Technology demonstration to be presented at the 2013 Annual Conference of the Prognostics and Health Management Society, New Orleans, LA (October 2013)
Vehicles such as aircraft and submarines, and large facility hardware systems such as compressors, generators, and transformers, are comprised of thousands of complex and tightly integrated components that function together under great demands of stress and continual use. Simple visual inspection of these components is often very time consuming and labor intensive, and may not detect very small defects that could in fact be indicative of diminished performance or imminent system failure. Components may have inherent flaws as a result of imperfect manufacture or they may develop defects over time as a result of continual use and routine maintenance. Defects such as cracks, pits, shaft misalignment, bends, bows, loose parts, and misaligned gear teeth can all lead to decreased system performance and possible failure. These risks can be mitigated by analyzing the vibration patterns from key components to detect and identify mechanical faults before they reach the point of failure. Sensors such as accelerometers, microphones, or laser vibrometers, can be placed in specific locations to monitor the vibration signatures of the component subsystems and detect fault indicators long before they lead to failure. For example, an increase in certain high frequency vibrations might indicate that a rotor is wearing down while an overall increase in low-frequency signal power might indicate gear misalignment. We present our Analytical Signal Processing Workbench as a tool to help identify and characterize faults in complex mechanical systems. Our extensible workbench provides researchers and engineers a means to develop, execute, test, visualize, and evaluate various signal processing, feature extraction, and classification algorithms in an offline environment for later online analysis of system health and fault detection.
The workbench uses a dynamic interface to select the desired training data to run experiments and train classifiers by filtering on available meta-data attributes. The signal processing parameters are customizable so users can specify custom window sizes, normalization parameters, threshold options, and logarithmic scaling. Users can also choose to use the PWelch, standard fast Fourier transform (FFT), or multi-taper method (MTM) to convert signals into frequency space. We support a large set of feature extraction routines and RELIEF-F and K-Means clustering algorithms to decide which features provide the best separability and how many unique signal modes are present. Classifiers include Naive Bayesian, K-Nearest Neighbor, and Fuzzy-ARTMAP classifiers, along with the novel Frequency Match Analysis (FMA) and Throttle Signature Match (TSM) techniques that we developed. Throughout each step in the process, signals, features, and classification results can be visualized to support analysis and help engineers build an intuitive understanding of the problem space. Once the classifiers are created and sufficiently tuned, the compiled versions can be run as part of online applications to execute against a stream of signal data and identify abnormalities and mechanical faults in near real-time.
Our workbench supports prognostic health maintenance by helping to determine what system components are failing. This allows maintenance technicians to be more effective and efficient in maintaining and servicing hardware systems, and also allows them to monitor the progression of potential defects over time and make informed decisions regarding maintenance schedules. Analysis of vibration sensor data for fault detection promises to reduce inspection costs and time, allowing near continual health monitoring of the system, and reducing the potential for catastrophic system failure.
For More Information
(Please include your name, address, organization, and the paper reference. Requests without this information will not be honored.)