SOM with neighborhood step decay for motor current based diagnostics

Embedded self-learning is a desired capability that can enhance autonomy in different types of unmanned systems. Autonomous diagnostics is an area of opportunity to deploy this capability, which allows for vehicle failure awareness and enables for other advantageous schemes such as fault tolerant control. In this paper, we present one subsystem of an ensemble of schemes that form the Enhanced Autonomous Health Monitoring System (EAHMS) designed to support NASA's Robotics, Tele-Robotics and Autonomous Systems Roadmap. The EAHMS is aimed to provide an integral framework to determine the operational condition of on-board sensors (odometry), actuators, and power systems. Within the EAHMS context, this paper outlines a method for diagnostics of a robotic vehicle mechanical mobility subsystem by motor current and vibration signature analysis based upon Self Organizing Maps (SOM) using an enhanced neighborhood step decay algorithm. The learning algorithm was tested for different learning rate functions and was applied to different training set cases. The resulting algorithm was used for conducting failure diagnostics in a testbed, where three types of transmission/motor mechanical failures were considered: (a) damaged chain link; (b) motor gearbox damage; and (c) damaged sprocket. A core goal of this diagnostic approach is to enhance a novel methodology called the embedded Collaborative Learning Engine (eCLE), which combines supervised and unsupervised learning synergistically to process new emerging data signatures. This technique for system enhancement and application results are described in this paper.