C. Rajagopalan

Improving the evaluation sensitivity of an ultrasonic pulse echo technique using a neural network classifier

In this paper, the use of an artificial neural network (ANN) for classifying weak ultrasonic signals has been attempted. The limitations of using a single conventional parameter for signal detection and classification (namely peak amplitude alone) are highlighted. Use of a multi-parameter approach is suggested. The ANN used is a multi-layered, feedforward, error-backpropagation network. Results are compared with those of conventional approaches.

Acoustic emission studies on welded and thermally treated AISI 304 stainless steel during tensile deformation

The present investigations are planned to study the influence of prior martensites formed due to cold treatment as 77K in AISI 304 SS welded specimens, on strain-induced martensites occurred during tensile deformation using AE technique. AE parameters like count rate and root mean square (r.m.s.) voltage have been used to characterize AE activities generated during tensile deformation process in as-welded and welded-treated samples. Frequency spectrum analysis of AE signals captured from the samples has been done to understand the dynamic behavior of the martensite phase formation. Tensile properties of these samples have also been reported. Volume fraction of the magnetic phase (martensite and delta ferrite) formed in these samples are measured before and after straining. X-ray diffraction (XRD) technique has been used to support the presence of delta ferrite (formed during welding) and martensite in the weld region.

A soft-computing framework for fault diagnosis

Abstract The field of fault diagnosis (FD) is undergoing rapid change in the methodologies it employs, and the vast array of materials/components/structures where it is applied. The field has grown steadily because of (a) the importance attached to fail-safe mechanisms; (b) ability to detect impending failure so that preventive and corrective action can be taken and (c) the opportunities opened up by automating the diagnosis process itself. FD as a tool has to draw knowledge and expertise from a variety of scientific fields and apply multi-disciplined knowledge to empirical domains in order to achieve its objectives. Application and use of automated FD concepts to empirical domains do not require merely crisp rules, but largely, a flexible and natural framework where “experience” often determines the success or failure of diagnosis. In this paper, we describe a flexible framework, which incorporates the ideas and tools of soft computing in its methods and approach.

On-line prediction of quality and shear strength of spacer pad welds of nuclear fuel pins by applying neural network analysis of acoustic emission signals

Abstract In the present study, the development of an acoustic emission technique (AET) based methodology is reported for online prediction of quality and shear strength of spacer pad welds of nuclear fuel pins of pressurised heavy water reactors (PHWRs). The quality evaluation of spacer pad welds was made by classification of different weld categories using cluster analysis and artificial neural network (ANN) study of acoustic emission signals generated during welding. The ANN approach was also effective in arriving at the quantitative estimation of percentage correct classification between any two classes. For assessment of shear strength of individual coins of spacer pad welds by ANN, the properties of basic sigmoidal function were exploited and this could predict the strength of each coin with an accuracy of 97%. The results assume significance because instrumentation methodology is suitable for online application and complement the currently followed statistical quality control approaches for spacer pad weld assessment.