Fatigue crack monitoring of helicopter fuselage through sensor network

Developing a holistic approach for looking at the structure’s integrity through real-time Structural Health Monitoring (SHM) defines the core of the project. This work argues the implementation of the real-time SHM for an isolated panel of the rear fuselage of a helicopter. This work initially presents an effective statistical approach to detect structural damage utilizing a novelty detection method based on multidimensional outlier analysis (OA). Going further into the hierarchy of damage identification, together with detection, quantification of the damage intensity and localization of the damage site are performed adopting a multi-layer perceptron (MLP) neural network structure. Prior to the establishment of algorithms of any method (OA or MLP) a thorough study regarding the identification of the provided database is lunched. Concerning outlier analysis, the necessary database is acquired from experimental tests, consisting of sampled strain measures for both undamaged and damaged scenarios coming from a network of Fiber Bragg Grating (FBG) optical sensors. Due to the high expenditure of experimental test, comprising the costs of both operating equipments and the one-time-use test specimen, the only available experimental database regarding the damaged case is set to be a skin crack propagating from the center of the central bay. To extract the damage sensitive features out of the sensors readings, a normalization is performed where changes in sensor readings caused by damage are separated from those caused by varying operational and environmental conditions. Moreover, an initiative is taken to investigate the influence of possible localized stress variations (this can occur due to localized loads and temperature variations present within the panel) over the damage indices. This is of a great use as it firstly allows for a visual representation of the influence and secondly becomes of a use later in outlier analysis to check the flexibility and efficiency of the novelty detection algorithm. In this thesis, the outlier analysis (OA) is conducted mainly for lowest level of fault detection (i.e. damage detection) so that the method is simply required to detect signal deviations from normal condition; i.e., the problem is the one of novelty detection. Therefore, the concept of discordancy from the statistical discipline of outlier analysis is used to identify signal deviance from the norm. Since the acquired database of the case study is multivariate, the discordancy test is performed by using Mahalanobis square distance measure to calculate the novelty values relevant to each observation of the database. In applying OA, the novelty values are finally compared against a threshold which allows to group novelty indices as novel or normal corresponding to damaged and undamaged cases respectively. To develop a diagnostic unit that includes quantification and localization levels of a damage state, as well as the damage detection level, MLP neural networks are implemented. Here, the first challenge is to provide the ANNs with sufficiently large training database which can be considered as a good representative of all possible damage cases that can occur on the panel. For that reason, a database simulated through a Finite Element (FE) model is used. When applying artificial neural networks as the performing algorithms of diagnostic unit, there are several parameters that are considered at each diagnostic level to optimize the performance of the diagnostic algorithm. The first parameter to be optimized is the structure of an ANN and in specific the number of hidden layers and nodes. Two other optimizing parameters are considered which are “introduction of additive Gaussian noise to the training database” and “optimal training database size”. Considering the classification algorithm for anomaly detection, the concept of “optimal noise” is applied upon redefining the state of damage, while the effect of additive noise in the performance of damage quantification an localization algorithms is simply carried out by checking the corresponding Root Mean Square Error (RMSE) in predictions. Finally, to verify the algorithms applicability to the real diagnostic system, a real experimental centre crack propagation is fed to all three layers of the diagnostic algorithm.