Divide, Propagate and Conquer: Splitting a Complex Diagnosis Problem for Early Detection of Faults in a Manufacturing Production Line

Abstract

The following paper elaborates the procedure followed by the HIRUTEK team to solve the data challenge proposed by the PHM 2021 organisation. This challenge deals with a manufacturing line that continuously tests fuses and suffers from several malfunctions. The solution addresses the diagnosis of the faults; the efficiency of the diagnosis; the identification of the signals related to each fault type; and, the identification of different operation settings that occur during the non-faulty conditions.
The proposed problem presents some difficulties that are common to machine fault diagnosis or manufacturing line monitoring; such as the class imbalance; the high amount of missing values; multicollinearity and high dimensionality; and, experimental noise. Additionally, the evaluation criteria presents further challenges such as the consideration of chronology and the detection of operation states (also referred in the literature as context awareness). The consideration of all these factors turns this exercise in a very representative and challenging problem.
The solution here proposed, that obtained the highest score in the contest, relies on the combination of decision tree algorithms and a propagation system. The trees provide observation-wise diagnoses while the propagation system deals with chronology by adding a Kalman style filter that updates the probabilities, resulting in a more reliable result.