Int J Adv Manuf Technol (2002) 20:241–247 Ownership and Copyright  2002 Springer-Verlag London Limited Flank Wear Estimation in Face Milling Based on Radial Basis Function Neural Networks P. Srinivasa Pai 1 , T. N. Nagabhushana 2 and P. K. Ramakrishna Rao 3 1 Department of Industrial and Production Engineering, NMAM Institute of Technology, Karnataka, India; 2 Department of Computer Science and Engineering, S. J. College of Engineering, Karnataka, India; and 3 Department of Mechanical Engineering, S. J. College of Engineering, Karnataka, India This paper presents an estimation of flank wear in face milling operations using radial basis function (RBF) networks. Various signals such as acoustic emission (AE), surface roughness, and cutting conditions (cutting speed and feed) have been used to estimate the flank wear. The hidden layer RBF units have been fixed randomly from the input data and using batch fuzzy C means algorithm, and a comparative study has been carried out. The results obtained from a fixed RBF network have been compared with those from a resource allocation network (RAN). Keywords: Batch fuzzy C means; Flank wear; Radial basis function; Resource allocation network 1. Introduction Manufacturing industries all over the world are trying to minimise cost and maximise productivity. The efforts in this direction have led to large-scale automation, with minimum human intervention. One of the major obstacles has been the condition of the cutting tools. The cutting tool in any machining process is subjected to wear, fracture and chipping, which are the various forms of tool failure. These factors will affect not only the condition of the cutting tool, but also the workpiece and sometimes even the machine tool. Various signals including acoustic emission, vibration, or force are reported in the litera- ture, which are used widely for monitoring the condition of the cutting tool. In the last decade, neural networks have been applied to monitor machining processes. Neural networks are mathemat- ical tools, which simulate the information processing mech- anism of the human brain. Tool wear is a very complex process, and modelling of it, and understanding the relationship Correspondence and offprint requests to: Dr T. N. Nagabhushana, Department of Computer Science and Engineering, Sri Jayachamarajen- dra College of Engineering, Mysore 570 006, Karnataka, India. E-mail: tagdurhotmail.com between the various interacting parameters is very difficult. Neural networks are able to model this highly nonlinear relationship between the tool wear and the sensed signals. Neural networks learn from meaningful example data by train- ing the network via a learning algorithm. The knowledge abstracted from training is generalised for the interpretation of novel sensor signals in terms of the level of tool wear. The generalisation makes correct interpretation possible, even when information from the sensor signals is incomplete and noisy. This is a characteristic that is advantageous in the performance of the diagnostic function in a manufacturing environment [1,2]. In this work a radial basis function (RBF) neural network has been trained with acoustic emission (AE) parameters, sur- face roughness parameters and cutting conditions to predict tool wear. The hidden units were fixed using two methods: 1. Using random selection from the input data. 2. Using the batch fuzzy C means algorithm. The network was trained and the weights obtained after conver- gence was used for testing. The results of flank wear estimation obtained from this study have been compared with those obtained from a resource allocation network (RAN). 2. Machining Experiment and Data Generation Face milling experiments were conducted for evaluation of the RBF network. The experiments were conducted on a Bharat Fritz Werner vertical milling machine. En-8 (medium carbon steel) was the workpiece material and uncoated carbide insert, SEKN 12 03 AFN TTMS (WIDIA hard metal grade) was used. Cutting speed was varied in the range 70–176 m min -1 , and feed tooth -1 was in the range 0.09–0.4 mm. The depth of cut was kept constant at 0.5 mm. There were three levels of tool wear : initial (0.0 to  0.2 mm flank wear), normal ( 0.2 mm to  0.4 mm) and abnormal ( 0.4 mm). The flank wear was VB max which was measured using a tool maker’s microscope. Acoustic emission (AE) data was collected