forward a novel suite of intuitive and explainable technologies that can help reduce or eliminate unexpected downtime and help digitally capture and transfer expert knowledge from the retiring workforce.This Small Business Technology Transfer (STTR) Phase I project proposes a novel, deep learning approach to machinery prognostics. Many existing deep learning approaches focus on the most likely failure scenarios given a set of training data. However, monitored equipment may not exbibit behavior covered in that training set, leading to low-confidence predictions. This approach will not only predict the remaining useful life of a machine component, but it will also quantify the uncertainty of a prediction through an ensemble of models and a temporal fusion of predictions. As a result, maintenance decisions can be made from a risk-based perspective, eliminating unnecessary maintenance stemming from low-confidence predictions. Furthermore, many existing deep learning approaches lack the ability to intuitively explain their predictions to human users. In critical applications where bad predictions have serious consequences, maintenance personnel must understand and trust an artificially intelligent predictive maintenance partner. The proposed solution produces an intuitive visual explanation for the model’s prediction by highlighting and animating the segments of a raw data signal that are contributing most significantly to the prediction. This will allow trained personnel to quickly make optimal maintenance decisions by fusing data-driven insights with their existing domain expertise.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.