Soft computing for tool life prediction a manufacturing application of neural - fuzzy systems

Abstract

Tooling technology is recognised as an element of vital importance within the manufacturing industry. Critical tooling decisions related to tool selection, tool life management, optimal determination of cutting conditions and on-line machining process monitoring and control are based on the existence of reliable detailed process models. Among the decisive factors of process planning and control activities, tool wear and tool life considerations hold a dominant role. Yet, both off-line tool life prediction, as well as real tune tool wear identification and prediction are still issues open to research. The main reason lies with the large number of factors, influencing tool wear, some of them being of stochastic nature. The inherent variability of workpiece materials, cutting tools and machine characteristics, further increases the uncertainty about the machining optimisation problem. In machining practice, tool life prediction is based on the availability of data provided from tool manufacturers, machining data handbooks or from the shop floor. This thesis recognises the need for a data-driven, flexible and yet simple approach in predicting tool life. Model building from sample data depends on the availability of a sufficiently rich cutting data set. Flexibility requires a tool-life model with high adaptation capacity. Simplicity calls for a solution with low complexity and easily interpretable by the user. A neural-fuzzy systems approach is adopted, which meets these targets and predicts tool life for a wide range of turning operations. A literature review has been carried out, covering areas such as tool wear and tool life, neural networks, frizzy sets theory and neural-fuzzy systems integration. Various sources of tool life data have been examined. It is concluded that a combined use of simulated data from existing tool life models and real life data is the best policy to follow. The neurofuzzy tool life model developed is constructed by employing neural network-like learning algorithms. The trained model stores the learned knowledge in the form of frizzy IF-THEN rules on its structure, thus featuring desired transparency. Low model complexity is ensured by employing an algorithm which constructs a rule base of reduced size from the available data. In addition, the flexibility of the developed model is demonstrated by the ease, speed and efficiency of its adaptation on the basis of new tool life data. The development of the neurofuzzy tool life model is based on the Fuzzy Logic Toolbox (vl.0) of MATLAB (v4.2cl), a dedicated tool which facilitates design and evaluation of fuzzy logic systems. Extensive results are presented, which demonstrate the neurofuzzy model predictive performance. The model can be directly employed within a process planning system, facilitating the optimisation of turning operations. Recommendations aremade for further enhancements towards this direction.