Incipient Engine Failure Predictor, 03-9182Printer Friendly Version
Inclusive Dates: 03/01/00 - 06/01/00
Background - The technology investigated in this project can be used to solve problems for commercial and military clients. On the civilian side, increasingly stringent emission regulations place a de facto requirement on long-term engine performance. The onset of any performance degradation often increases exhaust, crankcase emissions, or both beyond legislated levels. In the military environment, combat readiness is always a prime concern. However, scheduled maintenance based on simple statistical analysis results in significant levels of unnecessary maintenance, while missing some early failures. The unnecessary maintenance causes logistical problems, particularly for field operations, while early failure can be catastrophic for a crew, or even an entire unit, in a combat situation. Commercial and military applications require a method of predicting engine degradation and failure in real-time, on an individual vehicle basis.
Approach - The goal of the project was to investigate reciprocating engine failure prediction using vibration signals. Vibration data were collected from five accelerometers on a medium-duty diesel engine that was subjected to a 250-hour endurance test under varying speed and load conditions. Data from several degraded engine states were collected for comparison to data collected from a good engine under the same conditions. Failure prediction was to be achieved by mathematical comparison of good engine vibration content to the bad engine's content for a given load case.
The unique aspect of this approach was application of genetic algorithms (GAs) to vibration signal processing. This method permitted rapid, automated optimization of signal processing to differentiate the signals from good and degraded engines. The automated nature of the process permitted investigation of a wide array of signal-processing strategies, aided by the robustness of the GA, which can accommodate and compare discontinuous data and signal-processing methods. An example of discontinuous data would be comparison of data from different accelerometer locations. To maintain continuity from data-acquisition through signal-processing stages and optimization, all data acquisition and signal processing were performed using SwRI-developed software and algorithms.
Accomplishments - In this project, approximately 1.5 Gb of data were obtained to fully characterize good, degraded, and failed engine states. A functioning prototype of the GA for signal-processing optimization was coded and found to converge. The signal-processing methods were encoded into a 14-bit genome consisting of five alleles. The alleles were used to select the best combination of firing cycles per cyclic average, number of cyclic averages, windowing method, number of asynchronous frequency domain averages, and accelerometer location. Convergence was achieved in only three generations with an initial population of 100. The time required to evaluate each 100-member generation on a 500-MHz PIII computer was approximately 1,500 seconds. The genetic algorithm for this project was coded so as to be modular and generic in nature. For application to other optimization problems, only the cost function and the allele translation portions of the code need to be changed.