Maximum a posteriori deconvolution of sparse ultrasonic signals using genetic optimization

Deconvolution of sparse spike sequences has received much attention in the field of seismic exploration. In certain situations in ultrasonic non-destructive testing (NDT) of materials, similar conditions as those found in seismic exploration occur. One example is the problem of detecting disbonds in layered aluminum structures. The reflection sequence convolved with the impulse response of the transducer results in masking closely spaced reflections. Deconvolution of these signals may reveal the reflection sequence and thus make the interpretation easier. In this paper we use the Bernoulli-Gaussian (BG) distribution for modeling the signal generation. This relatively simple model allows maximum a posteriori (MAP) estimation of the reflection sequence. A derivation of the MAP criterion is given for clarity. We propose a genetic algorithm for optimizing the MAP criterion. The genetic algorithm approach is motivated by the fact that the criterion is non-convex, implying that the criterion may have more than one local minimum point. The probability of obtaining the global optimal solution is increased by using the proposed genetic algorithm. One of the key features in genetic algorithms, the so-called cross-over operator, has been modified and adapted to the structure of the BG deconvolution problem to improve the efficiency of the search. The algorithm is tested on simulated data using the probability of detection (PD) and probability of false alarm (PFA) as evaluation criteria. The algorithm is also tested on real ultrasonic data from a layered aluminum structure. The results show considerable improvements in the possibility of interpreting the signals.