Machine Fault Diagnosis of Fused Filament Fabrication Process with Physics-Constrained Dictionary Learning
Yanglong Lu, Yan Wang
Abstract
Sensors have been widely applied in modern manufacturing systems to monitor the processes and machine health conditions in order to control product quality. Processing a large amount of sensor data becomes a new challenge on the efficiency of diagnosis. In this paper, a novel physics-constrained dictionary learning approach is proposed to simultaneously improve the efficiency of data collection with compressed sensing (CS) and perform diagnosis with the classification of sensor data. Two-stage optimization is performed. At the first stage, measurement matrix is optimized to determine the time stamps of collected data points with a fixed basis matrix. This is solved based on a constrained FrameSense algorithm. At the second stage, the basis and classification matrices are optimized with the fixed measurement matrix based on the K-SVD algorithm. The above two optimization steps are repeated until the optimal measurement, classification, and basis matrices converge without further improvement. The recovered signals can be classified more accurately based on the learned classification matrix for specific data. The proposed approach for machine fault diagnosis is demonstrated with acoustic emission signals collected in the fused filament fabrication process.