Bandsaw diagnostics by neurocomputing-two are better than one!

Abstract

In industrial sawmills, bandsaws must work at a high production rate. Two major factors which limit cutting performance are cracking and instability of the saw blades. This paper describes the results from the development of a diagnostic system which monitors blade vibration and blade tension sensor data to estimate crack length using neurocomputing techniques, to help predict blade failure. It was found that a multi-layered feedforward artificial neural network with two hidden layers produces the most reliable results. The results indicate that this system should enable the detection of cracking in blades while in a running but unloaded (between cuts) state. This may help allow longer run times to be planned with confidence increasing production uptime and minimising maintenance. Introduction.Industrial bandsaws are used for reducing logs to lumber and are considered the most important machine in the sawmilling industry because of their thin kerf and high production rate. Two major factors which limit cutting performance are cracking and instability of the saw blades.In New Zealand one current procedure is to shut down the sawmill every four hours for preventative maintenance on the saw blades. Longer production times could result if reliable blade condition information was available on-line.Neurocomputing has been successfully used for solving otherwise intractable problems by modelling them without the need for algorithmic development and the pre-emptive parameter limitations of traditional techniques. Motivation has come from the study of neuroscience [l] which offers intrinsic pattern recognition capability and learning ability, together with the an important property of robustness in response to noisy or incomplete data. Successful examples can now be found in the literature in which neurocomputing has contributed to developing solutions in manufacturing processes and their support enterprises [2; 31.Also applying neurocomputing to the classification or interpretation of data which has been transformated into the frequency (spectral) domain by a Fourier operation is not new. Work published in the late 1980s has described the successful classification of military targets detected by sonar or radar, however it has been more recently applied to a much wider range of spectral data with equal success. One such example of this is in the quantitative assessment of pyrolysis mass spectrometry measures on complex mixtures of proteins and bacteria [4].Because the spectra &om the vibrational modes of a bandsaw blade are complex 151 and those of a cracked blade even more so, the recent successes reported on spectral data interpretation have prompted the work described here. A literature survey found no applications in this area. This paper describes the development of a preliminary detection system which uses blade vibration and tension sensor data to estimate crack length to help predict blade failure. It shows that neurocomputing could interpret the data and allow an investigation of the cracking process in situ, whi...