Approaches for efficient tool condition monitoring based on support vector machine

The direct sensing method estimates tool conditions through the measurement of tool geometry directly, such as shape or position of cutting edge, optical scanning of the tool tip, electr

The major objective of a sensor-based TCM system is to determine the cutting tool conditions (such as tool wear, breakage etc.) from the sensor data Much research (Elbestawi et al., 1991; Dornfeld, 1990; Tansel and McLauglin, 1993; Wong et al., 1997) has been undertaken in these fields, since cutting tools are both an important factor in manufacturing costs and the quality of the workpiece (Pfeifer and Wiegers, 2000) Despite intensive research during the past two decades, successful and effective TCM in automated machining systems remains an engineering challenge (Li and Mathew, 1990) The developed systems often have narrow ranges of performance, require substantial training or setup time to function correctly (Byrne and Dornfeld, 1995) Therefore further research is needed

In the following section, the basic architecture of sensor-based TCM systems is

Figure 1.1 Information flow and processing scheme in TCM

1.2.1 Signal acquisition by sensing methods

Sensing is the first part of the information-driven TCM system, which provides the primary information inputs The basic requirements in the selection of sensing signals are:

1 The signals should directly or indirectly provide information that is closely related to the changes in the tool conditions

2 The signals should have high signal to noise ratio (SNR), and not interfere with the machining process

3 The acquired sensing information should indicate or detect all significant events

in the cutting process

1.2.2 Signal processing

Signal processing is the core function of the information-driven TCM system,

Tool Condition Signal Selection and

Acquisition Information Processing and Refinement Decision Making

Chapter 1 Introduction

which includes feature extraction and feature selection It basically performs a transformation process in which a large flow of sensor signals is streamlined to a compact tool-condition-informative feature vector in time and frequency domain The key challenge in this technique is to derive features, which contain not only as much tool condition information as possible, but also compact in nature

Feature extraction

Since sensed signals are typically noisy, these signals have to be further processed i.e feature extraction, to yield useful features that are highly sensitive to tool conditions The widely used feature extraction approaches include:

1 Time domain analysis such as derivative of signal (Li and Mathew, 1990), statistical value of waveform (Kannatey-Asibu and Dornfeld, 1982)

2 Advanced signal processing techniques such as neural network (Tansel and McLauglin, 1993), wavelet analysis (Tansel and McLauglin, 1993, Wu and Du, 1995)

3 Power spectrum analysis such as FFT, cross spectrum (Emel and Asibul, 1988)

Kannatey-4 Time series analysis, such as autoregressive (AR) and autoregressive moving average (ARMA) (Liang and Dornfeld, 1989)

Feature selection

Feature selection is to select an optimum subset of features from potentially useful features which are available in a given problem domain (Gose et al., 1996) It is a challenging task to select the characteristic features that not only represent the characteristics of the process (information), but also contains less noise This method outputs a subset of all available features, therefore, the dimensionality of the final input feature set may be reduced Its intention is not only to discover all the features

Chapter 1 Introduction

relevant to the concept and determine how relevant they are, but also to find a minimum feature subset for effective classification with good generalization performance In addition, feature selection may also speed up the classifier for time critical applications, and make feature discovery possible

The optimum feature subset has been defined as the subset that performs the best under a classification system (Jain and Zongker, 1997) "Performs the best" here may

be explained in two slightly different ways:

1 The subset of features which gives the lowest classification error (an unconstrained combinatorial optimization problem); or

2 The smallest subset of features for which the classification error proportion is below a set threshold (constrained combinatorial optimization) (Siedlecki and Sklansky, 1988)

The latter is widely employed in many practical applications including this research

1.2.3 Decision making

The decision-making strategy is to map the signal features to a proper class (machining tool conditions) i.e pattern recognition (Li and Mathew, 1990) The output of the decision-making process includes one or more of the following:

1 Identification of tool conditions (such as tool wear/breakage etc.)

2 Evaluation of the severity of certain abnormal tool conditions

3 Prediction of tool conditions and control of machining process

This research focuses on the first item, i.e binary tool conditions identification (fresh or worn) and multiclassification of tool conditions (sharp, workable and worn) The robustness of decision-making depends not only on the identification techniques,

1.3.1 Overview of sensing method

To achieve greater reliability and robustness in turning operation, both single and multiple sensing, coupled with various signal processing and pattern recognition techniques, have been investigated for single or multiple tool condition identification

As aforementioned, the potentially most economical scheme for TCM is to employ a single-sensor approach for multiple tool conditions identification from the viewpoint

of information utilization

The sensing methods in TCM can be categorized into direct or indirect methods according to the signal obtained (Micheletti et al., 1976) The direct sensing method estimates tool conditions through the measurement of tool geometry directly, such as shape or position of cutting edge, optical scanning of the tool tip, electrical measurement of the contact resistance between the tool and workpiece, and radioactive analysis of the chip, analyzing the vision of the tool, measuring the volume of wear particles or the distance between workpiece and tool or tool holder The limitation of these methods lies in that it is difficult to collect the relevant

Chapter 1 Introduction

information under actual cutting process

The indirect methods are those concerned with detecting some process-borne signals about tool wear and establishing the relationship between these signals and tool wear (Elbestawi et al., 1991) Indirect methods include measurement of cutting force (Elbestawi et al., 1991; Hong et al., 1996; Santanu et al., 1996; Bao and Tansel, 2000), acoustic emission (Diei and Dornfield, 1987; Sampath and Vajpayee, 1987; Liu and Liang, 1991; Zizka, 1996; Wilcox et al., 1997; Niu et al., 1998; Xu, 2001), vibration of tool or tool post (Lee et al., 1987; Elwardany et al., 1996; Moore and Kiss, 1996; Li and Dong, 2000), ultrasonic vibration (Ultrasonic Energy) (Hayashi et al., 1988; Coker and Shin, 1996; Abuzahra and Yu, 2000), acoustic wave (sound) (Takata

et al., 1986), current of spindle or feed motor (power input) (Matsushima et al., 1982; Rangwala and Dornfeld, 1987; Altintas, 1992; Lee et al., 1995) and optical signal (Cuppini et al., 1986; Oguamanam et al., 1994; Wong et al., 1997) These indirect methods have the advantages of less complexity and suitability for practical application (Byrne and Dornfeld, 1995), thus they have been used by many researchers Of all the signals, acoustic emission (AE), and cutting force are most commonly used An introduction about them is provided as follows

AE sensing

AE signals reflect the microscopic activities (friction, fracture etc.) of the cutting process It naturally contains multiple tool condition information such as tool wear, fracture etc Through proper processing, it can be more economically (compared with multi-sensing approach) used for multiple tool condition identification The merit of using AE to detect tool wear lies in its frequency range is much higher than that of the machine vibrations and environment noises (Sata et al., 1973) Hence, relatively precise signal can easily be obtained by applying high-pass filter Moreover, AE can

Chapter 1 Introduction

be obtained by using a piezoelectric transducer mounted on the tool holder which does not interfere with the cutting operation, thereby makes continuous monitoring tool condition possible However, other researchers held a different idea They believed that AE signals cannot be independently used to provide reliable tool wear detection in TCM Blum and Inasaki (1990) performed experiments to determine the relationship between flank wear and AE signals They were particularly interested in the use of the AE mode, a parameter describing the ‘whole’ characteristics of the cutting process, and then concluded that extracting tool wear information from the AE signal was difficult The reason causing the two opposing views did not lay on the sensing technology, but on the ensuing analysis (Lister, 1993) Based on this opinion, this thesis first discusses the application of AE signals in TCM system when steel is used as workpiece

Cutting force sensing

Measuring cutting forces is one of the most common techniques to monitor tool condition, since they are more sensitive to tool wear than vibration or power measurements (Lee et al., 1989) The reliability of force measurements is another factor for their popularity in tool wear monitoring applications

As a cutting tool shears the workpiece, high stresses and strain rates give rise to forces with dynamic behavior across a broad spectrum of frequencies The relationship between tool wear (e.g flank wear and crater wear) and increasing cutting force is well known for a long time (Dornfeld, 1990; Oraby and Hayhurst, 1991; Lee et al, 1992; Ravindra et al., 1993b; Tarng et al., 1994)

Although many investigators agreed that the change of cutting forces represents an accurate and reliable approach to estimate tool condition, they still argued which component is the most sensitive; dynamic component, static component, or both of

Chapter 1 Introduction

them Cutting condition is also an argued issue Cuppini et al (1990) implemented a continuous monitoring method and established relationships between wear and cutting power without cutting conditions While, Choudhury and Kishore (2000) believed that cutting speed, feed and depth of cut should be taken into account in tool condition recognition

This work has tried to clear up the above arguments according to cutting force from titanium machining

1.3.2 Overview of signal processing

AE signal processing

Due to AE signals’ high frequency nature and sensitivity to the micro-structural behavior of material, it is widely employed to extract the useful information in TCM Iwata and Moriwaki (1977) pioneered the method of using AE signals to monitor tool wear condition in a cutting process, and they found that the power of spectrum of

AE signals up to 350kHz increased with tool wear and then it reached saturation Since the AE signal associated with the tool flank wear is stationary in nature, fast fourier transform (FFT) is still the best tool for the analysis of this type of signals The spectral density of AE signals has been found to be the most informative feature for TCM in turning (Emel and Kannatey-Asibu, 1988) Naerheim and Arora (1984) used continuous and discontinuous AE in turning operations to test gradual wear and intermittent degradation of cutting tools, respectively Roget et al (1988) concluded that AE parameters such as root mean square (RMS), mean, and peak values and their corresponding variance, kurtosis and skew could provide sufficient warning information of tool breakage and tool wear in various cutting condition Jemielniak and Otman (1988) considered that the skew and kurtosis to be better indicators of tool

Chapter 1 Introduction

failure than RMS values Another approach for improving the reliability of the wear related AE signal was proposed by Blum and Suzuki (1988) A feature called “AE mode” has been observed to be quite sensitive to tool wear condition Time series analysis focuses on the stochastic nature in the dynamics process of AE generation Liang and Dornfeld (1989) employed time series modeling techniques to extract AE features such as autoregressive (AR) parameters and AR residual signals for testing and monitoring tool wear Moriwaki and Tobito (1990) proposed statistical features (mean, variance and the coefficient of RMS) as inputs of a pattern recognition system

to identify and predict the ensuing tool life for coated tool life estimation Zheng et al (1992) used an optic fiber sensor and a commercially available PZT AE sensor to conduct drilling and milling operations Results from the two experiments showed a reasonable degree of agreement Using AE features, König et al (1992) performed tests to monitor small drillings and detect tool fracture, and reported that AE features were sensitive to tool chipping Dornfeld (1992) presented compelling reviews on the application of AE sensing techniques in tool wear detection in machining He observed that the changes in skew and kurtosis of AE RMS signals could effectively indicate tool wear Kakade et al (1994) reported that AE parameters (ring-down count, rise time, event duration, event rate and frequency) could distinguish clearly the cutting action of a sharp and worn or broken tool Kannatey-Asibu and Dornfeld (1982) found that the changes in skew, kurtosis and of the AE RMS signal effectively indicate the tool wear in machining Kamarthi et al (1995) considered that AE features extracted by the wavelet transform were very sensitive to gradually increasing flank wear The magnitude of the AE in the frequency domain was employed by Li and Yuan (1998) to monitor the change of tool states Choi et al (1999) fused AE and cutting forces to develop a real-time TCM for turning operations

Chapter 1 Introduction

The recorded data were analyzed through a fast block-averaging algorithm for features and patterns indicative of tool fracture Similar work was conducted by Jemielniak and Otman (1998), who used a statistical signal-processing algorithm to identify RMS, skew and kurtosis of the AE signals and detect catastrophic tool failure Inspection of the results indicated that the skew and kurtosis were better indicators of catastrophic tool failure than the RMS values

Cutting force processing

Shi and Ramalingam (1990) investigated the feasibility of different force components, and observed that the feed force to cutting force ratio was sensitive to flank wear but insensitive to process changes (cutting speed and depth of cut) Dornfeld (1990) and Ko and Cho (1994) focused on the dynamic characteristic of the cutting force, due to the friction variation between tool and workpiece in tool wear process Oraby and Hayhurst (1991) developed a model to build the relationship between the feed force, radial force and flank wear in a turning operation Elbestawi

et al (1991) employed FFT to compute the sensitivity of cutting harmonics (cutting force signals) to flank wear Lee et al (1992) found that the components of feed and tangential dynamic force bore a good relationship to flank wear trend Lister (1993) analyzed the power spectra of dynamic cutting forces and found that the power level

of certain frequency band increases with tool wear In the orthogonal milling, Caprino

et al (1996) concluded that both the horizontal and vertical forces undergo large variations with tool wear Lee et al (1989) analyzed the dynamic force signals of a coated grooved tool by FFT, and found that the percentage increase of dynamic tangential force could give a promising threshold for the prediction of tool failure Choudhury and Kishore (2000) observed that the ratio between the feed force and cutting force provided a practical method to quantify tool wear in turning Dimla and

Chapter 1 Introduction

Lister (2000b) reported that both static and dynamic cutting forces were effective in TCM The former was the most sensitive to cutting condition changes, while the latter was good at tracking changes of tool wear He also observed the vertical components (z-direction) of both cutting forces and the vibration were the most sensitive to tool wear Generally, investigators agreed that the cutting forces could provide a good and reliable indication of tool conditions, although differed in the relative effectiveness of cutting force components

From the above review, both AE and cutting force are widely used in TCM and different extracted features from them are preferred in their individual TCM systems However, neither has been shown to always provide better identification performance than the other Also, features collected from the independent data sources are not equally informative as certain features may correspond to noise, not information; others may be correlated or not relevant for the task to realize, since any technical indicators and statistical information related to the tool state can also be used as the predictors Furthermore, the success of a decision making method depends largely upon how well the monitoring features describe the characteristics of the process conditions Thus, selecting a suitable feature set is the critical factor in TCM

1.3.3 Overview of decision making

The methods employed in this field can be classified into three categories: based method, statistical-stochastic method and artificial intelligence approach

model-Model-based method

Model-based method sometimes can be viewed as heuristic-based rules with apriori knowledge only of the process parameters (Dimla and Lister, 2000b), e.g

signals, Liang and Dornfeld (1989) developed an AR time series model to classify

cutting tool conditions AR parameters in a predefined function form could be adaptively modified with a stochastic gradient algorithm, so as to provide correlation information Danai and Ulsoy (1987) developed a linearized version of Koren and Lenz’s flank wear model (1972) so as to separate the effects of cutting conditions on measured force Lin et al (1996) built a regression model based on experimental data

to estimate tool wear, which is a function of average chip thickness, tooth number, normal force coefficient and friction force coefficient Abuzahra and Yu (2000) presented the relation between the acoustic behavior of ultrasound waves and the progressive tool wear in a mathematical form In high speed cutting processes, Molinari and Nouari (2002) modeled the diffusion wear so as to optimize the cutting processes in terms of tool life

However, all of these methods suffer from two significant limitations in industrial applications First, machining process is a non-linear time-variant system, which is difficult to model and correlate the actual tool status with the physical variables from monitoring system And secondly, the signals obtained from sensing are dependent on

Houshmand et al (1991) implemented linear and quadratic discriminate analysis of multivariate AR processes in AE spectral components Li et al (1999) used the AE and feed current signals to detect tool breakage by discrete wavelet transform In order to predict tool state, the envelope detection method is employed to calculate the second difference of each wavelet coefficient for comparison with the tolerance threshold

Nevertheless, their research produced good results in certain limited experimental situations First the prior assumption and its threshold scheme make the system sensitive to the different tool-machine-workpiece combination, which restrict its applications Then, the limitation of the inherent complexity in tool-wear tribology and its variability over numerous cutting conditions make this method far from practical and reliable Recently, these strategies have been widely integrated with artificial intelligence as the signal interpretation method

Artificial intelligence approach

The emergence of artificial intelligence techniques has seen their enormous applications to TCM system, such as simple decision logic (Rao, 1986; Altintas, 1992; Elbestawi et al., 1991; Wu and Du, 1996), pattern recognition (Hirotoshi et al.,

Chapter 1 Introduction

1993; Diniz et al., 1992; Ravindra et al., 1993a; Colgan et al., 1994), fuzzy logic (Li and Elbestawi, 1996a), wavelets (Du, 1995a; Niu, 1998), and NNs (Rangwala and Dorneld, 1990; Chryssolouris, 1992; Tansel and McLauglin, 1993b) Among them, neural networks (NNs) are the most popular and successful tools

There is extensive literature about the application of NNs in this field (Rangwala and Dornfeld, 1987; Dornfeld, 1992; Tarng et al., 1994; Hong et al., 1996; Tansel, 1993; Santanu et al., 1996; Wong et al., 1997; Niu et al., 1998; Xu, 2001)

In NNs’ learning process, synaptic weights can be adjusted in an interactive process In terms of this character, the learned knowledge is usually distributed over a large number of neurons, and can be retrieved almost instantaneously in practical application NNs can also perform decision making based on incomplete and noisy information, which makes it suitable for the diagnostic function in a manufacturing system (Rangwala and Dornfeld, 1990)

Rangwala and Dornfeld (1987) pioneered the use of Back-propagation (BP) to classify AE and force signal for tool wear monitoring Up to 97% reliability was achieved in identifying the worn state of a turning tool In order to compare the learning abilities, Chryssolouris (1992) simulated tool wear monitoring using both statistical fusion approaches and BP method Their results have shown that neural

network is superior to statistical fusion approach in TCM Using AE and cutting force

signal, Leem and Dornfeld (1995) designed a customized neural network for sensor fusion in on-line detection and achieved high accuracy rates with robustness in classifying tool wear to two and three levels After tool wear levels were topologically

ordered by Kohonen’s Feature Map, input features of AE and force were transformed

via input feature scaling Niu et al (1998) applied a local wavelet packet

decomposition method to analyze AE signals in turning process, and separated the

Chapter 1 Introduction

signals into transient and continuous components To identify tool wear status, six features (mean and standard deviation of skew, kurtosis and bandpower) from spectral and statistical analysis techniques were used as inputs to adaptive resonance theory (ART2) network Silva et al (1998) developed two types of NNs (self-organizing map (SOM) and ART2) to classify tool wear in terms of 15 features collected from five sensors In order to improve the two networks’ performance, an expert system on the basis of Taylor’s tool life equation was used to identify and eliminate outlier Li et al (2000) implemented the Multiple Principal Component (MPC) and Fuzzy Neural Network (FNN) for TCM Force, vibration, and spindle motor power signals were fused in MPC to give a highly sensitive feature space, and the flexible structure of decision tree and the uncertainty measurement of fuzzy logic were utilized to perform decision making Xu (2001) applied Radial-Basis Function (RBF) network to perform real time monitoring of tool wear, in which an unsupervised Kohonen map was used

to select self-organized centers

Despite some successful applications and satisfactory characteristics, these algorithms also have their weaknesses such as a large number of controlling parameters, generalization problem (over-fitting problem, local minima) For instance,

BP learning algorithm susceptibly stays in local minima and converges slowly in large-size problems Unfortunately, both AE data and cutting force data in TCM are complex and involve large flow of information Due to the use of clustering theory and empirical risk minimization, RBF usually suffers from the low generalization performance and demanding computational task on testing samples (Zhang, 2002) Since no prior knowledge of tool wear is utilized in unsupervised algorithms such as ART2, these algorithms impose a great challenge for feature extraction techniques (Niu et al., 1998) However, extracting compact tool-wear-sensitive but condition-

Chapter 1 Introduction

independent features is still an ongoing research issue

In this thesis, support vector machine (SVM) is proposed to learn the correct tool wear information in the extensive cutting conditions Compared with other learning algorithms, the SVM possesses a firm background and excellent features, such as minimizing the system complexity, yielding a significant gain in classification accuracy

In conclusion, most TCM methods employ indirect sensing which generally reflects certain physical characteristics of the cutting process The needs for reasonable on-line TCM demand high requirements on the efficiency of feature extraction techniques This refers to using the minimal computation task to derive the most complete informative tool-condition correlated features NNs based on knowledge learning and prediction is the most popular method to perform decision making on tool conditions These issues are further discussed in great details in the following chapters

The value of this research is to improve the application of NNs-based methods in TCM so as to realize reliable tool condition identification over a range of cutting conditions

1.4 Research objective and contributions

The objective of this work is to improve tool condition identification so as to achieve efficient and reliable TCM for industrial applications It integrates the above mentioned research progress and identifies the key problems in applying NNs to TCM

In a TCM system, various features from suitably processed sensing signals are

Chapter 1 Introduction

utilized by researchers However, not every feature is equally informative in a specific monitoring system Hence, the issue of feature selection is discussed, and a method is introduced to identify important features

When NNs are utilized to identify tool states in machining processes, the main interest is often on the recognition ability Nevertheless, a higher classification rate from pattern recognition does not agree with the lower manufacturing loss in practical manufacturing systems Thus, a new evaluation function is proposed by manufacturing loss consideration so that the recognition ability of TCM can be considered by accounting for the economic impact more reasonably A nonstandard

NN method is then utilized to perform the recognition task

In metal cutting, there is a different between rough cutting and finish cutting In rough cutting, the main consideration is to effectively remove material to shape that is close the desired dimension In such process, the surface roughness is not important

In contrast, finishing cut requires precise cutting and stringent surface roughness requirement These imply that the wear limit for roughing and finishing is different, and the latter is usually less than that of the former In other words, a worn tool in finishing could still be used in roughing Hence, there is a need to differentiate different category of tool wear conditions Therefore this tool condition identification method is extended to multiclassifying tool conditions Finally, a framework which generalizes sensing signal selection, feature analysis, performance evaluation and decision making is proposed in this study Two case studies are provided to demonstrate the above proposed methods: one based on AE signal from machining steel and the other based on cutting force from machining titanium

In short, the major contributions of this thesis include:

(1) Develop a method to identify feature set from various extracted features

by reducing potential manufacturing loss

(5) Suggest a procedure to select an effective training data set in TCM system

(6) Propose a framework to generalize sensing signal selection, feature analysis, performance evaluation and decision making

1.5 Organization of thesis

Chapter 1 gives a literature review of TCM scheme First, the importance of TCM

in unmanned manufacturing system is introduced; then the basic components of TCM sensing method, signal processing and decision making) are described; finally a detail literature review about them is presented

Chapter 2 introduces four kinds of tool wear mechanisms (abrasive wear, adhesive wear, diffusion wear, fatigue wear), two types of tool wear according to the wear position (flank wear, crater wear), tool life and its criteria The generation of AE and cutting force signals from a tool wearing process are also discussed

The experimental setup is introduced in chapter 3, followed by the design of experiment A detail introduction about the AE sensing, cutting force sensing, tool inserts and workpieces used in this study is presented In the design of experiment, three levels of cutting speed and feed, and two levels of workpiece and cutting depth are selected in the provided case study Therefore, only nine experiments are needed

to study the entire machining parameter space using L9 orthogonal array

Chapter 1 Introduction

In chapter 4, an intelligent feature analysis method, which integrates feature extraction and feature selection, and decision making was designed and implemented This study comprehensively takes all these known signal features and aims to identify the most effective set that can give robust and reliable identification of tool condition Automatic relevance determination (ARD) approach coupled with SVM is employed

to rank these features, prune redundant information so as to realize effective identification of tool flank wear This method can not only reduce the data processing requirement as fewer feature sets are involved, but also provide robust and feasible recognition in an efficient way AE from machining steel is utilized as sensing signals and the commonly extracted AE features are adapted to discuss this subject

Chapter 5 firstly discusses the existing problems when NNs and related methods are used to classify tool conditions Then two kinds of losses caused by misclassifying tool states are analyzed: one is using a worn tool to machine workpiece; the other is the early replacement of a workable tool Based on the relationship of two kinds of losses, a new evaluation performance is proposed as the criterion to evaluate the recognition performance of TCM system Finally, a modified SVM approach with two regularization parameters is employed which can adjust the recognition ability for each tool condition separately Experimental results show the proposed approach can reduce the overdue prediction and lower potential manufacturing cost The following topics are also involved: generalization performance, evaluation criteria, training data selection and parameters tuning in tool condition identification An effective training data selection procedure is proposed in this chapter, which leads to a considerable reduction in the necessary training samples without the loss of classification performance Experimental results also show that this selection strategy can provide

Chapter 7 presents a framework to generalize sensor signal selection, feature analysis, performance evaluation and decision making in TCM system The application of this framework for candidate signals (AE and cutting force) in titanium machining is discussed Firstly, three basic requirements to sensor signals are proposed, and feature analysis is utilized to select the most effective feature set to identify tool conditions From the analyzed results, cutting force sensing is considered as a suitable monitoring signal, and its effective feature set is used as input for tool condition identification in titanium machining Then, a performance evaluation function with manufacturing loss consideration is introduced to determine whether a tool is due for replacement other than merely the extent of wear of the cutting tool Finally, a modified SVM algorithm

is utilized to minimize this evaluation function Experimental results demonstrate that this study provides an effective and reliable framework to implement tool condition identification in titanium machining It can also be exploited to monitor other cutting processes

Conclusions are given in chapter 8 together with recommendation for future research

Chapter 2 Tool Wear and Sensing Signals

CHAPTER 2

Tool Wear and Tool life

2.1 Tool wear mechanism

In order to achieve an economical tool life, various tool angles, cutting speeds, and feed rates are adopted in metal cutting However, the accurate tool life estimation is considered more important in improving the productivity of computer-integrated machining systems (Kramer, 1986) In practical machining processes, because of the non-homogeneities in both tool material and workpiece, not only uniform wear but also other unexpected wear states can be observed (e.g breakage, chipping) In order

to estimate tool state accurately, the fundamental forms of tool wear mechanism and tool failure are introduced

The basic mechanisms of tool wear are controlled by the cutting conditions (cutting speed, feed rate, etc), the mechanical and physico-chemical properties of the work piece and tool Tool wear is also the result of load, friction, and high temperature between the cutting edge and the workpiece In the following section, some major forms of tool wear mechanism are described, such as abrasion, adhesion, diffusion and fatigue

Abrasive wear is a primary wear mechanism in metal cutting Debris, which is created by plastic deformation, enters the clearance space between two moving surfaces and acts like cutting tools to remove materials from the surface Friction between the chips and the cutting edge leads to the formation of build-up edge When

Chapter 2 Tool Wear and Sensing Signals

this edge reaches an unstable size, it breaks away in small pieces or fracture (in the latter case, it causes damage to the cutting tool – adhesive wear) The ability of cutting edge to resist abrasive wear is related to its hardness

Adhesive wear occurs mainly at low temperature Excessive load and low speed often reduce the oil film thickness to a point where intimate metal-to-metal contact occurs, thus adhesive wear is formed

Diffusion wear mainly affected by chemical factors during the cutting process, is characterized by the smooth worn surface without plastic deformation It is produced when the tool or more specifically the atoms on its surface are diffused and are carried away in the form of chips The material transfer towards the chip eventually leads to the forming of a crater on the tool rake face Generally, higher temperature may result

in faster tool wear

Fatigue wear occurs as continuous sliding, rolling, impacting motions subject a surface to repeat stress cycling The stress cycle starts with small cracks on or near the surface, and then the cracks eventually become large enough to cause discrete regions near the surface to be ejected as debris Leading causes of fatigue wear include insufficient lubrication, lubricant contamination, and component fatigue

2.2 Tool wear and tool life

2.2.1 Tool wear

In this section, two types of tool wear according to the wear position are introduced: flank wear and crater wear Figure 2.1 shows a typical demonstration of flank wear and crater wear at the cutting edge This kind of classified visual angles is used throughout the whole thesis

In a normal machining process, the motion of the tool’s flank face against the

Chapter 2 Tool Wear and Sensing Signals

surface of the workpiece causes a “wear land” on the flank of the cutting tool, named flank wear Flank wear, resulting from the combined effect of abrasive wear and adhesive wear, is found to increase steadily with cutting time and speed When tools are used under economical conditions, the flank wear is usually the controlling factor (Boothroyd and Knight, 1989)

Figure 2.1 Typical demonstration of tool wear (Boothroyd and Knight, 1989)Crater wear caused by a chemical interaction between the hot chip, workpiece and material, is characterized by a concave wear pattern on the rake surface of an insert Under high-speed cutting conditions, crater wear is often the key factor to determine the life of a cutting tool Excessive crater wear weakens the cutting edge, inhibits proper chip flow, increases heat and pressure on the tool and eventually leads to tool fracture The crater depth KT is the most commonly used parameter in evaluating the rake face wear, KB and KM are the crater width, crater centre distance, respectively Both flank wear and crater wear belong to gradual wear, while chipping and breakage are two kinds of catastrophic wear Chipping happens when the edge line

Chapter 2 Tool Wear and Sensing Signals

breaks away from a tool’s cutting edge, rather than wear The chipped pieces may be

very small, or relatively large Intermittent cutting and thermal fatigue are key reasons

of chipping, meanwhile gross inconsistencies in the workpiece material composition

or structure may also cause chipping

2.2.2 Tool life

ISO 3685 (International Standard, 1993) defines tool life as the time elapsed until a defined amount of wear has occurred in the rake face or flank face of the cutting tool When a tool is used under normal cutting conditions, flank wear is usually the primary factor that determines the life of an insert, while crater formation is more important under high-temperature and high speed cutting conditions (Boothroyd and Knight, 1989) In this experiment, due to cutting conditions within the normal range, only flank wear is considered to determine tool life This is also because of the more direct influence that flank wear has on the accuracy of the product

In practical machining operations the flank wear is not uniform along the active cutting edge; therefore it is necessary to specify the locations and degree of the wear

As shown in Figure 2.1, on the active cutting edge, the flank wear at the tool corner tends to be more severe than that in the central part, because of the complicated flow

of chip material in that region The width of the flank wear land at the tool corner (zone C) is designated as VC At the opposite end of the active cutting edge (zone N)

a wear notch often forms, because the workpiece tends to be work-hardened from the previous processing operation in this region The width of the wear land at the wear notch is designated as VN

In the central portion (zone B), the wear land is fairly uniform The average land width in this region is designated as VB or VB, and the maximum wear-land is

Chapter 2 Tool Wear and Sensing Signals

designated as VBmax

The criteria recommended by the ISO 3685 for sintered-carbide tools are:

VB=0.3 mm or; VBmax=0.6 mm if the flank is irregularly worn

2.3 AE signals and tool wear

AE signals which reflect the microscopic activities (friction, fracture etc.) during cutting processes and naturally contain multiple tool condition information (tool wear, fracture), have been proven to effective in TCM Compared with multi-sensing approach, AE sensing can be more economically Thus, it is used as references for developing tool condition identification system

During metal cutting the workpiece undergoes considerable plastic deformation as the tool pushes through it Within the deformation zones, the low amplitude, high frequency stress wave generated by a rapid release of strain energy is commonly referred to as AE Other sources of AE include phase transformations, friction mechanisms (tool-workpiece contact), crack formation and extension fracture

Figure 2.2 Schematic illustration of a two-dimensional cutting process

Figure 2.2 shows a typical orthogonal machining operation The primary AE source is the shear zone since shear is accompanied by large scale dislocation motion which

Chapter 2 Tool Wear and Sensing Signals

releases high frequency, continuous type of AE signals The amplitude level of this

AE component is dependent on the shear rate and volume of the plastically deforming zone Chip fracture produces high amplitude decaying bursts in AE signals, whereas the friction between the tool-chip interface and tool-work-piece interface produces lower frequency AE which varies with the contact area and relative sliding velocities

AE signals from the cutting process consist of continuous and transient (or burst) signals, each with distinct characteristics The continuous-type AE signals are generated in the shear zone, the tool-chip interface and the tool flank-workpiece interface The discontinuous transient signals are generated due to tool fracture, chipping or chip breakage (Emel and Kannatey-Asibu, 1988)

Progressive tool flank wear: In the fresh tool stage, AE signals mainly reflect the

plastic deformation in the shear zone A continuous-type AE signals are produced which is similar to stationary random noise In the gradual tool flank wear stage (figure 2.3 and 2.4), the main source of AE is due to the rubbing between tool flank and workpiece The typical AE produced during gradual tool flank wear is also of a stationary feature In the intensive tool wear stage (as shown in figure 2.3 and 2.5), although the AE signals are still stationary, both the magnitude and frequency range become larger than those in the gradual wear stage

Tool chipping: The chipping of a cutting tool generally refers to a small amount of

damage to the tool edge, and their corresponding AE is a short-impulse signal (as shown in figure 2.6 and 2.7) One point worth noting is that the range and magnitude

of the frequency components at the location of the chipping-induced burst may vary with the level of tool chipping

Chapter 2 Tool Wear and Sensing Signals

Chip breakage: When cutting with a grooved insert under the normal cutting

conditions, chip breakage is a significant AE source, because the grooves on the rake face of the insert acts as a chip breaker This kind of AE contains one or several small bursts The existence of chip breakage can affect the reliability of the tool flank wear monitoring, since their corresponding AE signals may become more dominant than that of flank wear source

Tool fracture: When tool fracture occurs, much of the tool edge is broken away

such that a strong AE burst-signal is produced This AE burst has very large and very wide frequency-band components at the location of the burst with its intensive energy (as in figure 2.6 and 2.8)

intensive tool wear

Figure 2.3 AE signals, gradual and intensive tool wear

Figure 2.4 Image of gradual tool wear Figure 2.5 Image of intensive tool wear

Chapter 2 Tool Wear and Sensing Signals

Figure 2.7 Image of tool chipping Figure 2.8 Image of tool fracture

In summary, typical AE signals from the tool flank wear process are of stationary

or continuous nature, and AE signals from the tool fracture, chipping and chip breakage include transient type signals Generally, the AE signals from machining process may contain both types because of the co-existence of different tool conditions

2.4 Cutting forces and tool wear

Understanding the basic force trend in the metal-cutting process that is related to tool wear or failure will enable one to know tool conditions The variation of cutting forces to tool wear has been widely established (Oraby and Hayhurst, 1991; Lee et al., 1992; Lister, 1993; Ravindra et al., 1993b; Dimla and Lister, 2000) As the cutting tool ploughs through the workpiece, the resulting cutting force can be resolved into

Chapter 2 Tool Wear and Sensing Signals

the tangential (z-axis, Ft), feed (x-axis, Fx) and normal (y-axis, Fy) components (Figure 2.9) In cylindrical turning, they are defined as follows:

Tangential force: This acts in a direction tangential to the revolving workpiece

and represents the resistance to the rotation of the workpiece In a normal operation, tangential force is the highest of the three forces and accounts for about 98 percent of the total power required by the operation

Feed force: Feed force acts in the direction parallel to the axis of the work and

represents the resistance to the feeding tool This force is usually about 50 percent as great as tangential force Since feed velocity is usually very low in relation to the velocity of the rotating workpiece, this force accounts for only about 1 percent of total power required

Radial force: Radial force acts in a radial direction from the center line of the

workpiece The radial force is generally the smallest of the three, often about 50 percent as large as feed force Its effect on power requirements is very small because velocity in the radial direction is negligible in cylindrical turning

In practice, application of these cutting forces has been concentrated on studying the dynamic characteristic and interpreting its relation to tool wear levels This can largely be attributed to the fact that the dynamic characteristic becomes important in worn tool conditions as a result of friction variation between cutting tool flank and workpiece Hence, existing force-based TCM systems typically collect absolute force levels, and measure the relative change of force as a new tool wears (Gould, 1998)

Chapter 2 Tool Wear and Sensing Signals

Figure 2.9 Forces acting on a cutting tool in orthogonal metal cutting

Chapter 3 Sensing Signals and Experimental Design

Mic ros c ope

Chapter 3 Sensing Signals and Experimental Design

The original signals from AE sensor are first amplified by 20/40 dB gain, then pass through band-pass filter (200 kHz to 1 MHz) to minimize the influence of noise signals and aliasing errors The filter and amplifier are integrated into a AE Piezotron Coupler 5125B (as shown in figure 3.4)

Figure 3.2 AE sensor 8125B Figure 3.3 Mounting of AE sensor

The refined signals are fed into a HP VXI data acquisition system which consists of

a HP E1429B digitizer and two HP E1562B SCSI disks Firstly, the sampled data are stored in the hard disks via a local bus on real time basis, and then transferred to a computer through a VXI link The maximum sampling frequency of this system is 2MHz

Figure 3.4 AE sensor 8152B and coupler 5125B

HP VEE software is used to manage and control the E1429B digitizer and E1562B SCSI disks as well as the data acquisition process The HP VEE is a visual programming language which has strong instrument control capability A user-friendly data acquisition system using HP VEE has been developed as Figure 3.5

Chapter 3 Sensing Signals and Experimental Design

Figure 3.5 Data acquisition window using HP VEE

The experiments are conducted on an OKUMA CNC lathe The tool flank wear value is measured by an OLYMPUS microscope and the tool insert images are captured by a PANASONIC digital camera All the components needed in this experiment are listed in table 3.1

Table 3.1 Components of experimental system

1.OKUMA LH 35 – N turning machining

2 Kistler AE Sensor 8152B and Couple 5125B (AE measurement)

3 HP VXI System: E1483A, E1429B, E1562B, E1401B high power mainframe

4 Lab DC Power Supply

5 Olympus measuring microscope STM and Panasonic digital camera WV-CL350/G

6 Computer: PENTIUM-S CPU 90MHz

7 Sony CD writer, empty CDs and two external hard disks

Chapter 3 Sensing Signals and Experimental Design

The developed experimental scheme may allow for the continuous acquisition of the sensor signals, and ensure no loss of sensing information during the cutting process

So the complete information reflecting the cutting process characteristics is gathered

3.1.2 Force sensing

The experimental set-up for condition monitoring is illustrated in figure 3.6 and its components are included in table 3.2 The force sensor, KISTLER 9121 dynamometer (as in figure 3.7), is mounted under the tool bar The cutting force is captured by a dynamometer in the form of charges and the acquired signals are first passed through a charge amplifier, and then continuously recorded on a 4-channel Sony Instrumentation Cassette Recorder PC204AX as in figure 3.8 The recorded signal was sampled at 12 kHz through a PC-Scan 2 data acquisition system

A / Dsampler

ch arg e amp lifer

Figure 3.6 Schematic diagram of experimental setup

Chapter 3 Sensing Signals and Experimental Design

Figure 3.7 Dynamometer Figure 3.8 Sony cassette recorder

Table 3.2 Components of experimental system

1 OKUMA LH 35 – N turning machining

2 Kistler 9121 Quartz 3-Component Dynamometer

3 Kistler 5019A Multichannel Charge Amplifier

4 Sony Cassette Recorder PC204AX

5 Olympus measuring microscope and Panasonic digital camera

6 Personal Computer with PC scan 2 software

3.2 Workpiece and tool

3.2.1 Steel

The tool inserts used in this turning operation are Sumitomo SNMN120408 (A30), SNMG120408 (AC3000), and their detailed descriptions are listed in table 3.3 Two types of medium carbon steel ASSAB705 and ASSAB760 are used as the workpiece material with the range of diameter from 200 to 70 mm Their specifications are listed

in table 3.4

Table 3.3 Description of tool inserts

SNMN1204080 (uncoated ungrooved)

SNMG1204080 (coated grooved)

A30 AC3000

rake angle =-5o, relief angle = 5 o, cutting edge angle = 15 o, nose radius = 0.8 mm

Table 3.4 Specification for ASSAB705 and ASSAB760

Workpiece Hardness

(HB)

Tensile strength (MPa)

Yield strength (MPa)

Chapter 3 Sensing Signals and Experimental Design

3.2.2 Titanium

Titanium alloy is the widely used aeroengine alloys with properties of high temperature strength, hardness and chemical wear resistance Although these properties are desirable in practical applications, they pose a greater challenge to machining tools due to the following issues:

High cutting temperature: High cutting temperature is generated close to the

cutting edge of the tool when machining titanium alloy Ti-6Al-4V Ezugwu (2003) reported that nearly 80% of this heat is conducted into tool in the machining of titanium alloy, while only 50% of the heat is absorbed by the tool when machining steel Moreover the thermal conductivity of steel is about 6 times that of titanium alloys These effects result in the concentration of high temperature at the tool-workpiece interface, which weakens the bonding strength of the tool substrate, thereby acceleration tool wear

Chemical reactivity: Titanium is very chemically reactive, and therefore has a

tendency to weld to the cutting tool It also may react with most tool materials in excess of 500°C resulting in accelerated tool wear, and its high hot hardness and strength cause the deformation of cutting tool

Adhesion: Adhesion at the tool-chip interface and tool-workpiece material interface

is another important reason for the rapid tool wear Adhesion takes place after coatings worn out, thus leading to the initial chipping and the breakage of the cutting edge

Titanium alloy (Ti-6Al-4V) is used as the workpiece material with the range of diameter from 200 to 170 mm, and its chemical composition is listed in Table 3.5 SNMG120408N-UZ tool insert of material AC3000 is used to machine this workpiece,

Chapter 3 Sensing Signals and Experimental Design

which accounts for 60% of the total titanium production Flood cooling is used to minimize the temperature in this machining

In titanium machining process, the typical tool failure modes are notch wear, nose wear, crater wear, flank wear chipping and catastrophic tool failure The coating material AC3000 includes three layers: fine particle Al2O3 coating with heat resistance, thick TiCN layer with good wear and chipping resistance, tough cemented carbide substrate coating zirconium with good wear and fracture resistance

Table 3.5 Chemical composition of Ti-6Al-4V-Gr5 (in wt.%)

chemical

composition

strength (MPa) Ti-6Al-4V 0.045 0.27 6.2 4.2 0.02 0.11 0.0019 992

In this study, the Taguchi method of experimental design was employed so as to comprehensively cover the entire parameter space with consistent and reproducible results through few experiments With the use of this method, the number of experiments could be reduced while being able to capture the characteristic of parameters space

3.3.1 Taguchi method

In laying out an experiment and developing strategy, simple logic will usually be sufficient to establish all possible combinations of factors along with allowable ranges

Chapter 3 Sensing Signals and Experimental Design

But for engineering projects involving many factors, the number of possible combinations is prohibitively large It means a great deal of time and money is spent in experiments Taguchi observed this problem and started to develop new methods to optimize the process He clearly defined a set of orthogonal arrays (OAs), each of which can be used for many experimental situations The combination of standard experiment design and analysis methods in Taguchi’s approach produces consistency and reproducibility rarely found in any other statistical method So this method overcomes the limitations of traditional experimental design methods and improves the experimental processes depending on many factors To summarize, parameter design using the Taguchi method includes the following steps:

1 Identify the performance characteristics and process parameters to be evaluated

2 Determine the number of levels for the process parameters

3 Select the appropriate orthogonal array and assign the process parameters to the orthogonal array

4 Conduct experiments based on the arrangement of the orthogonal array

5 Analyze the experimental results using corresponding methods

3.3.2 Case study

A case study is provided when SNMN120408 insert of material A30 is used to machine steel ASSAB705 and ASSAB760 Three machining parameters were involved: cutting speed, feed rate and depth of cut, with two kinds of workpiece For each kind of workpiece, three levels of cutting speed and feed rate, and two levels of cutting depth were selected as shown in table 3.6 Normally, for every set of workpiece and tool insert, 18 (3*3*2) sets of cutting condition combinations are needed to cover the entire machining parameters space But it would be too expensive to carry out all

of these experiments

Chapter 3 Sensing Signals and Experimental Design

In Taguchi method, the experimental layout for machining parameters uses L9 orthogonal array with 4 columns and 9 rows This array can handle two two-level process parameters and two three–level process parameters at most Each machining parameter is assigned to a column and nine machining parameter combinations are required With L9 orthogonal array, only nine experiments are needed to study the entire machining parameter space, which are listed in table 3.7 In practical cutting experiments, these combinations were run in a random order for three times to avoid the influence of experimental setup

Table 3.6 Machining parameters and their levels

Depth(mm) 1 2

Table 3.7 Cutting conditions based on L9 orthogonal array

No f (mm/r) v (m/min) d (mm) Workpiece

Chapter 4 Feature Extraction and Selection

on the shop floor This, as discussed in Du et al (1995a), is due to the fact that monitoring signals (such as vibration, force and AE) are typically affected by:

• process defects (such as chatter and tool wear);

• process working conditions (e.g change of cutting speed);

• process noise (e.g inhomogeneity of the workpiece), structural vibration, sampling noise and other environmental noise

In TCM system, various extracted features from suitably processed sensing signals are utilized by researchers However, not all of these features are equally informative

in a specific monitoring system: certain features may correspond to noise, not information; others may be correlated or not relevant for the task to be realized This study comprehensively takes all these known signal features and aims to identify the most effective set that can give robust and reliable identification of tool condition In this chapter, automatic relevance determination (ARD) under Bayesian framework and SVM are coupled together to perform feature selection A detail introduction