A neuro-fuzzy approach to the prediction and control of surface roughness during grinding

Grinding is a complex process in which many variables affect the desired surface quality. Fuzzy logic is an effective technique for the prediction and control of complex systems in the absence of accurate mathematical models. Neuro-fuzzy is a relatively new technique that overcomes limitations of the pure fuzzy approach. The adaptive neuro-fuzzy inference system (ANFIS) is one such neuro-fuzzy approach.; This thesis describes the work done to design and develop a neuro-fuzzy based system for the prediction and control of surface roughness in edge grinding. Simulation and experimental tests were conducted to validate the system. A pneumatic gantry robot was used for the grinding experiments. The work was organized into four components: (1) ANFIS tuning of a PID force controller, (2) ANFIS off-line identification of roughness, (3) ANFIS on-line prediction of roughness and (4) supervisory control of roughness with fuzzy clustering.; ANFIS was used to model the relationship between the gains of a PID force controller and the target output response as specified by the desired percent overshoot and settling time. This ANFIS based input-output model was then used to tune on-line the PID gains for different response specifications. The ANFIS based controller was able to meet the response specifications with an accuracy of 95%.; ANFIS was used to identify the roughness off-line and predict the roughness on-line. Grinding force and feed rate were used as inputs. For validation purposes, surface roughness was measured directly off-line with a roughness gauge. For on-line prediction, a piezoelectric accelerometer was used to generate an indirect measure of roughness. The power spectral density of the accelerometer signal was used as an on-line input to the ANFIS. Experiments were conducted to compare the actual and ANFIS identified and predicted values of roughness. Off-line identification accuracy was 96% and on-line prediction accuracy was 91%.; Three different fuzzy-based approaches were taken to the design of a supervisory roughness controller: (1) fuzzy-C means clustering, (2) fuzzy subtractive clustering and (3) ANFIS. It was found that fuzzy subtractive clustering (not ANFIS) gave the best combination of high accuracy (98%), low computational cycle time (0.08 s) and minimal tuning effort.