INTRODUCTION
A time series is a set of observation x_{t}, each one being recorded
at a specified time t (Brockwell and Davis, 1991). In statistics
and signal processing, a time series is a sequence of data points, measured
typically at successive times spaced at uniform time intervals (Pollock,
1999). Methods for time series analysis may be divided into two classes:
frequencydomain methods and timedomain methods. The time domain method is
suitable for simple structures where only a few hot spots need to be investigated
(Wang and Sun, 2005). Fatigue analysis of such structures
generally use frequency domain methods. In the frequency domain, fatigue life
is calculated from the Power Spectral Density (PSD) of the stress response.
Frequencydomain methods include spectral analysis and recently wavelet analysis;
timedomain approaches include autocorrelation and crosscorrelation analysis.
Fatigue life prediction is important in the design process of vehicle structural
components and involves a set of observations of a variable taken at equally
spaced intervals of time (Anderson, 1996). Since fatigue
is one of the major causes in vehicle component failure, its life prediction
or fatigue damage has become a subject of discussion in many texts (Nadot
and Denier, 2004). Svenssona et al. (2005)
conducted fatigue life prediction based on variable amplitude testspecific
applications, while Rahman et al. (2009) conducted
the fatigue life behaviour of an Aluminium alloys suspension arm using the strainlife
approach. The strainlife (εN) approach was used for this analysis as
the case study related to low cycle fatigue, which is a suitable approach to
analyze random data collected from automotive components.
This study discusses some issues related to the analysis on VA loading strains
data behaviour of a lower suspension arm by using amplitude Probability Distribution
Function (PDF), Power Spectral Density (PSD) function and the signal statistical
parameters which are included the mean, the RootMeanSquare (RMS), the skewness
and the kurtosis values. The signal cross correlation value was then calculated
using the Cross Correlation Function (CCF) to determine the similarity of the
two signals pattern. Finally, the fatigue life analysis was taken using three
models of strainlife method which are CoffinManson, Morrow and SmithWatsonTopper
(SWT). Material fatigue properties were obtained through using the Glyphworks^{®}
software. The goal of this study is to observe the suitability of these techniques
and methods in detecting the pattern behaviour in terms of energy and probability
distribution in a strainbased time series. These kinds of signals were collected
from a real road test of an automobile component.
Signal statistical analysis: A signal is a series of numbers obtained
through measurement, typically obtained using some recording method as a function
of time. In the case of fatigue research, the signal consists of a measurement
of a cyclic load, i.e., force, strain and stress against time (Meyer,
1993). In actual applications, mechanical signals can be classified to have
stationary or nonstationary behaviour. Stationary signal behaviour shows that
the statistical property values remained unchanged over time. For nonstationary
signals, common in fatigue analysis cases, however, the statistical property
values of a signal are dependent on the time of measurement (Bendat
and Piersol, 1986).
In mathematics, a Probability Density Function (PDF) represents a probability
distribution in terms of integrals. In amplitude probability analysis (Ncode,
2005) the Time at Level analysis is useful as it is more common to represent
the data as amplitude PDF. This is simply a way of normalising the time at level
plot so it does not change shape when the length of time history changes or
the number of bins is changed. Time at level analysis is useful for determining
the statistical amplitude content of a signal and can be used to detect anomalies
such as signal drift, spikes, clipping, etc.
Frequency analysis data is typically presented in graphical form as a Power Spectral Density (PSD) function. Essentially a PSD displays the amplitude of each sinusoidal wave of a particular frequency. The mean squared amplitude of a sinusoidal wave at any frequency can be determined by finding the area under the PSD over that frequency range. In PSD plot, each frequency step value is characterised by amplitude, A_{k}, as in the following equation:
where, S(f_{k}) is underlying PSD of the Gaussian signal and f_{k} is the harmonic frequency. The PSDs are useful for detecting resonance in components, aliasing in the data, frequency interference, etc.
The objective of time series analysis is to determine the statistical characteristics
of the original function by manipulating the series of discrete numbers. In
normal practice, the global signal statistical values are frequently used to
classify random signals (Abdullah et al., 2008a).
The most commonly used statistical parameters are the mean value, the RootMeanSquare
(RMS) value, the skewness and the kurtosis (Nopiah et
al., 2010). For a signal with a number n of data points, the
mean value will be:
The RootMeanSquare (RMS) is the square root of the mean square value (Draper,
2007). The RMS value, which is the 2nd statistical moment, is used to quantify
the overall energy content of the signal. For a zeromean signal the RMS value
is equal to the SD value. For discrete data sets the RMS value is defined as:
This measure is also used as a simple quality check on the measured data as
any change in mean or range will be reflected in this parameter (Ncode,
2005).
The skewness, the signal 3rd statistical moment, is a measure of the symmetry of the distribution of the data points about the mean value. The skewness for a symmetrical distribution such as a sinusoid or a Gaussian random signal is zero. Negative skewness values indicate probability distributions that are skewed to the left, while a positive skewness value indicates probability distributions that are skewed to the right, with respect to the mean value. The skewness of a signal is given by:
Kurtosis value is mathematically defined in Eq. 5, which is the signal’s 4th statistical moment that highly sensitive to the spikiness of the data. For a Gaussion distribution the kurtosis value is approximately 3.0. Higher kurtosis values indicate the presence of more extreme values than should be found in a Gaussian distribution. Kurtosis is used in engineering for detection of fault symptoms because of its sensitivity to high amplitude events.
Correlation function: Correlation is a mathematical tool used frequently
in signal processing for analysing functions or series of values, such as time
domain signals. It is a mathematical tool for finding repeating patterns, such
as the presence of a periodic signal which has been buried under noise, or identifying
the missing fundamental frequency in a signal implied by its harmonic frequencies.
Correlation is the mutual relationship between two or more random variables.
Autocorrelation is the correlation of a signal with itself (Parr
and Phillips, 1999), unlike cross correlation, which is the correlation
of two different signals.
The autocorrelation function (ACF) is the plot of autocorrelations and is very
useful when examining stationarity and when selecting from among various nonstationary
models. Autocorrelation is one of the major tools in time series modeling (Minitab
StatGuide, 2007). In autocorrelation, lag is a time periods separating the
ordered data and used to calculate the autocorrelation coefficients. The maximum
number of lags is roughly n/4 for a series with less than 240 observations or
(√n + 45) for a series with more than 240 observations, where, n is the
number of observations or for time series the number of data points.
An important guide to the properties of a time series is provided by a series
of quantities called sample autocorrelation coefficients, which measure the
correlation between observations at different distances apart (Chatfield,
1996). Let the time series of length N be x_{t} , t = 1,...,N. The
lagged scatterplot for lag k is a scatterplot of the last Nk observations against
the first Nk observations. Eq. 6 can be generalized to give
the correlation between observations separated by k time steps:
where:
is the overall mean. The quantity r_{k} is called the autocorrelation
coefficient at lag k. The plot of the autocorrelation function as a function
of lag is also called the correlogram.
The Cross Correlation Function (CCF) is the plot of cross correlations between elements of two time series and is useful in determining if a given series leads another series and by how many periods. The CCF of two time series is the productmoment correlation as a function of lag, or timeoffset, between the series. The lags are the time periods separating data in one time series from data in the other time series when calculating the cross correlation coefficients. By default, the lags range from (√n + 10) to +(√n + 10), where n is the data points of time series.
The following equations are used to calculate the coefficient of cross correlation
(r_{xy}) of two time series x and y, each with N data points (Chatfield,
1996):
if,
where, k is a number indicating a time shift of one signal with respect to
the other; k=0 for the original two time series synchronized in time. The correlation
r_{xy}(0), calculated from c_{xy}(0), will give an indication
of pattern similarity between the two sets of data.
Fatigue life assessment: It is common that the service loads acquired
on components of machines, vehicles and structures are analysed for fatigue
life using crack growth approaches. This approach is suitable for high capital
valued structures, such as large aircraft, space shuttle, pressure vessels and
oil rigs (Dowling, 1999). A fatigue life estimation based
on the related strainbased approach is usually used in these cases (Dowling,
1999; Wu et al., 1997). The strainlife fatigue
model relates the plastic deformation that occurs at a localised region where
fatigue cracks begin to effect the durability of the structure. This model is
often used for ductile materials at relatively short fatigue lives but can also
be used where there is little plasticity at long fatigue lives. Therefore, this
is a comprehensive approach that can be used in place of the stressbased approach.
Current industrial practice for fatigue life prediction is to use the PalmgrenMiner
(PM) linear damage rule (Draper, 2007). For strainbased
fatigue life prediction, this rule is normally applied with strainlife fatigue
damage models. The first strainlife model is the CoffinManson relationship
(Coffin, 1954; Manson, 1965),
defined as:
where, E is the material modulus of elasticity, ε_{a} is true strain amplitude, 2N_{f} is the number of cycles to failure, σ_{f}’
is the fatigue strength coefficient, b is a fatigue strength exponent, is a fatigue ductility coefficient and c is a fatigue ductility exponent.
Some realistic service situations will involve nonzero mean stresses. Two
mean stress effect models are used in the strainlife fatigue damage analysis,
i.e., Morrow and SWT strainlife models. Mathematically, the Morrow model (Morrow,
1968) is defined by:
where, σ_{m} is the mean stress. The SWT strainlife model (Smith
et al., 1970) is mathematically defined by:
where, σ_{max} is the maximum stress for the particular cycle.
The fatigue damage caused by each cycle is calculated by reference to material
life curves, i.e., stress life (SN) or strain life (εN) curves (Abdullah
et al., 2008b). The N_{f} value for each cycle can be obtained
from Eq. 9 to 11 for all three models and the fatigue damage
value, D, for one cycle is calculated as:
and therefore fatigue damage values have the range (0, 1) where zero denotes no damage (extremely high or infinite number of cycles to failure) and 1 means total failure (one cycle to failure). Failure occurs when the fatigue damage ∑(n/N_{f}) = 1, so the fatigue life is the inverse of the damage as expressed below:
MATERIALS AND METHODS
This study was conducted at field road test on Selangor, Malaysia in November
2009. A Variable Amplitude (VA) strain loading was used in this study as an
input signal for the lower suspension arm of midsized sedan car. The fatigue
data acquisition system, SoMat eDaQ Data Acquisition, was used for the strain
data measurements. The strain loading was measured using two strain gauges of
5 mm size which were located at the maximum stress area and the minimum stress
area, notated as for channel 1 and 2, respectively. The signal was measured
on the right lower suspension arm of a car travelling over a pavé road
of 40 km h^{1} and a highway route of 70 km h^{1}.

Fig. 1: 
Flowchart of the fatigue road test and fatigue data analysis 
The strain data was collected twice for each road in order to provide a variety
of data sets pattern between channel 1 and channel 2. The data was collected
in the unit of microstrain (με) for 60 sec at a sampling rate of 500
Hz, which gave 30,000 discrete data points.
A flowchart describing the fatigue road test, analyzing strain data processes using data distribution function, global signal statistic parameters, cross correlation function and fatigue damage calculations is shown in Fig. 1. In first stage, the collected strain signal was then analyzed for the distribution function and global signal statistics parameter using the GlyphWorks^{®} software. Then the distribution pattern and signal statistical parameters will be compared between the pavé road and highway route data sets. The second stage analyzes the strain signal of Run 1 and 2 using the cross correlation method using Minitab 15^{®} software. The purpose of this method is to select the strain signal which had the higher cross correlation function value, which is then used in the final stage. The last stage is the fatigue life analysis which is achieved by calculating the fatigue damage values using three models of strainlife, which are CoffinManson, Morrow and SmithWatsonTopper (SWT) that are contained in the GlyphWorks^{®} software.
RESULTS AND DISCUSSION
Fatigue time series: The data plot in Fig. 2ad
was collected on a straight pavé road at Wilayah Putrajaya and a highway
route of Sistem LingkaranLebuhraya Kajang (SILK) of Selangor.

Fig. 2: 
Time history plot of (a, b) straight pavé for Run 1
and 2 and a (c, d)highway route for Run 1 and 2 
These strain signals, having a VA pattern in the strain format, were measured
on the front left lower suspension arm of midsized sedan car travelling on
public road surface. The road conditions were in the surface type of a stretch
of highway road for representing mostly consistent load features and a stretch
of brickpavéd road for representing a noisy but mostly consistent load
features.
The captured of nonstationary strain loading data was analyzed in the time domain signal plots. It showed the high strain amplitudes for the pavé road were recorded at 295 and 75 με for channel 1 and 2, respectively. For highway route, high strain amplitudes were recorded at 321 με for channel 1 and 81 με for channel 2. The peak in strain signal or spike, occurs when vehicles is braking or driven on an uneven road.
Global signal statistics: Figure 3ad
show the amplitude PDF diagram for straight pavé and highway route sharing
the shape of classic symmetrical bell curve following the Gaussian Normal distribution.
This shaped curve is expected for random vibration signals. This PDF diagram
of channel 1 has a wide range of bell shaped compared to channel 2. This is
related to the amplitude range of it signal, which the range is wider than the
similar trend of channel 2. In addition, the signal of channel 1 is more skewed
to right side from origin point, but the peak of channel 2 is around origin
point. It means that the distribution of channel 1 is found to be a nonGaussian
distribution. For channel 2, the distribution is closely to the Gaussian distribution.
The diagram below shows the PSD produced from pavé and highway route
strain signals. The PSD diagram of channel 1 shows a broad band signal while
channel 2 shows nearly as narrow band signal. A broad band signal covers a wide
range of frequencies and it might consist of a single, wide spike or a number
of distinct spikes as shown in the PSD diagram, a narrow band signal only cover
a narrow range of frequencies; this occurrence can be seen in Fig.
4ad.

Fig. 3: 
Graph of amplitude PDF of (a, b) straight pave for Run 1 and
2 and a (c, d) highway route for Run 1 and 2 

Fig. 4: 
Graph of PSD function of (a, b) straight pavé for Run
1 and 2 and a (c, d) highway route for Run 1 and 2 
Table 1: 
Global signal statistical parameter 

This PSD pattern of Fig. 4 represents the texture of surface road that is mostly noisy. The PSD is also used to measure the power of a signal by converting the signal from the time domain to the frequency domain. It shows where the average power is distributed as a function of frequency. In this case, power distribution of channel 1 has a greater range of frequency compared to channel 2. This is related to channel 1 having more vibration compared to channel 2 when the vehicle is moving.
Statistical analysis is concerned with reducing a long time signal into a few numerical values that describe its characteristics, therefore the global statistical parameter values were calculated and the results were shown in Table 1. Generally the mean value is the average value or the centre of area of the PDF about the xaxis. From this table, mean value of channel 1 is significantly larger than channel 2. The mean value represent the point of highest peak in PDF diagram. When the mean value is not equal to zero, it means the distribution is a nonGaussian distribution. It is clearly seen in Fig. 3, where PDF diagram on channel 1 shows nonGaussian distribution, but not for the channel 2 signal.
The RMS can be said as the vibration signal energy in a time series and the kurtosis value represented the amplitude range in a time series. The RMS value of channel 1 is higher than channel 2 which means the maximum stress area creates more vibration compared to the minimum stress area. With refering to the PSD diagram in Fig. 4, it was found that the power distribution on channel 1 is wider than for channel 2.
Skewness is a measure of the degree of asymmetry of a distribution. As we can see in Table 1, the skewness value of pavé road and highway route on channel 1 is lower than channel 2. Both channels of pavé road have positive skewness, which means the data distribution is skewed to the right. For highway route, at Run 2 the skewness has a negative value and this describes a data distribution skewed to the left. By skewed left, it means the left tail is long relative to the right tail. Similarly, skewed right means the right tail is long relative to the left tail. This is useful for identifying signal drift, cyclic hardening, etc.
The kurtosis value of channel 1 is lower than channel 2. This is because kurtosis was highly sensitive to the spikiness of the data. The differences of kurtosis value on pavé road is higher than highway because the pavé road surface represent more noisy even mostly consistent load features compared to highway surface.
Cross correlation analysis: This analysis will provide a correlation between two time series or two strain signals. The observations of one series are correlated with the observations of another series at various lags and leads. Crosscorrelations help identify variables which are leading indicators of other variables or how much one variable is predicted to change in relation to the other variable. There are three possible outcomes in the Pearson’s Product Moment correlation r; Positive correlation (r = +1), where as one variable rises the other variable is predicted to rise at a similar rate, Zero (r = 0) or no correlation at all, or Negative correlation (r = 1) where as one variable rises the other falls. The crosscorrelation test of two timeseries data sets involves many calculations of the coefficient r by timeshifting the one data set relative to the other data set. A typical crosscorrelation graph shows enough lags in both negative and positive directions, as Fig. 5, to show the cyclical relationship of the two sets of data.
Figure 5a show the cross correlation function (also known as correlogram) graph of pavé road for run 1 and 2 and it shows the correlogram pattern was similar for both runs. While for Fig. 5b, the correlogram shows different pattern for both runs. This might occurs when the driver could not maintain the velocity of vehicle. On pavé road, the r value of run 1 is 0.861 while the value of r at run 2 is 0.850, both at lag 0. These numbers described the strain signals at run 1 is nearly similar pattern compared to run 2. In Fig. 5b, at run 1, the r value is 0.930 while the value of r at run 2 is 0.873, both at lag 0. These numbers also described the strain signals at run 1 is closely for both pattern signal compared to run 2.
In crosscorrelation analysis, the signals that alternate are outofphase
from each other and will have a negative relationship, whereas the signals that
are synchronous will be inphase and have a positive relationship. The strength
of the relationship between two signals will be perfect at ±1 and will
diminish to a minimum when approaching 0. A high degree of symmetry or stability
along the xaxis indicates a stable relationship between the two signals.

Fig. 5: 
Graph of cross correlation function for: (a) straight pavé
road and (b) highway route 
Table 2: 
Result of total fatigue damage and number of cycles 

Table 3: 
Result of calculated fatigue life 

However, as the relationship between signals varies, therefore creating decreasing
correlation values beyond zero lag, this indicates less stability in the relationship.
Fatigue life analysis: Table 2 shows the results of the total fatigue damage that based on strainlife models of CoffinManson, Morrow and SmithWatsonTopper (SWT) approaches using GlyphWorks^{®} software. It was found on channel 1 for straight pavé road gave the higher total fatigue damage that is 9.97x10^{3}, 10.2x10^{3} and 11.0x10^{3} for CoffinManson, Morrow and SWT model, respectively. These values are higher than highway for channel 1. On the other hand, the total fatigue damage for channel 2 is 0.0185x10^{3}, 0.0186x10^{3} and 0.0201x10^{3} for CoffinManson, Morrow and SWT approaches, respectively. These values are lower than highway on channel 2.
From Table 2, it clearly shows that the number of cycles of channel 1 is lower than channel 2 for both roads, it means that the number of cycles could be less on high stress area of a component and vice versa. In overall, the number of cycles from pavé road is less than highway. This can be related to pattern of surface road where a stretch of highway road represents mostly consistent load features, while a stretch of brickpavéd road represents noisy and mostly consistent load features.
Predicting fatigue life for variable amplitude signals requires a cumulative damage model to calculate the total damage caused by a series of applied cycles. Miner’s linear damage rule equates the damage caused by one cycle to the inverse of the number of cycles to failure for the given cycle N_{f}. Miner’s linear damage rule was applied to the variable amplitude test signals by using the average constant amplitude fatigue life, determined experimentally, for each loading condition as the basis for the predicted fatigue lives. By using Eq. 13, the fatigue life could be calculated and concluded as in Table 3 for pavé road and highway of each strainlife models.
Referring to Table 3, the fatigue life of pavé road are lower than highway for channel 1 which is the maximum stress area. On the other hand, the fatigue life for pavé road was higher than highway on channel 2, where stress area is minimum. The lower suspension arm that was driven over the damagesurface road experienced the shortest fatigue life although the velocity of the car was the lowest among the velocity on the other roads. In overall, we can see that the fatigue life from CoffinManson model is the highest follows the Morrow model and finally SWT model.
CONCLUSIONS
This study discussed the capability of amplitude PDF and PSD techniques in
investigating the pattern behaviour in terms of energy and probability distribution
in a signal. The study also presents an analysis on statistical parameters that
were contained in VA loading strains data and detecting the similarity of pattern
signal using cross correlation method. A set of case study data consisting of
nonstationary VA loading strains data that exhibits a random behaviour was
used in this study. For nonstationary signal, the engineeringbased signal
analysis is important to explore the characteristics and behaviour of the signal
and the outcomes of these can also be related to the fault detection and analysis.
From the analysis, the nonstationary VA strain loading data shows strain amplitude
range for the pavé road are smaller compared to highway for channel 1
and 2, respectively. The amplitude PDF diagram illustrated symmetrical bell
shaped curve following the Gaussian Normal distribution and this shaped curve
is expected for random vibration signals. From PDF and PSD diagrams, the result
can be concluded that the nonGaussian distribution can be related to a broad
band signal, while for Gaussian distribution normally for a narrow band signal.
From this study, it was found that higher kurtosis values indicate the presence
of more extreme values than should be found in a Gaussian distribution. In cross
correlation analysis, the r value of run 1 is higher than run 2 for both roads.
In fatigue lives analysis, fatigue lives of pavé road are lower than
highway for channel 1 but higher than highway on channel 2. The findings from
this study are expected to be used to determine the pattern behaviour that exists
in nonstationary VA signals.
ACKNOWLEDGMENT
The authors would like to express their gratitude to Universiti Kebangsaan Malaysia (UKMKK02FRGS01292009) for supporting these research activities.