raj@utkux1.utk.edu (Rajendra Patil) (06/16/91)
Hi n-net world: Few months back there was a post related to neural net usage for vibration analysis of rotating machine, this post is related. We (3) have been working on this for last few months. Two of the members are using classifier approach to recognize fault spectrum of the machine. We have some arguments and I hope someone could tell us their views. The objective is to develop NN based system to diagnose faults in the rotating machine (for now basic faults). I say that classifier approach though feasible, is not practical for following facts. To train a classifier we will need reliable data from the machine. Classifier trained on data from one machine may not classify the spectral patterns from other machine due to difference in environmental, operating, and physical conditions. Collection of data under different faults with different degrees of severity is not feasible for training purpose. Training a classifier is done based on the complete geometrical shape of the spectra and not on the feature. Fault in rotating machine is indicated very specifically in the spectrum and not by the complete spectrum. What happens if the training data collection setup has different parameters, gain, bias, filters, sensor locations then the testing data collection setup. The spectral signatures under the same faults will be significantly different. The classifier will not work. This is like 100 new cars of the same type used for 1 year are to be diagnosed. Classifier approach says that pick up any car, simulate the faults, collect data, train a classifier and expect this classifier to diagnose rest 99 cars used under different conditions. The assumption that all the machine are in same operating, environmental and physical condition is not practical. Machine on first floor will have different spectral signature as compared to machine on ground floor under the same fault condition. All this suggests that, for a classifier system to work reliably for vibration diagnostics purpose, take a machine, simulate different faults with different degree of severity, record the data acquisition parameters, train the classifier neural net for days (large data), and use it to test the same machine with same operating, environmental, physical and data acquisition setup parameter conditions. This whole thing is not practical. I have a feeling that above approach is wrong for a problem like vibrations. What else can be tried? ( I am working on connectionist expert systems and works good). What I need is your views about the classifier approach (will work, or not and Why?). I am having a hard time convencing my fellow members that it is not practical. Comments about anything said above are very much appreciated. Regards. Raj