Equipment Condition Monitoring
Recent advances in wireless networks and embedded computational devices have made it possible to monitor cost-efficiently and in real-time the vast majority of the electromechanical devices. This has opened the possibility for timely fault detection, diagnosis, and even prognosis of future abnormalities. Our objective is to develop fully automated machine learning algorithms for building probabilistic models of normal operation from normal data, and classifiers and predictors for fault disambiguation and prognosis.
Background & Objective: We have employed various machine learning methods for fault detection and diagnosis, and have achieved success with memory-based learning algorithms such a locally-weighted regression. Furthermore, we have developed original algorithms for fast abrupt change detection in sensor data streams based on the same memory-based approach.
Technical Discussion: The readings produced by sensors attached to electromechanical machines are random variables that fluctuate constantly depending on the operating mode of the equipment. The main technical challenge in equipment condition monitoring is to distinguish those normal variations from deviations due to abnormal behavior and/or faulty operation. The problem reduces to learning probabilistic models of conditional densities, where the conditioning is upon external driving variables, and test robustly for deviations from these densities in real time. We have been pursuing a memory-based learning approach, and have been able to exploit the repetitive computational structure of memory-based density estimates to propose novel algorithms for abrupt anomaly detection in sensor data streams that have excellent computational complexity. Our algorithms MB-GT and MB-CUSUM have complexity only quadratic in the size of the memory buffer of sensor readings, and allow sensor stream monitoring in real time. They have been implemented in C and are available for operation in embedded systems.
Future Direction: We are continuing work on expanding this technology to high-dimensional data streams, typically generated by sensors attached to rotating machinery. If successful, such technology would allow us to address the problem of monitoring of motors and generators and all their industrial uses, such as elevators, compressors, pumps, etc.
Contacts:
Daniel Nikovski
Ajay Divakaran
Kevin W. Wilson
| Technical Reports: | |
| A Comparison Between Polynomial and Locally Weighted Regression for Fault Detection and Diagnosis of HVAC Equipment | |
Technology Area: Sensor and Data Systems
Modification Date: October 2, 2007
