Biologically-inspired Acoustic Wear Analysis

Trying to predict the wear of a tool from the sound it makes, has a long history. People have always been interested in real-time monitoring of machinery to detect faults as and when they occur, rather than wait until the next maintenance period. This way, unnecessary maintenance, as well as long runs in a faulty condition, can be avoided. In the case of a cutting tool, trying to cut with a blunt tool can lead to the breakage of the tool and degradation of the job, while pulling the tool off for frequent assessments are expensive in terms of the machinist’s time. It is of interest to develop a method that can give an estimate of the wear from easily observable signals. This estimate, along with some kind of a confidence measure, can be used by the machinist to help his own intuition. To a machinist, the most important cue is the sound of the tool. Our goal in this work is to develop a system of classifying sounds according to the wear level of the tool that makes it. This problem has many parallels with speech recognition, but there are some difficulties unique to the tool monitoring problem. Classifying the sound is not the final goal here, as it is in the case of speech recognition. Here the aim is to classify sounds and then correlate these classes to the physical state of the tool. Efficient ways of using wear measurements are of the utmost importance. In addition, modifications are needed because the training data is very sparsely labeled (2-3 wear measurements per lifetime of tool). Previous work on estimating tool wear or damage from acoustic emissions include using the power density spectrum ([7], [8],[6] and [4]). Other approaches include looking for high energy transients in the sound signal [2], and using torque and thrust information in addition to vibration data [5] . Determining the effect of wear on the acoustic emissions of a piece of machinery is complicated by the fact that machine tools have very complex vibration modes. Usually such machines can be modeled accurately only as 3dimensional, non-linear, distributed systems, whose outputs (measured vibration)depends on the inputs (tool and job surfaces) in a very complex way. Non-linear phenomena like chatter are evidence for this [1]. Since simple models of the tool surface vibration relationship are not available, one is forced to look for nonparametric solutions to this problem. Our approach in this paper, is to first extract a feature vector from the sound, and then do a non-model-based classification using Vector Quantization [12]. Obviously, the selection of the feature vector is of paramount importance for this approach to give any good results. In this investigation we use filters based on a model of mammalian audition, followed by a tree structured classifier, based on vector quantization.