On-line surface roughness estimation in cylindrical turning using neural networks

In recent years a direct method of surface finish quality detection by electrical resistance, optical, image processing and dial indicator methods have been proved to be quantitatively unreliable. The cutting tool wear undergoes a gradual increase and the failure at the end of useful life is decided on the extent of the wear growth. This irregular tool wear trend proportionally causes the quality of the surface finish to be unpredictable. Therefore there is a great need for a reliable quantitative method to establish the quality of a work materials surface finish, online. From an industry point of view, it is often necessary to identify the production quality of work materials to avoid expensive losses on machines. A worn or wearing cutting tool will cause a gradual decline in the surface quality of a component, causing unfavorable circumstances. For example, machining of alloy wheels in Ford, Chrysler and GMH have a specific surface roughness to achieve for effective coating of the wheels. The quality of the wheel coatings and their 'scratch resistance' depend on the surface profile produced during the machining operation. While many industries adopt a 'direct modeling' approach, where the machining process variables are given as inputs to the process to estimate the surface roughness, there is no evidence available of the 'inverse modeling' where the operating conditions are estimated based on the target values of the surface roughness. Nevertheless, using the direct modeling the surface roughness can only be measured 'off line' as a quality control exercise to meet the specifications of the wheel. There is a general disagreement with the manufacturer specifications of the surface roughness, on line, with the constant changes in the cutting tool conditions and wear. Hence there is a need for 'on line' determination of the surface roughness while carrying out the machining operation. This will give an indication on the extent of tool wear growth. Further, this study will also give quantitative values of surface roughness 'on line' and its application to the manufacturing process. In this study a neural network model is proposed for surface roughness detection 'online' in cylindrical turning operations. Neural network architectures will be used as 'direct' modeling techniques to estimate the surface roughness. As 'inverse' modeling tools the operating conditions will also be predicted using the surf ace roughness and cutting tool wear as inputs. Extensive turning experiments on a lathe will be carried out covering a comprehensive range of cutting conditions to generate the knowledge base for the training stage of the neural network algorithms. The process parameters measured during the experimentation for identification of surface finish quality includes the forces, cutting tool vibrations and surface finish during cylindrical turning operation. These problem addresses one of the most pressing needs of modem automobile manufacturing industry where an 'on line' estimation of surface roughness is seen as an important parameter. This project while improving the understanding of the machining parameters and their influence on surface finish will also identify necessary neural network tools for application. This work is seen as a step towards establishing intelligent tools for machining performance estimation, while addressing the mathematical and scientific basis of machining science.