Monitoring the health of aeronautical structures
Traditional non-destructive tests (NDTs) for assessing damage to aeronautical structures typically include ultrasonic waves or x-rays. These techniques have the disadvantage that they require the disassembly of the part to be tested, expensive equipment, and expert operators.
A new NDT, based on vibrational measurements, avoids these drawbacks.1 The test meets most of the mandatory requirements for effective health-monitoring systems while at the same time reducing the complexity of the data analysis algorithms and experimental instrumentation used. And because it obviates the need to disassemble components, the test can even be used for parts that are inaccessible.
The method we have developed relies on acquiring and comparing the frequency-response function (FRF). This is the ratio between the Fourier transform of the signal used to excite one locus on the structure being monitored, and the Fourier transform of the response to the signal acquired by a sensor at another locus, before and after damage occurs (see Figure 1). Damage modifies the dynamic behavior of a structure, affecting its mass, stiffness, and damping. Consequently, comparison of the FRF of a damaged structure with the that of a sound one makes it possible to identify, localize, and quantify the injury.
The research highlighted in this article focuses on a new FRF-processing technique that determines a representative ‘damage index’ for evaluating variation in the FRF of a structure owing to injury.2 The index calculates the ratio between the deviation in the FRF of the structure before and after damage, and the FRF of the initial (sound) structure. In this way, it provides a direct measure of the damage.
We have also developed a dedicated neural network algorithm for ‘recognition-based learning.’3 The algorithm teaches the neural network to recognize only ‘positive’ samples and to discard ‘negative’ ones. For purposes of the structural NDT, a positive sample means that the analyzed structure is ‘healthy.’ Conversely, a negative sample indicates a ‘damaged’ or perturbed state. Just as in formulating the damage index, we trained the neural network using the FRF of the healthy structure.
Once the network has been implemented and trained, an additional statistical procedure determines the threshold for operatively differentiating the two sample classes (positive or healthy, and negative or damaged). Then the network is ready to ‘classify’ the health status of a structural component.4
This new health-monitoring system has been used to identify and analyze damage to real-scale aeronautical structural components such as reinforced fuselage panels and aeronautical composite panels, and an actual ATR-72 aircraft. Corrosion, failure of linking rivets, simple cracks, impacts, and other sorts of damage have been induced in test articles.
Piezoceramic patches have been evaluated as actuators and sensors for the excitation and response signals needed to generate FRFs. The behavior of the patches has been validated using a laser-scanning vibrometer system. Experimental tests have shown the ability of both the damage index and the neural network to identify various types of injuries to a number of structures and to quantify how likely the damage is to propagate. In addition, the damage index pinpoints the location of injury to a structure.
These new techniques avert having to deliberately damage structures, to use finite element methods, or to calculate eigenvalues and eigenvectors. As such, the techniques are independent of structure and damage. They represent a step forward in implementing an automatic health-monitoring system able to identify structural damage in real time, improve safety, and reduce maintenance costs.
