Adaptive Correction of Errors in Cnc Turning Using Artificial Neural Networks

This paper presents the development of an intelligent adaptive correction tool for improving the profile accuracy of parts produced on CNC turning centers. Actual dimensions and shape (form) produced on components during machining differ from the nominal dimensions commanded in the CNC program. The error if predicted apriori can be used to correct the ideal CNC code and thus improve the part accuracy in CNC machining. In the present work Artificial neural networks (ANN) are used to capture the complex relationship between the errors on the component and process conditions. The predicted profile error was analysed and the ideal CNC code was adaptively revised. Corrected CNC code was found to significantly reduce the profile errors on the component. INTRODUCTION Manufacturing industries worldwide, are facing several challenges such as shorter product life cycles, frequent changes in product design, small volume of production and demand for higher productivity and quality of products. During the last decade several CAD/CAM/CIM tools have been developed to automate and integrate various phases of the product cycle. Among them CNC technology assumes a special significance as it provides a means for rapid conversion of design models into physical prototypes. The accuracy and utilization of CNC machine tools is dictated to a large extent by the availability of an error free optimum CNC program (code). Despite significant developments in CNC program generation & verification tools and CAM software development [1], CNC code generation depends primarily on ideal (nominal) part geometry, disregarding tool/work deflections, static and dynamic compliance and inherent errors present in the machine tool.[2] The dimensions and form of the part produced are thus different from ideal (commanded) tool/work movement in the CNC code. Accuracy of parts produced on CNC machine tool can be enhanced if one has an apriori knowledge about the pattern of errors/deflections on the machine tool and the factors causing them. A need therefore exists to intelligently capture this complex relationship and use it in a predictive manner for automatic and adaptive correction of ideal CNC code. The present research work is an attempt in this direction. Research Scholar, Email: [email protected] †Professor, Email:[email protected] and corresponding author LITERATURE REVIEW Scant research work has been concentrated on issues to study and control the part profile inaccuracies produced during CNC turning and milling processes. Several on-line measurement and adaptive control systems were reported to compensate thermally induced errors and elastic deflections caused by cutting forces employing system responses like laser scanning signals force signals etc.[2, 3, 4] These systems use complex and sophisticated measurement systems for monitoring on-line the dimensional changes making them suitable only as laboratory setups.[3] Literature reveals that recently some approaches using ANN have been reported to correct the part errors produced during CNC turning. Ziegert et. al., built an ANN tool for the mapping of nominal coordinates of cutting tools and corresponding tool point errors.[5] The simulated test results show that the error was reduced to 1 10 th of the original dimensional error. However the kinematic error model used to generate the training sample considers only the positioning error compensation.[5] The ANN system reported by Wei-Ren Chang [6] also does not consider the effect of process parameters (cutting speed, feed rate, depth of cut etc.,), tool-work material combinations, tool parameters (tool geometry), and workpiece geometry etc., on the dimensional accuracy of the part profile. It uses simulated results of a turning process model for training ANN. All the works reported so far [4, 5, 6, 7] have considered only dimensional errors and do not address form errors. This paper reports the systematic experimental investigations carried out to study the effect of process parameters on the part profile accuracy to capture the complex functional relationship between them using Artificial Neural Networks (ANN) technique. Further the trained network is used in a predictive manner to compute profile error likely to be produced on the component and revise the CNC code adaptively. The following section describes the methodology of adaptive CNC code generation. METHODOLOGY Two stages are envisaged in adaptively correcting the CNC code for improving part accuracy. Stage I comprises of training and testing phases of Artificial Neural Network (ANN) model to capture functional relationship between the profile errors and the process parameters. In the training phase ANN model is trained using input patterns in terms of process parameters , part geometry etc., and output as the form error produced during actual machining. The network is trained till the error converges to the acceptable limit. In the testing phase trained network was tested for its prediction accuracy by using a set of data. Stray samples which are causing inefficiency are often eliminated and training continues till the network has reasonably learnt the relationship between input and output variables. This phase is iterated till a satisfactory prediction accuracy is obtained. This methodology of iterative training and testing of ANN is explained at length else where.[8] Stage II is the application phase where in the trained network is used to predict the profile errors likely to be produced on the component. The ideal CNC code can be revised to account for these profile error at each tool position. This significantly improves the profile accuracy of components. Our previous research papers have presented at length this methodology.[8, 9] This extensive error correction strategy, however, increases the CNC code length compared to the uncorrected one. In addition in some cases the magnitude of errors may also be within acceptance tolerance limits in certain zones of part length. It is, thus, clear that if the error correction is adaptively applied in certain zones only, it will provide a good trade off between CNC code length and accuracy improvement. This strategy is the subject of the present research paper. In essence the predicted profile is critically analysed where the tool position needs to be corrected to reduce the error magnitude. An OOPS based software AdaptCNC is developed to predict and analyse the profile and further generate adaptively the revised CNC code. EXPERIMENTAL PROGRAM It was decided to carry out the experiments by machining various part profiles consisting of cylinder, taper, arc (both convex and concave) and compound (combination of cylinder, arc, taper profiles) profiles in the overall part size envelope of 35mm diameter and 60mm length. Pilot tests were conducted to identify stable machining conditions for all the part profiles. The components were machined on Cyclone CNC turning center with Fanuc OT-B controller. The process conditions, tool/workpiece details used for machining are as under. Cutting speed V : 60, 80, 90, 110m/min Feed rate F : 0.1, 0.15, 0.2 mm/rev. Part size : 15, 20 and 25mm for cylindrical components and 10 and 15mm at free end and 30mm at the clamped end for taper, convex, concave profiles Raw material : Aluminium (size: diameter 35mm and 100mm length Cutting tool : Insert of type CCMT with nose radius 0.2mm Various programming utilities of Fanuc controller, such as Multiple & pattern repetitive cycles, Finishing cycle, TNR compensation and constant surface speed were used in generating the ideal CNC code. 60 components were machined, in all, including those with repeated process conditions.[9] The machined components were inspected on Nikon measurescope to measure the dimensions produced. The inspection datum chosen was the same as the machining datum (part zero), so that the correction can be effected from the free end of the component. Each component profile was inspected at 25 locations along its length to measure the form and dimensional error compared to the ideal programmed geometry. For a chosen axial location, three readings were recorded by orienting the component approximately by 120 deg, to average the direct measurement errors. The average of these three readings was finally used as the process response at a particular axial location (Zpos) for training the ANN. The profile error was expressed as % error and calculated as under. ep = (Di Da) Di 100 (1) Where ep = part error (%) Di = Ideal (programmed) diameter, mm Da = Actual diameter, mm ep was calculated at each axial position (Zpos) inspected. The radial error of the profile in microns is calculated as under. er = ep Ri (2) Where er = radial profile error in μm Ri = ideal profile radius, mm The profile error data (ep) obtained after inspection of components was normalized to generate the .... Cutting speed X