A major stake of the Industry 4.0 is in the exploitation of the data that are emitted by production machines. When these data are becoming heterogeneous and voluminous, Big Data and Data Mining approaches are typically used. The institute of Complex System (iCoSys) at HEIA-FR has made of this topic an important axis of research.
In the Process4Plastics P4P-2 project, iCoSys has applied such technologies with good success on the experimental PICC platform for plastic injection molding. The first step of such process is to collect and store the data from the sensors embedded in the machines. Once collected, the data are transferred to DAPLAB, a Big Data cloud-like infrastructure recently assembled in the premises of HEIA-FR for data storage and processing. The second step involves Data Mining technologies that include data preprocessing (cleaning, interpolation, filtering) and Machine Learning based modelling. The principle of Machine Learning is to infer automatically complex mathematical models on exemplary data, doing a partial or total abstraction of the underlying physical phenomena. Such approaches allow for predictive and prescriptive analysis. Predictive models are used to predict future data of machine processes such as quality indices. Prescriptive models are used to help decisions such as for the initialization of machine parameters to values that are likely to bring the process in acceptable outputs. In the P4P-2 project, different use cases were identified, analyzed and validated on data capture from the experimental PICC platform. A first use case was about automatic anomaly detection. A model was automatically trained using Machine Learning on historical data where the machine was in stable and regular production mode. This model was then used to detect abnormal deviations of the data streaming out of the machine in live production. Once an alarm triggered, a message was automatically sent to the operator. The second use case was about the prediction of quality indices using a time frame of recent data emitted by the machine. This use case shown the potential to reduce time consuming human quality inspection while providing a non-stop quality control mechanism. Other use cases have also shown potential interests such as machine parameter optimization and sensor redundancy analysis.
Co-Head of the iCoSys Institute