A flexible method to detect the fault of components in an injection group of a diecasting machine

This work presents a simple method to detect the fault of components in an injection group of a diecasting machine. A usual problem in predictive diagnostics in industrial applications is the lack of replicable failure and fatal data: the parameters of the diecasting process are often changed to adapt the machine to a new production cycle and this makes difficult to identify faults using automated data analysis. The proposed method is based on an algorithm able to re-train itself when a change in production is detected. The final prediction of each fault condition is performed by combining classical machine learning analysis and experts' knowledge of the field by identifying discriminant weights to insert in the machine learning analysis. These weights are quantities that represent the experts' knowledge and the algorithm take into account.