A scientific paper entitled “Unveiling Trends and Predictions in Digital Factories” has been authored by UBITECH and is presented at the International Workshop on IoT Applications and Industry 4.0 (IoTI4 2019) that is part of the annual International Conference on Distributed Computing in Sensor Systems (DCOSS 2019), hosted between May 29-31, 2019 in Santorini, Greece. In this paper, Karagiorgou Sophia, Vafeiadis Georgios, Ntalaperas Dimitrios, Lykousas Nikolaos, Vergeti Danae and Alexandrou Dimitrios propose a failure prediction system for complex IT systems in the steel industry. The novelty of their work lies in the exploitation of Deep Learning techniques from streaming operational sensor data, enabling earlier failure predictions through a Neural Networks approach [in particular, through Long Short-Term Memory Networks (LSTM) that is a Recurrent Neural Network (RNN) architecture]. This predictive maintenance framework consists of three components: the Sense Module, the Detect Module and the Predict Module. To evaluate the proposed framework, real-life data are collected and analyzed based on daily operational and maintenance activities within the production line. They further demonstrate the framework’s potential by presenting some early results in modeling and predicting the complex and dynamic behavior in the manufacturing settings.
The Sense Module consists of a set of functionalities that are responsible for collecting data streams that are typically being generated from sensors installed at the industry’s production line. The sensor data are stored in real-time in the framework’s central InfluxDB instance. Analytics and insights generated in real-time are also stored as InfluxDB data in the form of time-series. The integration points between the sensors and the framework have been achieved by using an underlying Kafka infrastructure. Along with the centralized storage infrastructure for sensor data, categorization filters are also available for efficient distribution of each variable to the category that belongs to. In such a way, variables are categorized under specific parts of the technical equipment, which facilitates the consistent identification of specific malfunctions in the production line.
The Detect Module consists of a sequence of State Detection and Health Assessment steps. It includes real-time statistical/machine learning algorithms embedded in an appropriate and continuously processing software in order to recognize the presence of an unusual (and potentially hazardous) state within the behaviours or activities of the monitored system, with respect to some model of ’normal’ behaviour which are either modelled from domain experts or learned from real-time data observations. The diagnostic models continuously learn from the actual equipment behaviour by updating and improving the incorporated diagnostic models by using a LSTM model.
The Predict Module includes state prediction of a whole system or components with respect to mechanical system, i.e. prediction about the time-to-failure and the probability distribution function of the failure occurrence against to the projected performance level for each component of the system assuming no maintenance actions. The analysis is carried out by different algorithms and essentially by combination of data-driven algorithms in conjunction with physical models. Therefore, a LSTM model is used taking into account the data and information gathered in the InfluxDB along with actions performed in order to continuously update the prognostic models.
