Development of an autonomous cutting force data acquisition and prediction system for CNC turning
Student thesis: Doctoral Thesis
The present study attempts to develop a new solution for cutting force prediction that each machine tool on the shop floor can implement autonomously. The proposed strategy, Autonomous Cutting Force Data Acquisition and Prediction System (ACFDAPS), is based on a combination of three different areas: sensing, modeling and learning. Four features characterize the system: (i) autonomous extraction of cutting force data from the machine’s routine shop-floor experiences (ii) parallel application of a range of plausible process models each capable of providing an independent force estimate, (iii) progressive refinement of the model coefficients on the basis of further cutting experiences, and (iv) identification of the most effective model for each work-tool material combination by applying the principle of “survival of the fittest”. The new system leads to the possibility of implementing distributed machining databases where each machine could autonomously compile its own machining database based on its routine shop floor experiences. Knowledge of cutting force magnitudes is a prerequisite for the modeling of temperatures, tool wear and machining errors, etc. Traditional cutting force prediction approaches have relied on machining databases consisting of direct mappings between input conditions and output forces obtained via off-line experimentation conducted at remote laboratories. The databases tend to be voluminous because of the large input space and sparse because of the expense involved in manually collecting the data. The latter feature often leads to significant interpolation or extrapolation errors during the prediction phase. Further, the method cannot take into account machine-specific effects. Finally, the method does not yield any physical insights into the process (e.g., the magnitudes of shear angle, chip-tool friction coefficient, etc.) necessary to perform downstream modeling of cutting temperatures, etc. ACFDAPS is designed to overcome all these limitations. In ACFDAPS, cutting force data are collected autonomously and on-line for each specific machine via motor current sensing. The method is inexpensive and shop floor friendly because the Hall-effect current sensors used cost only a few dollars and are easily slipped round the input cables of motors. Data are collected only with respect input condition ranges of relevance to the particular machine in the particular shop floor context. The problem of sparse databases is solved by attempting to recognize the input-output patterns implicit in the data rather than through direct input-output mapping. The patterns will be different for different work-tool material combinations. They could also be different under different input condition ranges. One method of recognizing the patterns is to use a neural net. But this method yields no physical insights. A similar problem exists with empirical modeling where cutting forces are expressed as analytical functions of input conditions. In contrast, much deeper physical insights can be obtained by utilizing an analytical model. Analytical modeling of a practical operation (e.g., turning, or end-milling) requires identifying the equivalent single edge operation at each cutting point. However, there are many single edge models reported in literature. We do not know in advance which model is most appropriate with respect to a given work-tool material combination and range of input conditions. These problems are solved in ACFDAPS by including as many of the plausible process models (including neural, empirical and analytical models) as available and letting the system learn through experience which models are the fittest under which conditions. The system stores the model coefficients for each new work-tool material combination and updates (refines) them each time when the same combination is encountered on the shop floor. One major problem with the use of many analytical single edge models is the need for prior knowledge of the magnitude of shear angle. Traditionally, this problem has been sought to be overcome through the development of shear angle solutions. However, the problem has proved to be intractable despite the formulation over fifty shear angle solution. This need has so far been the major hurdle to the industrial application of analytical models of machining. This problem is however totally sidestepped in this thesis. This is achieved assuming a new Extended Linear Shear Angle Solution (ELinSAS) and determining the coefficients involved in the solution by means of a novel optimization technique called ‘Minimize the Variance of the Model Invariant (MVMI)’. This thesis includes the various analytical components of ACFDAPS with respect to turning (including profile turning) and end-milling. Extensive turning experiments (and some preliminary end-milling experiments) have been conducted to verify the different components of ACFDAPS. The results indicate that the accuracy of the predicted force from ACFDAPS is of the order of 30-40N. It appears that, except in the case of fine machining, the prediction accuracy compares very favorably with that generally obtainable through conventional databases using direct input-output mappings or index sets derived from analytical process models. Also, the proposed system is able to provide additional information about the cutting process such as shear angle solution, etc.
- Cutting, Machining