We have been researching the application of artificial neural network (ANN) technology to analysis of tribological behavior. The report "Neural Network Models of Simple Mechanical Systems Illustrating the Feasibility of Accelerated Life Testing", NASA TM-107108 by Steven P. Jones, Ralph Jansen, and Robert L. Fusaro explores the design and training of various networks for modeling wear in three tribological tests. The ultimate goal is to ascertain whether ANN models can be used for accelerated-life testing.
These models were constructed from "Input Layer Dampened Recurrent Networks" with supervised backpropagation training. A number of architectures were investigated systematically before arriving at this design. The GRNN was actually very accurate at modeling the training set according to the R^2 coefficient, but it could not generalize well.
You can access the
abstract and ordering information for NASA TM-107108.
(The report is not available for download from the Glenn report server.)
|Architecutral Design||Rub Shoe Model||Pin-on-Disk Model||Four-Ball Model|
|3 Layer Backpropagation||0.76||0.89||0.67|
|4 Layer Backpropagation||0.78||0.89||0.69|
|5 Layer Backpropagation||N/A||0.88||0.66|
|Input Layer Dampened Recurrent Network||0.84||0.93||0.92|
|Hidden Layer Dampened Recurrent Network||0.82||0.91||0.82|
|Output Layer Dampened Recurrent Network||0.77||0.89||0.60|
|2 Hidden Layers with different activation functions||0.63||0.90||0.69|
|3 Hidden Layers with different activation functions||0.74||0.89||0.67|
|2 Hidden Layers with different activation function and jump connection||0.76||0.90||0.67|
|3 Layers with jump connections||0.77||0.84||0.61|
|4 Layers with jump connections||0.76||0.84||0.54|
|5 Layers with jump connections||0.77||0.84||0.54|
|Input Variable||Contribution Strength|
From an ANN standpoint, the problem with the bearing data is that they are few. In bearing tests it is impractical to collect huge amounts of data. This makes it very difficult for the ANN to model such a sophisticated relationship - bearing systems are extremely sensitive to a number of inputs, including load, speed, sliding distance, lubricant, temperature, atmosphere, manufacture, etc. Our sets have consisted of about 50 data, and that's quite a large sample, to give you some idea. We take some of the above input parameters and calculate a single output - wear volume.
We had two ideas for accelerated-life testing. The first was to perform simple extrapolation, but these nets seem to be quite bad at this (worse than other statistical methods tried). The other idea was to focus on detecting anomalies in the experimental data which might indicate long-term failure. For this we might use an unsupervised network (Kohonen, to be exact) to detect anomalous behavior. This is our current area of exploration.
You also maybe interested in the comp.ai.neural-nets newsgroup and its FAQ (Frequently Asked Questions document). Unfortunately both the newsgroup and FAQ are ill-suited to beginners.
|Space Mechanisms Project Page | Mechanical Components Branch|
Some of the documents in this page are provided
in PDF format and require Adobe's Acrobat Reader.
Download the Adobe plug-in from www.adobe.com.
Curator & responsible official: Phillip.Abel_AT_nasa.gov.
Last modified 28 Nov 2006