Feature Engineering for Digital Twin Design: Industrial Applications

Predictive maintenance (PdM) is one of the key points in the Industry 4.0 revolution, and its characteristics of increased resilience, sustainability, control, and human safety are of key importance for Industry 5.0. PdM was already established in most industry scenarios, but the advent of Machine Learning (ML) and Internet of Things (IoT) technologies has completely revolutionized the sector. Degradation processes are very complex and dependent on external factors, so constant and sparse data collection enables the Machine Learning algorithm to identify relationships between different data sources, capturing diagnostic information that could not have emerged from a purely physical/technical analysis. In the literature, can be observed many works concerning the generation of new ML techniques or improvements to existing techniques, with a constant desire to overcome some benchmark. However, the real key to the success of an ML application lies in the preliminary problem-setting process, which involves a series of technical choices related to the data collection step and the pre-processing operations required to provide this data as input to one (or more) ML algorithms. This process is called the feature engineering Process, and this industrial PhD thesis will present two main research contexts regarding PdM-related applications in which the feature engineering process was crucial for effective real-world implementation. In the first context, the problem of hardware reduction for smart chemical sensors equipped with AI units for health and quality monitoring will be addressed. The feature engineering process allows us to save hardware resources, evaluate the appropriate classification algorithm, and increase sensor useful life in case of a corrosive environment. The second context regards activities related to two research projects, ASSIOMI and E-ADAPTIVE. Here, a complete physical health management system implementation for a CNC milling machine is presented, detailing all the crucial steps, from the data communication performed by an infrastructure to the data-driven alerting unit modeling (also including the design of the experimental setting for training dataset collection). The activities also include an application of early exiting in Classification Tree inference to decrease the rate of data acquisition operation and an interesting feature engineering solution to preventively detect rotational machine eccentricity faults in the startup phase. Again, the cost reduction purpose of the sensor choice will be addressed by increasing the complexity of the feature engineering process.