Design of a Physics-Based and Data-Driven Hybrid Model for Predictive Maintenance

The maintenance process is crucial in any system that is prone to failure or degradation, particularly in manufacturing operations. In fact, maintenance costs can reach up to 40% of the cost of production in certain industries. In the era of Industry 4.0, maintenance methods can maximize the use of components predicting the remaining useful life. These methods are identified as Predictive Maintenance and include several innovative technologies, such as IoT for deploying sensors that monitor machines and AI that provides the algorithms to interpret the data collected. The information generated from sensor data allows for more accurate predictions using statistical models that are sensitive to the peculiarities of an individual tool set on a particular machine and used by a certain operator. These models, unlike traditional methods based on physical laws, increase in efficiency as the data increases, and therefore are not efficient or usable when a sufficient bank of data is not available. This work proposes a hybrid model that, being based on both classical physics and data-drive models, demonstrates how it is possible to obtain a prediction method that estimates the state of the tool even in the absence of historical data and that increases its accuracy as such data increases. The proposed model is evaluated by using a public experimental milling dataset.