TY - GEN
T1 - Performance comparison of machine learning algorithms and feature reduction techniques for tool condition classification during end milling
AU - Binsaeid, Sultan
AU - Asfour, Shihab
AU - Cho, Sohyung
AU - Eltoukhy, Moataz
PY - 2007/12/1
Y1 - 2007/12/1
N2 - In fully automated machining environments, the immediate detection of condition state of the cutting tool that is identified as worn, chipped, or broken tool is essential to the improvement of productivity and cost effectiveness. In this paper, a multi-sensor tool condition monitoring system (TCM) via machine learning (ML) approach is proposed to identify the instantaneous condition of a multi-layer coated and multi-flute carbide end mill cutter when machining 4340 steel based on extracted sensors signal features. Using data acquisition and signal processing for force, vibration, acoustic emission, and spindle power sensor, 135 different features were extracted from all sensory signals in the time and frequency domain. Then, these features along with machining parameters are evaluated for significance using different feature reduction techniques. Two feature extraction methods were investigated; independent component analysis (ICA), and principle component analysis (PCA) and three feature selection methods were studied; Chi square, Information Gain, and ReliefF. Five classifiers were used to assess the effectiveness of the different feature reduction techniques. These classifiers were Naïve Bayes Tree, C4.5, Bayesian Network, Nearest Neighbor, and K-star algorithms. These selected ML algorithms classify not only flank wear level of the multi-flute end mill but also breakage and chipping. The criterion for selecting the best model depends on classification accuracy and simplicity of the implemented system where the number of features and sensors are kept to minimum in order to increase the efficiency of the online acquisition system. In this paper, all feature selection methods have relatively outperformed the extraction methods using ICA and PCA. In addition, the utilization of simple classifiers such as Nearest Neighbor and K-star classifier has proven to be effective in classifying tools abnormalities in end milling.
AB - In fully automated machining environments, the immediate detection of condition state of the cutting tool that is identified as worn, chipped, or broken tool is essential to the improvement of productivity and cost effectiveness. In this paper, a multi-sensor tool condition monitoring system (TCM) via machine learning (ML) approach is proposed to identify the instantaneous condition of a multi-layer coated and multi-flute carbide end mill cutter when machining 4340 steel based on extracted sensors signal features. Using data acquisition and signal processing for force, vibration, acoustic emission, and spindle power sensor, 135 different features were extracted from all sensory signals in the time and frequency domain. Then, these features along with machining parameters are evaluated for significance using different feature reduction techniques. Two feature extraction methods were investigated; independent component analysis (ICA), and principle component analysis (PCA) and three feature selection methods were studied; Chi square, Information Gain, and ReliefF. Five classifiers were used to assess the effectiveness of the different feature reduction techniques. These classifiers were Naïve Bayes Tree, C4.5, Bayesian Network, Nearest Neighbor, and K-star algorithms. These selected ML algorithms classify not only flank wear level of the multi-flute end mill but also breakage and chipping. The criterion for selecting the best model depends on classification accuracy and simplicity of the implemented system where the number of features and sensors are kept to minimum in order to increase the efficiency of the online acquisition system. In this paper, all feature selection methods have relatively outperformed the extraction methods using ICA and PCA. In addition, the utilization of simple classifiers such as Nearest Neighbor and K-star classifier has proven to be effective in classifying tools abnormalities in end milling.
KW - Feature selection
KW - Machine Learning
KW - Tool condition monitoring
KW - Tool wear
UR - http://www.scopus.com/inward/record.url?scp=84885993394&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84885993394&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:84885993394
SN - 9781627486811
T3 - 37th International Conference on Computers and Industrial Engineering 2007
SP - 2378
EP - 2389
BT - 37th International Conference on Computers and Industrial Engineering 2007
T2 - 37th International Conference on Computers and Industrial Engineering 2007
Y2 - 20 October 2007 through 23 October 2007
ER -