Prognostics in Manufacturing Systems Using DeepSurv

Abstract

Prognostics in manufacturing systems help forecast the future state using the data logs from the integrated sensors. Prognostics can be especially helpful in the current plant floor automation scenario where IoT devices are being utilized to the best possible extent. These devices generate large volumes of data that could provide value to the business when used in conjunction with analytic tools and allow businesses to take preventive measures before a breakdown could halt the plant machinery. Survival Analysis could be useful to learn the trend and forecast the event propagation using the data logs from these devices. This thesis attempts to apply modern machine learning techniques to learn the fault propagation trend from historical timestamps of the data and flag any impending breakdowns to take preventive action. DeepSurv, a python module intended for the treatment recommender setting in survival analysis, is used as a prognostic model for the manufacturing plant use case. The method is applied to the data collected from a 25-machine manufacturing plant floor. The results show that the model can forecast the future state of a machine using the historical patterns of the data logs with a Brier score of less than 0.15.