Optimized Recurrent Neural Network-based Tool Wear Modeling in Hard Turning

Hard turning has been proved to be more effective and efficient than traditional grinding operations in machining hardened steels. However, rapid tool wear is still one of the major hurdles affecting its wide implementation in industry. Better modeling of the tool wear progression helps optimize cutting conditions and/or tool geometry to reduce tool wear, which can make hard turning a feasible technology. The objective of this study is to design a recurrent neural network (RNN)-based estimator for tool wear modeling in hard turning. The extended Kalman filter (EKF) algorithm is applied to train the network. Its structure is further optimized using a pruning algorithm. The proposed estimator is tested in modeling the cubic boron nitride (CBN) tool wear progression in hard turning. Due to its optimized recurrent network connections, the estimator can better model nonstationary and dynamical systems, such as the tool wear progression, in terms of the training speed and generalization ability. INTRODUCTION The hard turning process is defined as the single point turning of materials with hardness higher than 50 HRc under the small feed and fine depth of cut condition. It offers possible benefits over the process of grinding in the context of lower equipment costs, shorter setup time, fewer process steps, greater part geometry flexibility, and the elimination of cutting fluid use (Konig et al. 1984; Tönshoff et al. 2000). Among the available tool materials, cubic boron nitride (CBN), second to diamond in hardness and inert to steel materials, has been recommended as the best tool material and widely used for hard turning operations. However, one of the major hurdles preventing the wide implementation of hard turning in industry is the severe wear of CBN tools. The cost of hard turning tools and the tool change downtime due to rapid tool wear can impact the economic viability of precision hard turning. For a given tool and workpiece combination, the ability to estimate the tool wear as a function of cutting conditions cutting speed, feed rate, and depth of cut, is critical to the overall optimization of a hard turning process. Tool wear modeling as a general research topic has been researched by numerous Transactions of NAMRI/SME 213 Volume 37, 2009 researchers and its overall progress can be mainly classified into four categories as reviewed in Wang et al. (2008a): analytical model-based approach which models tool wear as a function of cutting conditions, cutting environment, tool geometry and property, and/or workpiece property; computational methodbased approach which applies the finite element method (FEM) to model tool wear development process; artificial intelligence (AI)-based approach which includes artificial neural networks (ANNs), fuzzy logic, and support vector machine; and parametric model-based approach such as the Taylor tool life model and the regression model. With respect to CBN tool wear modeling in hard turning, the main endeavors include analytical model-based approach (Huang and Liang 2004a; Huang and Liang 2004b); AI-based approach (Ozel and Karpat 2005); and the Taylor tool life equationbased approach (Poulachon 2001; Dawson 2002). Although analytical models can provide better insight to the underlying physical wear progression mechanisms, they sometimes are less accurate because they often apply simplifications and assumptions in their modeling development process (Scheffer et al. 2003). The FEM approach also provides insight to the process; however, it is too time consuming and not suitable for optimization using the current computing technology. Time series (Boyd et al. 1996) and regression models (Ozel and Karpat 2005) are typically less accurate when compared with ANNs. If both accuracy and speed are of interest instead of the underlying wear progression mechanisms, AI-based modeling approaches are favored for real applications. Among the AI-based approaches, ANN is a viable, reliable, and attractive approach for tool wear modeling (Chryssoluouris and Guillot 1990; Haber and Alique 2003) due to the following reasons: 1) ANNs have strong strength in modeling nonlinear process which makes them suitable for modeling the tool wear process; 2) The data driven feature of ANNs makes them powerful in parallel computing and capable of handling large amount of data; and 3) ANNs are good at fault tolerance and adaptability and suitable for modeling tool wear during metal cutting process which always subjected to noisy environment. Among the ANNs approaches, the multilayer perceptron NN (MLP) is the most frequently used one. Previous research has found that a fully forward connected NN (FFCNN) exhibits better performance in terms of the generalization ability, training speed, and structural robustness than that of an MLP (Wang et al. 2008a). The FFCNN can be further modified into a fully connected recurrent neural network (RNN), accommodating internal or external recurrent connections among neurons. Unlike an FFCNN, a RNN can store information from past states, making it more capable of modeling nonstationary and dynamical phenomena. The RNN excels the FFCNN in faster training speed and better generalization ability in modeling nonstationary and dynamical systems. The goal of this study is to develop a RNNbased estimator with application to CBN tool flank wear modeling. The extended Kalman filter (EKF) algorithm is utilized to train the network. Network connectivity optimization is achieved by disconnecting some less important connections based on a connectivity pruning algorithm. The modeling performance of RNN and optimized RNN is compared with that of an MLP, an FFCNN, and an optimized FFCNN. The RNN approach has proved to be faster, more accurate, and more robust. The study will contribute to better optimize the cutting conditions and tool geometry for hard turning as well as other metal cutting processes. The paper first introduces the theoretical background of the proposed RNN modeling approach and then compares the performance of different ANN approaches including RNN in CBN hard turning experimental studies (Huang and Liang 2004a). Finally, some conclusions are made regarding the proposed approach.