8.05 Machine learning (ML) models and technological solutions to support predictive maintenance, quality & energy efficiency monitoring, control and product improvement in industrial applications and multi-energy systems

Energy efficiency of energy–intensive industries as well as local energy distribution networks and multi-energy systems goes through the definition and implementation of innovative monitoring and control systems that can transform the collected data into correlated and usable information by means of a sustainable, well-designed, and upgradable energy management information system (EMIS).

Furthermore, Energy efficiency is strictly linked to Operational Excellence through the implementation of a Prognostic and Health Management System (PHMS) capable to monitor critical productive assets, facilities and areas, to detect anomalies, identify fault causes and predict remaining useful life in order to support decision in real time driven by the state-of-the-art condition-based technology techniques and strategies for reliability centered maintenance.

Both EMIS and PHMS combine software, hardware, and data to support people in managing energy at the process, system, facility, and enterprise levels. Their function can be grouped into four layers, devoted to:

1. Data Acquisition, Integration and Adaptation
2. Data Classification, Transformation, and Storage
3. Data Correlation, Analysis
4. Data visualization

Layer1 includes the collection of the sensor data needed for the analysis, through proper IoT technologies and platforms, their integration and their adaptation to the format required by the subsequent operations. Layer2 requires the use of innovative Data Mining techniques, and the use of Machine Learning (ML) models is strictly required to develop – in Layer3 – smart monitoring functions devoted to both efficiency analysis and condition monitoring for fault detection, predictive maintenance, and control. ML-based technologies allow to implement efficiently the targets presented in layer2 and in particular they allows to define tools for:

• Diagnostics: approaches to automatically analyze the status of complex machines, processes or products. In this category, Deep Learning-based approaches have proliferated for quality/defect monitoring. Moreover, due to typical last of labels in productive environment, unsupervised approaches for anomaly detection become an extremely popular tool for monitoring.
• Prognostics: approaches for predicting process degradations or equipment failures. Thanks to such technologies, advanced maintenance management policies, like Predictive Maintenance, can be enabled.
• Energy Efficiency and Resource planning: approaches to predict the critical behaviour of processes and machinery from the point of view of energy consumption and production and for the planning of the use of the resources including storages and self-production of energy.

In the project, the algorithms that will be developed for diagnostics, prognostics and Energy Efficiency, focus to:

1. the relationship between such approaches and control;
2. the interpretability of such solutions for enabling decision making;
3. the implementation of such approaches in the IoT scenario.
   An innovative challenge of this project is to complement EMIS/PHMS following the Digital Twin (DT) paradigm.

Development of a digital twin of the product through the identification of cross-correlations to support Design Failure Mode Effect Analysis, Design for Manufacturability, critical to quality measurements identification and range setting and advanced root cause analysis support.
The Product Digital Twin is based on the concept of linking all the measures of test report of the item with the measurements of tests carried out at a sub-assembly level, associated to the relative data of inventory to guarantee the operational performance required by the product.
A task of the project will be, starting from the information collected and evaluated by EMIS/PHMS, the development of methods and techniques for the implementation of Machine Learning algorithms on image recognition, defect identification, clustering and advanced statistical process control capable to exploit data retrieved by automated optical inspection machines.
Aim of the method is to develop a quality control process that detects defects and performs assessments based on machine learning models determining whether the target unit has any defects present that will negatively affect the resulting product.