2.4 Application of Soft Computing Technique
2.4.3 Cutting Force
signals from various sensors. They used two types of learning algorithm in neural network, one is adaptive resonance theory (ART) and other is self-organising map (SOM). The authors found the NN with SOM performs better than ART.
Balazinski et al. [2002] described three artificial intelligence (AI) based methods to estimate tool wear in turning. These are: (a) Feed forward back propagation neural network (FFBPNN), (b) Fuzzy decision support system (FDSS) and (c) Artificial neural network-based fuzzy inference system (ANNBFIS). The tool wear estimation was based on the cutting force, feed force and feed. Tool wear was found increasing linearly with respect to feed force. All the three AI models give a similar result in estimation of tool life. Chungchoo and Saini [2002] proposed an on-line fuzzy neural network (FNN) model for wear estimation. They found that the model is better for the estimation of average width of flank wear and maximum depth of crater wear. Fang and Jawahir [1994] presented a methodology for assessing the aspects of total machining performance encompassing surface finish, tool wear rate, dimensional accuracy, cutting power and chip breakability. They quantified the effects of influencing process parameters on total machining performance by fuzzy- set method and developed a series of fuzzy-set models to give quantitative assessments for any given conditions, including work material properties, tool geometries, chip-breaker types and cutting conditions. Several authors have applied wavelet transform of various signals for the tool wear or machining condition estimation [Luo et al., 2000; Li et al., 2005]. In many of these researches, only a limited amount of experimental validation was conducted. In some cases, the tools were artificially ground.
Spedding [1995] developed a neural network model to find out the cutting conditions for a given work material and required depth of cut to predict the cutting forces, surface roughness and tool life. With the help of a multi-layer perceptron network, Szecsi [1999] modelled the three components of cutting force in turning process as a function of process parameters, tool geometry, workpiece material and tool flank wear. He used 3200 training and 1500 testing data for designing the network. Very good prediction accuracy has been reported. However, there is no discussion about the statistical variation of the cutting forces. Ezugwu et al. [2005]
used NN model to correlate between process parameter and performance parameters.
The process parameters were speed, feed, depth of cut, cutting time and coolant pressure. The performance parameters were tangential force, feed force, consumption of spindle motor power, surface roughness, average flank wear, maximum flank wear and nose wear. The model gives a good performance and agrees well with experimental data.
A number of other authors have applied multi-layer perceptron neural network for the prediction of cutting forces in milling processes [Zuperl and Cus 2004;
Radhakrishnan and Nandan, 2005; Zuperl et al., 2006; Aykut et al., 2007]. Tandon and El-Mounayri [2001] modelled the forces in end-milling process by a multi-layer perceptron network. The modelling was limited to one tool-work material combination (HSS tool and aluminium work piece) and a total of 96 data were used for training and testing. The output of the neural network model consisted of forces as well as standard deviation. However, this model does not take into account the tool wears. Hao et al. [2006] predicted the cutting forces in self-propelled rotary tool using a multi-layer perceptron neural network. Cutting speed, feed, depth of cut and tool inclination angle were input parameters in the network and thrust force, radial force and main cutting force were the output of the network. For improvement of the performance of the network, the authors used hybrid of genetic algorithm (GA) and back propagation (BP) algorithm.
Zuperl et al. [2006] found the radial basis function neural networks superior to the multi-layer neural network in modelling of machining forces in ball-end milling.
Radhakrishnan and Nandan [2005] used neural network and regression model for prediction of cutting force. The regression model is used to filter out abnormal data
and the filtered data were used in neural network for better production. Briceno et al.
[2002] also compared a multi-layer perceptron neural network with a radial basis function neural network for the prediction of machining forces in milling. The radial basis function neural network was found to be superior to multi-layer perceptron neural network in some aspects. Considering the statistical variation, the neural networks were used for predicting minimum, maximum, mean and standard deviation of the forces.
Researchers have put a lot of effort for indirect estimation of tool wear using cutting forces. It has been observed that the cutting force values are more sensitive to tool wear than other signals such as vibration or acoustic emission [Jemielniak, 1999; Byrne et al. 1995]. Choudhury and Kishore [2000] developed a mathematical model for the estimation of flank wear and concluded that force increases linearly with tool wear land width. Flank wear has been deduced indirectly by measuring an easily measurable quantity, ratio of the feed force to vertical cutting force. Lin et al.
[2001] predicted cutting force and surface roughness using abductive neural network with speed, feed and depth of cut as input parameter. The abductive neural networks are composed of a number of polynomial function nodes organized into several layers and generate optimal network architecture automatically. It takes less iteration during training of the network.
Lin et al. [2003] employed radial basis function neural network in finding out the force-wear relationship in turning of composite materials. They compared the performance of the neural network model with a multiple regression model and observed that the neural network prediction is more accurate than the multiple regression model prediction especially when the functional dependency is nonlinear.
They have found that feed force data gives a superior result than the cutting force data. Zhou et al. [2003] presented a method for tool wear monitoring using ANN model. The tool flank wear of a CBN tool is monitored and correlated with some measured parameters such as passive (radial) force, frequency energy and accumulated cutting time and all these parameters are integrated using ANN model.
The review of the literature in the area of soft computing applications to machining indicates the great potential of the soft computing techniques for the prediction and optimization of machining performance. In particular, neural
networks have been found very suitable for the modelling of machining process.
However, the requirement of huge amount of training and testing data, the presence of the outliers and lack of one well-established standard method of training of neural network poses a challenge for the researchers.