A Framework for Fusion Application in Ndi

Historically, a variety of NDI technologies have been successfully utilized for the detection of corrosion damage and cracks in a variety of products. Each technology leverages a sensing theory over a particular problem domain. Fusion of multiple NDI technologies has been proposed as a means to attack new domains or improve the probability of detection (POD) in existing domains. Research and application of fusion has been inhibited by a lack of a cohesive framework for applying this computational technique. In this paper, we present a data fusion application methodology derived from the military applications of multi-sensor identification. We show that with the proper selection of NDI technologies and data fusion methodology, complementary information could be obtained and combined to produce more accurate and reliable corrosion assessment. NDI data fusion greatly improves accuracy in defect characterization and the overall performance of a NDI evaluation. In our framework for fusion application to NDI, we organize the processing into pipelines with options for fusion at three levels: the signal level, the feature level, and the interpretation level. The processing pipelines in this NDI framework consist of steps such as signal digitization, digital image processing, image transformation using wavelets, and classification and analytical technique application for interpretation, with the opportunity for fusion at each step. In this paper, we present the overall framework. We demonstrate the application of that framework using two of the possible approaches – Image Fusion and Interpretation Fusion. Our examples are based on developments to increase robustness and the probability of detection in the aircraft lap joint corrosion detection and damage assessment domain. INTRODUCTION There are various challenges and serious problems faced by both the civilian and military industries in monitoring corrosion in aging aircraft fleet. For the structural integrity of aging aircraft, lap joints are one of the major areas of concern for corrosion damage. The trapped moisture inside the lap joints of the aircraft is one of the major factors for triggering corrosion. Eastaugh et al. demonstrated that less than 5% of thickness loss could reduce the fatigue life in certain lap joint specimens up to 50%. Currently, there are no proven structural models that can predict the limits on corrosion level that can be tolerated for airworthiness. The FAA has directives, which prohibit the operation of aircraft with greater than 10% material loss in lap slice joints. The KC-135 aircraft contains some 900 feet of lap splice joints and 1000 square feet of doublers that are also possible sites for hidden corrosion. Reliable prediction of corrosion is one of the single most important aspects for aging aircraft systems. The inspection of all these areas on the fleet of more than 500 KC-135 aircraft operated by USAF will generate * President, [email protected] † Researcher, [email protected] ‡ Researcher, [email protected] § Project Manager, [email protected] an enormous amount of data. The cataloging and management of the assessment results is essential to fleet management and life-cycle engineering. A general trend in NDI technology is the use of multiple sensors of various kinds for inspection. Generally, NDI techniques are classified based on the energy source that is utilized for evaluation. For example, the NDI techniques can be categorized as acoustic, electromagnetic, optical, x-ray, and even physical methods such as liquid penetrants. Each sensor has its advantages and limitations, which may result in each capturing only a limited dimension of corrosion features. For example, Ultrasonic and Eddy Current are the two most popular and widely used sensors for corrosion evaluation. While the Ultrasonic method is based on wave propagation of energy, the Eddy Current inspection is an electromagnetic phenomenon, and it can be modeled as a diffusion process. Ultrasonic techniques provide good resolution for thickness loss in plates due to corrosion, but these techniques are sensitive to uneven scanning surface and coupling as the probe needs to have good contact with the test specimen. Eddy Current methods overcome some limitation of Ultrasonic techniques as it does not require contact thus making it insensitive to surface roughness. But, it has resolution problems, and it is sensitive to “lift-off’ variations. Thus, single sensor inspection is not adequate enough to cover all of the different dimensions of corrosion. However, the use of multiple NDI techniques requires a means of integrating the results into a coherent whole. Our research supports this fusion of the information from the various sources. Research has shown that fusion can be achieved at different stages in the process starting with the data level (where sensor data is fused to effectively provide a higher signal to noise ratio), at the feature level (where indicator features extracted from each technique are fused to provide more accurate feature-toevidence mapping), and finally at the evidence level (where evidence from multiple sources (including non-sensor based life-cycle information) can be combined to provide more reliable directed insight). The essential requirement for the fusion-based corrosion prediction approach is multiple NDI sensor data sets registered onto a single coordinate space. According to Hall & Llinas, there are three basic alternatives currently used for multi-sensor data fusion: (1) direct fusion of the sensor data at the signal level, (2) fusion of representative parameters of the sensor data via feature vector (FV) fusion, and (3) fusion of the decisions that are made with individual sensor input. The methods and goals of fusion are summarized below in TABLE 1. TABLE 1 FUSION METHODS Signal/Pixel Level Feature Level Decision Level Methods Estimation/Combination Correspondence, attribute combination Logical/statistical inference Improvement Expected variance, increase equality, performance Increase feature measurement, value of additional features Increase in truth or in probability values Sensor data or pixel level fusion is potentially the most accurate method, but this method is effective only for identical or commensurate data. The data at various resolutions cannot be directly fused at the sensors level. Methods based on decision level fusion can combine information from arbitrary sources and are most generic, but this approach, in a sense, “kills” significant information contained in the data acquired by the sensors. Our goal in this effort was to develop a compendium of synthetic techniques that combine mathematical and machine-learning methods based on all three approaches for extracting features of the corrosion pattern from multiple sensors (Ultrasound, Eddy current, Thermal, Xray). The features extracted from different sensors will then be fused for better corrosion classification. The objectives of the paper were (i) to develop a data fusion approach that will integrate various NDI data and derive useful knowledge that is not available from any single NDI technique, (ii) to develop wavelet-based approaches for image processing and feature extraction, and (iii) to develop a methodology in support of corrosion assessment based on data and model fusion The main results of the paper are described below. The fundamental idea behind our methodology is to increase robustness and the probability of detection of corrosion from multiple sensor feeds through data fusion.