Surface Roughness in Grinding: On-line Prediction with Adaptive Neuro-fuzzy Inference System

An on-line monitoring and prediction of surface roughness in grinding is introduced with experimental verification. An adaptive neurofuzzy inference system (ANFIS) is used to monitor and identify the surface roughness online. The system uses a piezoelectric accelerometer to generate a signal related to grinding features and surface finish. The power spectral density (PSD) of this signal is used as an input to ANFIS, which in turn outputs a value for the on-line predicted surface roughness. Different neuro-fuzzy parameters were adopted during the training process of ANFIS in order to improve the on-line monitoring and prediction accuracy of surface roughness. Experimental validation runs were conducted to compare the measured surface roughness values with the online-predicted ones. The comparison shows that the adoption of Bell-shaped membership function in ANFIS achieved a very satisfactory on-line prediction accuracy of 91%. INTRODUCTION Automated and intelligent grinding is used in the metal working industry to produce parts of high quality surface finish and geometry. Surface roughness is one of the most important factors for evaluating workpiece quality during the finishing process because the quality of surface affects the functional characteristics of the workpiece such as fatigue and fracture resistance and surface friction. Intelligent monitoring and control of grinding, and surface roughness identification has three important roles: (i) to detect problems in surface roughness which occurs during the grinding process, (ii) to provide information to optimize and on-line control the process, (iii) to contribute to establishing the database, which is necessary to determine the best operating conditions. Direct and on-line measurement of the surface roughness is difficult. Therefore, indirect measurement of surface roughness has become an active subject in the finishing operations field. Fuzzy logic and fuzzy inference system (FIS) is an effective technique for the identification and control of complex non-linear systems. Fuzzy logic is particularly attractive due to its ability to solve problems in the absence of accurate mathematical models. Surface roughness Transactions of NAMRI/SME 57 Volume 33, 2005 modeling in grinding is considered complex process, so using the conventional techniques to model the surface roughness in grinding results in significant discrepancies between simulation results and experimental data. Thus, this complex and highly time-variable process fits within the realm of neuro-fuzzy techniques. The application of a neuro-fuzzy inference system to prediction and identification is a novel approach that overcomes limitations of a fuzzy inference system such as the dependency on the expert for fuzzy rule generation and design of the nonadaptive fuzzy set. Prediction and identification of surface roughness has been the subject of many researches in the manufacturing field. A systematic approach for identifying optimum surface roughness performance in end-milling operations, was proposed by [Yang and Chen, 2001]. They used Taguchi parameter design in identifying the significant processing parameters and optimizing the surface roughness of endmilling operations. Inasaki, [1999] introduced a monitoring and controlling system for the cylindrical grinding process with experimental verification. Indirect measurement of the surface roughness by means of an acoustic emission sensor was proposed in this study. Tsai and Wang [2001] conducted a comparative study between neural network models and neuro-fuzzy model on predictions of surface finish in electrical discharge machining. A genetic algorithmic approach for optimizing of surface roughness prediction model was introduced by [Suresh et al., 2002]. They adopted a surface roughness prediction model for machining mild steel, using response surface methodology (RSM). Behrens and Ginzel [2003] introduces a process control system consisting of a fuzzy gap-width controller adapted by a neural network in electrical-discharge machining. A new method for predicting the surface roughness of the workpiece for the grinding process was developed by [Zhou and Xi, 2002]. The proposed method takes into consideration the random distribution of the grain protrusion height. Direct on-line measurement of roughness during the grinding process is difficult. On-line laser based systems are available but they are expensive and not suitable to the grinding environment. Indirect on-line measurement of the roughness by means of a piezoelectric accelerometer is the subject of this paper. Offline measurement of roughness was demonstrated in a companion paper [Samhouri and Surgenor, 2005]. However, it must be noted that an off-line ANFIS model has different inputs and structure from what is required for an on-line application. For on-line prediction, ANFIS presents an indirect measurement of surface roughness during edge grinding process by using the frequency domain analysis of the timeseries signal generated on-line by the accelerometer as inputs to an intelligent neurofuzzy identification system. Actually, the use of indirect measures of surface roughness (e.g. grinding force, applied force or feed rate) as roughness control feedback does not guarantee constant homogenous surface roughness. By contrast, measuring the surface roughness online as controller feedback should achieve constant, homogenous, and accurate surface roughness without the need to measure other grinding parameters. INTELLIGENT GRINDING AND SURFACE ROUGHNESS PREDICTION SYSTEM Demands to increase productivity and surface quality in grinding require to lead the process in a narrow window. Consequently, Intelligent monitoring has to be introduced in a growing extent. That means sensors, signal processing, and evaluation procedures have to be implemented. In general, there are four important operating parameters in grinding: the grinding path, the feed rate, the grinding wheel speed, and the contact grinding force. In this paper, we investigate the effect of changing the grinding force and feed rate on the generated acceleration signal by a piezoelectric accelerometer mounted on a grinding robot, thus, constructing an on-line intelligent neurofuzzy (ANFIS) identification system for the surface roughness, as an attempt to measure, monitor, and control the roughness on-line. The grinding robot used in this study is a pneumatically-based gantry robot built by Raoufi and Surgenor [2003]. This robot is shown schematically in Figure 1. It was employed to grind the edges of steel blanks after being cut and prior to stamping process, in order to eliminate or reduce the micro-cracks occurred from the cutting process by improving the surface roughness, thus, to reduce the rejection rate of the final product. Raoufi and Surgenor [2003] concluded in their study that regulating the applied force in grinding Transactions of NAMRI/SME 58 Volume 33, 2005 FIGURE 1. SCHEMATIC DIAGRAM OF THE GRINDING ROBOT at a constant value, results in a homogenous surface roughness, but this conclusion was not proved experimentally by measuring the resulted roughness of the blanks edges. Therefore, this paper introduces an on-line prediction system of surface roughness in order to be used as a feedback control signal rather than using the regulated applied force indirectly to control the surface roughness in grinding. Surface roughness, which is used to determine and evaluate the quality of a product, is one of the major quality attribute of a ground product. The arithmetic average height parameter (Ra), also known as the center line average (CLA), is used in this study as a measure of the surface roughness. A contact stylus-type tester was employed to measure the surface roughness offline in order to generate an off-line training data for the ANFIS prediction system. The grinding gantry robot, shown in Figure 1, is a three degree of freedom robot which has Xdirection motion to provide the feed rate (R), and Z-direction motion to provide the applied force of grinding (F). This grinding robot is required to follow the outside edge of the blank, while applying a grinding wheel to the inner corner with a constant force. The controllers on both X and Z variables are performed by nicely-tuned position/velocity (PV) and proportional-integralderivative (PID) controllers, respectively. Extensive studies have been carried out in identifying the surface roughness of the workpiece in the grinding process. The predicted values of surface roughness based on traditional methods showed a considerable difference between the actual and predicted values. This discrepancy results from the complex and nonlinear grinding operations. This explains why the models for mapping the grinding parameters to the surface quality are hard to be established accurately. Therefore, an on-line intelligent adaptive fuzzy inference system is built in this research to identify the roughness in grinding. ACCELEROMETER DATA ANALYSIS AND PRESENTATION On-line surface roughness prediction requires generating representative and useful information about the grinding process features by means of a sensor. Our approach to measure the surface roughness on-line is to mount a piezoelectric accelerometer on the end effector of the grinding robot in order to give a time-series signal which is supposed to contain useful information about the grinding conditions and surface quality attributes. Figure 2 shows a photograph of the grinding robot end effector with the piezoelectric accelerometer mounted on it. FIGURE 2. THE ACCELEROMETER MOUNTED ON THE GRINDING ROBOT’S END EFFECTOR In the analysis of time series signals, certain restrictions are imposed by the length of the data window (T), being analyzed and by the sampling rate (fs), used when digitizing continuous data [Rogers et al., 1997]. A sample of time series segment of length T = 20 sec is shown in Figure 3. This is needed by the grinding robot to start from its home position, go to the initial position of grinding, grind the blank’s edge, and then go back to the home position. Figure 3 shows that it is difficult to get useful information about the grinding process features by only considering the time-series signal. Transformation of data to the frequency domain is done to gain more insight about the grinding process and to help identify the generated surface roughness on-line. Analyses that transform data into the frequency domain result in displays of acceleration power spectral density (PSD) versus frequency. Two pieces of information define one segment of time series data: the length of the segment, T, and sampling Transactions of NAMRI/SME 59 Volume 33, 2005 FIGURE 3. A 20-SECOND WORTH OF TIMESERIES SIGNAL DURING GRINDING time, dT, used in the acquisition data. The sampling time used must be appropriate for the data of interest because it determines the highest frequency component which can be faithfully reconstructed in spectral calculations. This value is called the Nyquist frequency, fN, where,