Gopal T. Venkatesan

Automatic fault monitoring using acoustic emissions

Techniques for automatic monitoring of faults in machinery are being considered as a means to safely simplify or dispense with expensive periodic fault inspection procedures. This paper presents results from an ongoing investigation into the feasibility of using acoustic emissions (AEs) for automatic detection of microcrack formation/growth in machine components.

Detection and characterization of cracks for failure monitoring and diagnostics

An early indicator of the onset of material failure due to fatigue and/or stress is the appearance of microcracks. The generation of these cracks creates propagating acoustic signals in the material, termed acoustic emissions (AEs). This paper presents results from an ongoing investigation into the detection end characterization of microcracks using AE signals measured in the presence of strong correlated interference and white noise.

A deterministic blind identification technique for SIMO systems of unknown model order

In this paper we present a method for the deterministic blind identification of single-input multiple-output systems with unknown model order. The technique, that is applicable to both the FIR and IIR cases, requires only an upper bound of the model order. It is based on the special kernel structure of block Toeplitz matrices. When the model order is overestimated, this special structure entails the true solution to be embedded in the overestimated solution in a unique shift-chain form. This special shift-chain structure is then utilized to extract the true solution.

Time-frequency distribution kernels using FIR filter design techniques

Time-frequency (t-f) distribution kernels are obtained using finite impulse response filter design methods. Two methods are considered: the window method and the frequency sampling method. The ideal kernel is first specified in the ambiguity domain. At each time-lag, an FIR filter approximating the desired response is designed using one of the above methods. The individual responses along the different lags are then assembled to construct a finite extent t-f kernel with desirable characteristics. Two different classes of kernels are therefore introduced. While the window based t-f kernels are simple, noniteratively generated, and can be used for all practical considerations, the frequency sampling class of kernels is computationally efficient and leads to recursive computations. Kernels from the two different classes are presented along with simulation examples which illustrate their performance.

Detection and modeling of acoustic emissions for fault diagnostics

The formation of microcracks in a material creates propagating ultrasonic waves that are called acoustic emissions (AEs). These AEs provide an early warning to the onset of material failure. In practical cases, however, these AEs have to be detected at very low SNRs, amongst strong interference and random noise. This paper presents some preliminary results from an ongoing investigation into the modeling and detection of AEs as a viable technique for predictive diagnostics.

Acoustic emission transient detection based on linear model residuals

Acoustic emissions (AE), which are ultrasonic waves created by the formation/propagation of a crack in a material, provide a possible avenue for automatic fault monitoring in machinery. In metal components, these AE signals are very complex transients that exhibit rapid time-varying behaviour due to multiple modes and multipath propagation. We present a scheme for the detection of AE transients that requires minimal a priori assumptions on the signal model. The technique, that is amenable to both adaptive and block implementations, is based on the thresholding of the ratio of the residual energy in fitting the transient to a linear model, to the signal energy. The method works well in detecting transients even at low SNR.

Discrete powers-of-two kernels for time-frequency distributions

We introduce a new class of powers-of-two (PFT) kernels for fast real time implementations of time-frequency distributions. In this class, the local autocorrelation function is computed using a series of shifting and addition operations. PFT filter design techniques are not limited to the design of fixed kernels. They can also be used to design data-dependent kernels suitable for specific operating environments. In the time-frequency context, where the task is to identify the signal autoterms in the time-frequency domain, a discretized powers-of-two kernel shows little or no difference in performance from its infinite precision counterpart. A simple modification of the PFT design technique that significantly improves the approximation when small register lengths are used, is also introduced.

Fault monitoring using acoustic emissions

Automatic monitoring techniques are a means to safely relax and simplify preventive maintenance and inspection procedures that are expensive and necessitate substantial down time. Acoustic emissions (AEs), that are ultrasonic waves emanating from the formation or propagation of a crack in a material, provide a possible avenue for nondestructive evaluation. Though the characteristics of AEs have been extensively studied, most of the work has been done under controlled laboratory conditions at very low noise levels. In practice, however, the AEs are buried under a wide variety of strong interference and noise. These arise due to a number of factors that, other than vibration, may include fretting, hydraulic noise and electromagnetic interference. Most of these noise events are transient and not unlike AE signals. In consequence, the detection and isolation of AE events from the measured data is not a trivial problem. In this paper we present some signal processing techniques that we have proposed and evaluated for the above problem. We treat the AE problem as the detection of an unknown transient in additive noise followed by a robust classification of the detected transients. We address the problem of transient detection using the residual error in fitting a special linear model to the data. Our group is currently working on the transient classification using neural networks.