26.4 Signal Processing for Sensor-Based Tool Condition Monitoring

Back

Using the sensor information from the different sensor systems described in the previous section, a

decision must be made with respect to the tool condition. This decision is generally referred to as the data

classification. It is often better to combine sensory information to solve a complex problem such as TCM.

Such a combined approach is referred to as sensor fusion. Sick (2002) proposed a generic sensor fusion

architecture for TCM, which summarizes the various sensor fusion levels of a TCMS. These are:

* Analogue preprocessing

* Digital preprocessing

* Feature extraction

* Wear model

* Decision making

Fusion of sensor information can occur at any of these levels. Analogue and digital preprocessing are

activities such as signal amplification, conditioning, filtering, calibration, and temperature compensation.

The feature extraction step is probably the most important step, because here the sensor signals

must be condensed and reduced to only a few appropriate wear sensitive values. Many different methods

are available to achieve this. The wear model level establishes a relationship between the chosen features

and the tool condition. In many cases, neural networks (NNs) are used in this step, and sensor fusion

takes place within the NN. A decision level can also be included where a final decision is made with

respect to the tool condition, for instance a “competing experts” formulation if a TCMS is used in

conjunction with a tool-life equation. Discussions on the various techniques follow.

26.4.1 Feature Extraction

Most decision-making techniques for process monitoring are based on signal features. Through

appropriate signal processing, features can be extracted from signals that show consistent trends with

respect to tool wear. Features are mainly derived through processing in the time, frequency, or joint timefrequency

domain or statistical analysis.

26.4.1.1 Time Domain

Features extracted from the time domain are usually fundamental values such as the signal mean or RMS.

Other techniques include the shape of enveloping signals, threshold crossings, ratios between

Vibration-Based Tool Condition Monitoring Systems 26-11

© 2005 by Taylor & Francis Group, LLC

time-domain signals, peak values, and polynomial

approximations of time-domain signals. Examples

of time-domain features from an interrupted

cutting operation are shown in Figure 26.15.

It has been found that some of the time-domain

features show good correlation with tool wear and

are easy to implement (Scheffer, 1999). Bayramoglu

and Du¨ngel (1998) investigated the use of

several different force ratios (calculated from the

static cutting forces). It was found that certain

force ratios can be used to monitor tool wear

under a wide range of cutting conditions. Most

commercial TCMS rely on time-domain information.

However, time-domain features are

known to be sensitive to disturbances and should

be complemented with features from another

domain.

26.4.1.2 Frequency Domain

The power or energy of certain frequency bands

in the fast Fourier transform (FFT) is often

suggested as a feature for TCM. It is very

challenging to identify spectral bands that are

sensitive to tool wear. It is even more difficult to

determine exactly why these frequencies are

influenced by tool wear. Power in certain bands

will often increase due to higher excitation forces

because of the increase in friction when the tool

starts to wear. Sometimes a peak in the FFT will

also shift due to changing process dynamics as a

result of tool wear. An early frequency-domain

approach is reported by Jiang et al. (1987), in

which frequency-band energy is determined from

the power spectral density (PSD) function as a

feature for tool wear.

Some authors suggest that two frequency ranges be identified from the original signal (Bonifacio and

Diniz, 1994). The one range must be sensitive to tool wear, the other must be insensitive. For instance, if

the measurement was made from 0 to 8 kHz, it must be split (using appropriate filters) into a 0 to 4 kHz

signal and a 4 to 8 kHz signal. If the lower range is more sensitive to tool wear, a ratio between the two

ranges can be calculated. If this ratio exceeds a certain pre established value, it can be concluded that the

end of the tool life has been reached. This can also apply for a ratio between the signal recorded from a

fresh tool to that compared with a worn tool. Examples of frequency-domain features from cutting forces

are shown in Figure 26.16.

One difficulty with frequency-domain approaches is that the dynamics of the operation and

measurement hardware is not always fully understood. The fact that measurement hardware dynamics

instead of process dynamics are often measured was also recently identified by Warnecke and Siems

(2002). The response of a force dynamometer is influenced by its clamping condition, which may cause it

to experience nonlinearities at relatively low frequencies. There are also some uncertainties when using

these instruments, relating to their calibration and other varying parameters. A model for expressing the

uncertainties when collecting cutting forces with a dynamometer was proposed by Axinte et al. (2001).

These uncertainties might be responsible for the scatter of force components often reported in the

FIGURE 26.15 Simple time-domain features.

FIGURE 26.16 Frequency-domain features.

26-12 Vibration and Shock Handbook

© 2005 by Taylor & Francis Group, LLC

literature. An interesting study is also reported by Ba¨hre et al. (1997), concerning determination of the

natural frequencies of the machine tool components using the finite element method (FEM). These are

taken into account for interpretation of the vibration/AE signal.

26.4.1.3 Statistical Processing

In the case of statistical features, signals are assumed to have a probabilistic distribution, and

consequently, useful information can be extracted from the statistics of the distribution. Hence, the signal

is regarded as a random process. Generally, machining processes are nonstationary but are assumed to be

stationary for the short periods during which features are calculated. Several statistical features have been

investigated for TCM and can be applied to several machining operations. The main features are those

that describe the probability distribution of a random process (variance, standard deviation, skewness,

kurtosis, etc.) and coefficients of time-series models. There are also miscellaneous other statistical

features, such as cross-correlations, the coherence function, and the harmonic mean.

One useful approach is the use of autoregressive (AR) and autoregressive moving average (ARMA)

coefficients. AR coefficients computed for a signal represent its characteristic behavior. When the signal

changes during the cutting operation due to tool wear, the model coefficients also change and can then be

utilized to monitor the progressive tool wear. Baek et al. (2000) report on the use of an eighth order AR

model for tool breakage detection during end milling. It was found that the AR approach is somewhat

more accurate than the frequency band energy method. Yao et al. (1990) used the ARMA method to

decompose the dynamic cutting force signals, and wear-sensitive frequencies were identified. This

assisted in identifying the importance of certain vibration modes with respect to TCM.

The use of statistical process control (SPC) methods is also reported by some authors. Jun and Suh

(1999) considered the X-bar and exponentially weighted moving average (EWMA) for tool breakage

detection in milling. Jennings and Drake (1997) used statistical quality control charts for TCM. Different

statistical parameters are calculated and examples of one-, two- and three-variable control charts are given.

26.4.1.4 Time – Frequency Domain

The most common time – frequency domain processing method in TCM applications is wavelet analysis.

A comprehensive discussion on the advantages and disadvantages of wavelet analysis for TCM is

described by Sick (2002). It is often stated that wavelets are used because they provide information about

the localization of an event in the time as well as in the frequency domain. However, locating discrete

frequency-related events in the time domain is rarely of importance with respect to tool wear (which is a

gradually increasing phenomenon). In contrast, tool breakage will have a large localized effect in the time

domain, but this can be monitored more effectively using time-domain techniques. Furthermore,

wavelets are time variant and the exact contribution of a particular frequency at any given time can never

be determined accurately due to Heisenberg’s uncertainty principle.

Despite the above arguments, the use of wavelet analysis for TCM is reported in several publications.

Lee and Tarng (1999) use the discrete wavelet transform for cutter breakage detection in milling and find

that the technique is reliable even under changing machining conditions. Luo et al. (2002) published

results of a TCMS using wavelet analysis of vibration signals. In this case, the wavelet is used as a filter to

enhance wear-sensitive features in the signals. However, the results are not compared with conventional

digital filtering. A comparative study between wavelets and digital filtering for tool wear monitoring was

carried out by Scheffer (2002). It was found that, although the wavelet packets act as automated filters, a

very similar (if not better) result could be achieved with appropriate digital filtering. The use of wavelets

increase the complexity of the TCMS, which is a disadvantage for shop-floor implementations.

Furthermore, the results from digital filtering can be physically related to the machining operation and

tool wear, whereas the behavior of wavelet packets is more difficult to interpret.

Another method of time – frequency analysis that can be applied for TCM is spectograms (e.g., the

Gabor distribution). Spectograms are very useful to identify stationarity in dynamic signals, and

for detection of disturbances that may be time-localized in signals. The use of the Choi–Williams

time – frequency distribution for TCM during multimilling is described by James and Tzeng (2000).

Vibration-Based Tool Condition Monitoring Systems 26-13

© 2005 by Taylor & Francis Group, LLC

Wear-sensitive regions on the time – frequency

distribution are calculated and used as inputs to

a NN for wear classification. An example of a

change in the dominant chip curl frequency during

hard turning is shown in Figure 26.17. It is obvious

that, due to some disturbance (perhaps tool wear),

the dominant dynamic force frequency “jumps”

from 27 to 9 Hz.

26.4.2 Feature Selection

Various authors attempt to generate features that

are sensitive to tool wear but insensitive to

changing machining parameters. For most operations,

the machining parameters can be included

in the wear model and hence the sensitivity of the

features is not such an important issue. There are

also other techniques to normalize sensor data

with respect to machining parameters, for instance

the use of a theoretical model (Sick, 1998). This is

very useful if the machining conditions change so often that not enough data can be collected for training

or calibrating a model. Numerous techniques exist to select the most wear-sensitive features or to reduce

the input feature matrix to a lower dimension. The main techniques for feature selection and reduction

are listed below:

* Principal component analysis (PCA)

* Statistical overlap factor (SOF)

* Genetic algorithm (GA)

* Partial least squares (PLS)

* Automatic relevance determination (ARD)

* Analysis of variance (ANOVA)

* Correlation coefficient

* Simulation error calculations

Al-Habaibeh et al. (2000) presented a TCMS for a parallel kinematics machine tool for high-speed

milling of titanium. An interesting approach to feature selection is employed, called self-learning

automated sensors and signal processing selection (ASPS). This approach is based on an on-line selflearning

methodology, whereby a certain feature will be selected automatically based on a correlation

with tool wear. A linear regression is performed on each feature in the sensory feature matrix to detect the

sensitivity of each feature with respect to tool wear. A very interesting cost analysis is then preformed to

determine if the installation of a sensor justifies its costs.

Ruiz et al. (1993) proposed the use of a discrimination power for feature selection in a TCM

application. The method is similar to that of the SOF. An automated version is proposed that also checks

for linear correlation between features. It is difficult to assess the success rate of the automated procedure

because the experiments/simulations are not described in enough detail. Lee et al. (1998) describe the use

of ANOVA to determine the best force ratio for TCM statistically. Several ratios between the three main

cutting forces are computed and the influence of controllable parameters (e.g., machining conditions) on

these ratios are investigated by means of ANOVA.

Du (1999) describes the use of a blackboard system, which is a knowledge-based approach for feature

selection and decision-making. An advantage is the fact that a physical interpretation of a feature can be

linked to phenomena in the machining operation. The method is also flexible, but suffers from the

disadvantage of requiring a large quantity of data and expertise to establish the knowledge-based rules.

FIGURE 26.17 Time-frequency distribution of cutting

force signal. (Source: Scheffer, et al., Int. J. Mach. Tools

Manuf., Elsevier, 2003. With permission.)

26-14 Vibration and Shock Handbook

© 2005 by Taylor & Francis Group, LLC

Mdlazi et al. (2003) compared the performance of ARD and PCA for feature selection for two damage

detection case studies. It was found that the performance of the methods is similar, but one might

perform better on a particular data set. Generally speaking, the PCA yields better results for damage

detection problems. Scheffer and Heyns (2002b) compared several feature selection methods for TCM,

such as SOF, PCA, GA, ANOVA, and the linear correlation coefficient. It was found that the correlation

coefficient approach and the SOF should be preferred for TCM applications. PCA could also be of

assistance, but the feasibility of PCA for on-line applications is still questionable. The correlation

coefficient and SOF is expressed as percentages in Figure 26.18 (from Scheffer and Heyns, 2004) for 30

different wear monitoring features in a turning tool wear case study. Ideally, a feature with a high level of

correlation and SOF should be selected.

As a last step, engineering judgment is required for proper feature selection because automated

methods will often select features that are dependant on one another, thus not achieving the goals of

sensor fusion. The following rules can be used as a guideline for selecting features for TCM:

* Select features from the static and dynamic parts of force signals.

* Select features measured in different directions.

* Use time- and frequency-domain features.

* Features based on simple signal processing methods are preferred.

* There should be a reasonable physical explanation for the behavior of a feature with respect to

tool wear.

Навигация