An Agnostic Data-Driven Approach to Predict Stoppages of Industrial Packing Machine in Near

Abstract

As data awareness in manufacturing companies increases with the deployment of sensors and Internet of Things (IoT) devices, data-driven maintenance and prediction have become quite popular in the Industry 4.0 paradigm. Machine Learning (ML) has been recognised as a promising, efficient and reliable tool for fault detection use cases, as it allows to export important knowledge from monitored assets. Scientists deal with issues such as the small amount of data that indicate potential problems, or the imbalance which exists between the standard process data and the data inadequacy of the systems to make a high precision forecast. Currently, in this context, even large industries are not able to effectively predict abnormal behaviors in their tools, processes and equipment, when adopting strategies to anticipate crucial events. In this paper, we propose a methodology to enable prediction of a packing machine’s stoppages in manufacturing process of a large industry, by using forecasting techniques based on univariate time series data. There are more than 100 reasons that cause the machine to stop, in a quite big production line length. However, we use a single signal, concerning the machines operational status to make our prediction, without considering other fault or warning signals, hence its characterization as "agnostic". A workflow is presented for cleaning and preprocessing the data, and for training and evaluating a predictive model. Two predictive models, namely ARIMA and Prophet, are applied and evaluated on real data from an advanced machining process used for packing. Training and evaluation tests indicate that the results of the applied methods perform well on a daily basis. Our work can be further extended and act as reference for future research activities that could lead to more robust and accurate prediction frameworks.