Learning via acceleration spectrograms of a DC motor system with application to condition monitoring

In a highly automated manufacturing plant, the reliability of manufacturing equipment is critical for normal operation. A sudden machine breakdown can bring unexpected downtime, shorter lifespan of equipment, and lower operational efficiency. Breakdowns can lead to defective parts and consume extra energy-issues that are undesirable from an environmental standpoint-and also erode productivity and increase costs. To improve machine tool reliability, a machine may be continuously monitored to track its health condition. Monitoring a machine often provides large amounts of data that must be processed to distill useful information. Electric motors are found in many pieces of common manufacturing equipment. Deep learning methods can be combined with data collected on motors, e.g., acceleration time-frequency data, to identify motor condition. In this paper, three state-of-the-art deep learning architectures are evaluated for their ability to effectively monitor motor condition. Experiments are performed on a lab-scale motor test bed to secure condition data for several common motor faults. Tri-axial acceleration data are collected and converted into 2D images (spectrograms) using the power spectral density function. Some of these experiments are used to tune the deep learning algorithms, and others are used to test the proposed monitoring methods. The relative performances of the architectures are assessed, and it is demonstrated that the use of time-frequency images within a deep learning context can efficiently handle large amounts of data and effectively monitor the motor condition.