Dynamic Models of Machining Vibrations,

Designed for Classification of Tool Wear

Randall K. Fish

Abstract

The goal of this dissertation is to develop a machining tool-wear classification system which uses features drawn from accelerometers that respond to machining vibrations. Specifically, we use features from wide band accelerometer signals ina two stage dyanmic classifier estimating the flank wear on end mills cutting notches in either steel or titanium workpieces. Since no standard data set and test paradigm exists for this task, we introduce an experimental paradigm which incorporates new evaluation metrics not previously used in tool-wear monitoring.

Until recently, only static classifiers have been used for tool-wear applications. However, the process of increasing wear is dynamic. Individual wear events which occur at a changing rate and last for a few milliseconds gradually change a tool's cutting edge from sharp to dull. Our experiments also show that within an individual cutting pass the wear process changes as the cutter moves into and out of "regions of interest" which effect the sensor features used in classification. We select features which are sensitive to the dynamics of these various time scales. We demonstrate a single-rate dynamic classifier which models the dynamics of wear both within an individual cutting pass and also over the cutting life of the tool.

Our single-rate dynamic classifier captures the slowly varying wear phenomena by using sequential states in a hidden Markov model. To improve the modeling of the rapidly varying discrete wear events that last several milliseconds, we extend the single-rate dynamic classifier to a multi-rate classifier. The multi-rate classifier splits the task of modeling events at the two time scales into two state-coupled classifiers processing feature streams at different data rates. We demonstrate that coupling the two classifiers during classification gives better performance than combining the outputs of the separate classifiers in a second stage.

The availability of data in this application is limited. Data annotated with the correct level of wear is even more scarce. We demonstrate a method of using both labeled and unlabeled data to train model parameters. The broad range of cutting conditions encountered in actual industrial practice imposes the need for the classifier to generalize to cutting conditions not included in the model training. We demonstrate feature processing which allows us to generalize to a limited range of cutting conditions including the use of features drawn from accelerometers with different response characteristics.

Our classification system is not intended to be the sole arbiter of the decision of whether or not a cutter should continue to be used or be replaced. We present the information from the classifier in several different formats to assist the machinist in making an informed decision. Our system estimates the wear on the primary cutting edge at the end of each cutting pass. In addition to this estimate, we provide a measure of the confidence in the cutter wear having exceeded a predefined level considered to constitute the end of the cutter's useful life. Prior to cutting with a new tool, the useful life for the cutter is expected to be the average for this type of cutter being used under the present cutting conditions. At the end of each cutting pass, our system updates this estimate of the remaining cutter life. We incorporate the actual cutting behavior seen for the particular cutter in use resulting in a more accurate prediction than is possible with a simple average.

The accuracy of our single-rate classifier is 90% to 97% when classifying the wear on cutters milling steel. Even on the more difficult problem of classification when cutting titanium, our multi-rate classifier achieves accuracy of 94%.

The full thesis in pdf format.


Return to the SSLI Lab Graduate Students Theses Page.