5 On Partitioning of an SHM Problem and Parallels with Transfer Learning 43 3. What kind of damage is present (type/classification)? 4. How severe is the damage (extent/severity)? 5. How much useful (safe) life remains (prognosis)? A common approach to the first level is to observe the structure in its normal condition and try to find changes in features extracted from measured signals that are sensitive to damage. This approach is called novelty detection [10, 11], and it has some advantages and disadvantages. The main advantage is that it is usually an unsupervised method, that is only trained on data that are considered to be from the undamaged condition of the structure, without a specific target class label. These methods are thus trained to detect anychanges in the behaviour of the elements under consideration, which can be a disadvantage, since structures can change their behaviour for benign reasons, like changes in their environmental or operational conditions; such benign changes or confounding influences can raise false alarms. In this work a problem of damage localisation is considered (at Level 2 in Rytter’s hierarchy [9]); the structure of interest being a wing of a Gnat trainer aircraft. The problem is one of supervised-learning, as the data for all damage cases were collected and a classification model was trained accordingly. Subsequently, the classifier was used to predict the damage class of newly-presented data. The features used as inputs to the classifier were novelty indices calculated between frequency intervals of the transmissibilities of the normal condition of the structure (undamaged state) and the testing states. The transmissibility between two points of a structure is given by equation (5.1), and this represents the ratio of two response spectra. This feature is useful because it describes the response of the structure in the frequency domain, without requiring any knowledge of the frequency content of the excitation. The transmissibility is defined as, Tij = FRFi FRFj = Fi Fexcitation Fj Fexcitation = Fi Fj (5.1) where, Fi is the Fourier Transform of the signal given by the ith sensor and FRFi is the Frequency Response Function (FRF) at the ithpoint. The experiment was set up as described in [12]. The wing of the aircraft was excited with a Gaussian white noise using an electrodynamic shaker attached on the bottom surface of the wing. The configuration of the sensors placed on the wing can be seen in Fig. 5.2. Responses were measured with accelerometers on the upper surface of the wing, and the transmissibilities Fig. 5.2 Configuration of sensors on the Gnat aircraft wing [13]
RkJQdWJsaXNoZXIy MTMzNzEzMQ==