Fault Detection Techniques

– Er. Spandan Mishra

Fault detection technique has wide applicability in engineering problems ranging from quality control of a manufacturing process to structural health monitoring (SHM). Lots of efforts have been put in to identify the non-destructive smart damage detection technique that can reliably keep us updated about the integrity of the structure. Just like human skin which is sensitive to damage and external stimuli, researchers in the field of manufacturing engineering are working  to develop an intelligent material that is sensitive to external pressure and can even tell us when its damaged or will possibly damage. SHM is closely related with conditional monitoring. Technical challenges to adaption of SHM to common application include    i) developments of methods to optimally define number and location of sensors, ii) identification of features sensitive small damage severity, iii) the ability to differentiate  between changes in the signal caused by damage and those caused by environmental condition and iv) development of statistical methods to discriminate features from undamaged and damaged structure and perform a comparative study of different damage identification method applied to common data set (Farrar & Worden, 2007). The various fault detection techniques adapted till now are:

  1. Lamb wave based Structural Health Monitoring

Propagation of Lamb wave has been found to be sensitive to the structural damage. Inference about degree and location of damage can be made by comparing lamb wave propagation characteristics on a damaged structuring with that on pristine structure (Lu, Wang, Tang, & Ding, 2008). Lamb wave can be used to monitor large volume of the structure due to the fact that lamb wave have ability to travel considerable distance and have the ability to  cover large area along the line between transmitter and receiver(Diamanti, Hodgkinson, & Soutis, 2004). Lamb wave is characterized by two properties i) Phase velocity ii) Group Velocity. Phase velocity is referred to as propagation speed of the wave phase at a given frequency(Rose, 2004).

The feature that makes Lamb wave very useful for damage detection is that whenever it reaches a region of dissimilar wave speed a portion of wave is reflected proportionally to the difference in their stiffness and density (Seth S Kessler et al., 2002). Effect of through crack on lamb wave depends on the frequency at lower frequency like 50 kHz there is no change in the amplitude but phase difference. However at higher frequency reduction of signal amplitude is significant(Qing, 2006). A proper lamb wave mode should have following characteristics: i) Non-dispersion. ii)Low attenuation iii) High Sensitivity iv)Easy excitability v) Good detectibility.vi) Toiless selectivity (Su, Ye, & Lu, 2006).

Piezoelectric sensors are used to generate the lamb wave. When voltage is applied to the piezoelectric patch it expands and contracts parallel to the surface, inducing bending moment in the structure which results in the generation of wave(Valdés & Soutis, 2002). Though lamb wave are sensitive to damages there are various challenges associated with lamb wave based damage detection (Su & Ye, 2009b) including:

  • The signals collected may not represent actual state of the structure due contamination with environmental factors like humidity, temperature, pressure etc.
  • The vibration of the host structure may overshadow the actual signal.
  • A signal consists of multiple lamb wave modes with complex and unique dispersion behaviors.
  • Huge amount of data collected due to high sampling rate used tends to serialize the sampling.

The disadvantage of Lamb wave method include is that they require an active driving mechanism to propagate the waves , and the resulting requires mathematical interpretation .(Seth S Kessler et al., 2002)

  1. Control chart based condition monitoring

Control chart is process monitoring technique to assess stability of a process and distinguishing assignable cause variation from common cause variation. (Montgomery) Group of in control sample is taken to calculate the process mean and control limits. The control limits are driven by the natural variability of the process(Montgomery, 2005).Generally, the effective use of control chart will require periodic revision of the control chart limits. The multivariate statistical process control chart has been popularly used on the features extracted from the signal for fault detection. Univariate statistical process chart (USPC) can also be used to monitor the process but with rapid development of data acquisition system, many features of the process can be simultaneously monitored using Multi-Variant Statistical Process Control Chart (MSPC). Applying USPC independently to each component of multivariate data may give misleading results because of not allowing for some inherent relationship among the components.(Montgomery, 2005)

  1. Pattern extraction based fault detection approach

Humans have highly sophisticated ability to sense their environment and take actions accordingly. Developers have worked on similar sorts of algorithms which helps a machine to recognize patterns. Various Pattern recognition algorithms like Neural network, Principle Component Analysis (PCA) etc. are popularly used in Structural Recognition techniques have been successfully used to identify the type and the severity of damage. Damage detection usually is a three step process: 1) Data Preprocessing 2)Detection Algorithm 3)Damage Prediction (S. S. Kessler & Agrawal, 2007).  Once the data has been acquired, de-noising is done to remove all the unwanted noise and preexisting artifacts.

  • Outlier based methods for fault detection

(Worden, Manson, & Fieller, 2000) have used the concept of discordance to signal the deviance from norm.  This approach uses the lowest level of detection, i.e. It only answers if damage is present or not.

  • Autoregressive moving average based fault detection approach:

(Omenzetter & Brownjohn, 2006) used autoregressive moving average (ARIMA) models to analyze static strain data from a bridge during it construction and when the bridge was in service .However their model was not able to detect nature severity and location of the structural change. Autoregressive model for fault identification is one of the approaches. Nair et al (Nair, Kiremidjian, & Law, 2006) used  autoregressive moving average model (ARMA) to model the process ; they identified first three AR coefficients as damage sensitive features. Although the model was able to detect nature and location of the damage for numerical data, but the validity of the model on experimental data requires further work.

  • Wavelets:

Wavelet transform is a waveform with limited duration and its average amplitude is equals to zero(Su & Ye, 2009b). Wavelet transform allows detailed analysis of the signal by unveiling signal characteristic such as singularity or discontinuity. There are basically two types of wavelet transform: a) Continuous wavelet transforms b) Discrete wavelet transform. The wavelet based method for damage detection can be divided into three categories:

  • Variation of wavelet coefficient before and after damage.
  • Local perturbation of wavelet coefficient.
  • Reflective wave caused by local damage
    • Neural networks:

(Worden & Staszewski, 2000) have been able to train neural network to quantify and locate the impact when provided with feature of the process. They used genetic algorithm to optimize the location of the sensors. The use of neural networks is based on basic assumption that no damage occurs below certain energy threshold (Worden & Staszewski, 2000).  The diagnostic system used basically uses two neural networks one to predict the energy of impact and other to predict the damage location.  The major drawback of Neural Network based method is that it is totally dependent on features extracted from the data; inferior features may produce inferior results.

  • Principle Component Analysis:

Principle component Analysis has been widely used for dimension reduction in process monitoring and fault diagnosis. using the dependencies between variables to represent it in a lower dimensional form, without losing too much information and in a form that is easier to interpret (Jolliffe, 2002). It is used to extract useful pieces of information from the matrix by representing it in terms of new set of orthogonal variables.

  • Frequency domain based feature extraction:

Damage index of the signal can be computed from the features extracted in the frequency domain.  Features that can be used to calculate the damage index include figure of merit(FOM) (Keilers & Chang, 1995)(Choi, Keilers Jr, & Chang,1994)(Tracy & Chang,1998), spectral density(Halabe & Franklin, 1998),peak of FFT amplitude, FFT coefficient(Monnier,2006),variance and kurtosis (Staszewski, Boller, & Tomlinson, 2004)(Sohn et al., 2001).

So, in conclusion, depending upon the degree of accuracy and extent of information required various fault detection tools can be used. Basic statistics like RMSD, covariance or correlation can be used to trigger the out of control processes; these statistics don’t usually give the direction of change. Pattern recognition based tools like wavelets, principle component analysis, Fast Fourier transformation can be used to extract features from the original signal. Quality control tools like Shewart control chart or Hotelling’s control can then be used further calculate the degree of variance from in control process.