Performance Analysis of Acoustic Emission Hit Detection Methods Using Time Features

Acoustic emission (AE) analysis is a powerful potential characterization method for fracture mechanism analysis during metallic specimen testing. Nevertheless, identifying and extracting each event when analyzing the raw signal remains a major challenge. Typically, the AE detection is carried out using a thresholding approach. However, though extensively applied, this approach presents some critical limitations due to overlapping transients, differences in strength and low signal-to-noise ratio. In this paper, to address these limitations, advanced methodologies for detecting AE hits have been developed. The most prominent methodologies used are instantaneous amplitude, the short-term average to long-term average ratio, the Akaike information criterion, and wavelet analysis, each of which exhibits satisfactory performance and ease of implementation for diverse applications. However, their proneness to errors in the presence of non-cyclostationary AE wavefronts and the lack of thorough comparison for transient AE signals are constraints to the wider application of these methods in non-destructive testing procedures. In this paper, with the aim of making aware of the drawbacks of the traditional threshold approach, a comprehensive analysis of its limiting factors when taking into regard the AE waveform behavior is presented. In addition, in a second section, a performance analysis of the main advanced representative-methods in the field is carried out through a common comparative framework, by analyzing first, AE waves generated from a standardized Hsu-Nielsen test and second, a data frame of a highly active signal derived from a tensile test. In this paper, with the aim to quantify the performance with which these AE detection methodologies work, for the first time, time features as the endpoint and duration accuracies, as well as statistical metrics as accuracy, precision, and false detection rates, are studied.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.