54 G. P. Tsialiamanis et al. To further explain the content of the aforementioned classes, their description and some of their subclasses will be presented subsequently. The “analysis models” class contains anything that is used to analyse a structure. Since many types of models are used in various structural health monitoring techniques, like monitoring the modal assurance criterion (MAC) [13] or neural networks [14] to classify/localise damage. The first partition is into the classes of “physical models” and “data-driven models”. “Data-driven models” is a class that mainly contains machine learning techniques. These are models that simulate physics, but infer their output using existing data from structures. These models are partitioned mainly into regression models and classification models. The first of these classes is composed of models predicting values of variables (e.g. natural frequencies, displacements, etc.) and are more commonly used in simulating the behaviour of structures, whilst the second one is often used to discriminate data corresponding to damaged or undamaged states of the structure. Some of the included models are neural networks [15], K-means clustering [16], support vector machines [17], autoassociative neural networks [18] etc. Physics-based models are constructed by studying and explaining the physics of problems. Two major subclasses of this class are the analytical and numerical models. Numerical models contain finite element models, surrogate models, lumped mass models etc; they are the most often-used ones, since they can be simulated using computers. On the other hand, the analytical models are solved using calculations and are not perhaps as common as the numerical ones, since engineering problems often preclude analytical solutions; however, they are still included in the ontology for completeness. The “Data” superclass contains all the data that are acquired from structures. The data are partitioned in classes according to their type. For example, acceleration, displacement and velocity data, which belong to the time domain, but also data in the frequency domain. Furthermore, the data are also separated into data from sensors (“raw data”), data from processing the sensor data (“processed data”) and data from any models (“simulation data”). “Data” is probably the class that can be more easily transferred into ontologies describing other domains, since most processes nowadays produce data and their analyses are based on data. Additionally, data can be transferred from one SHM application to another. Analysing data from a structure and making inference about them, can help in understanding the behaviour of similar structures or materials used in another SHM application. Some subclasses included in this class are “accelerograms”, “contour plots”, “mode shapes”, “displacement simulation data”, “displacement sensor data”, etc. The next superclass is related to methods that are used to process the data. These methods mainly belong to the discipline of signal processing, from the simple and ubiquitous Fourier transform [19] and signal statistics extraction, to more complicated methods like wavelet decomposition of signals [20]. These methods are vital for SHM, because they may reveal unseen features of the signals that are sensitive to damage. It is quite common that a damaged state can be spotted by observing a spectrum (which is produced by performing a Fourier transform on the acceleration signal of a sensor) rather than the acceleration time-history itself. Methods included in this class are: “signal smoothing algorithms”, “Hilbert transform”, “principal component analysis”, etc. The fourth superclass is the “physical parts” that the all methods refer to. This class simply contains instances of the structures that are monitored and the sensors placed on them. The instances are just objects referring to an existing structure. The structures belonging to this class can be partitioned in substructures for a more detailed and modular representation of a greater object. For example a wind turbine can be partitioned into its blades and the tower and even deeper, the blade can be partitioned into the cell and the stiffeners existing inside. This class serves the purpose of registering all monitored physical objects but also separating them into classes to facilitate search for data from a specific structure or type of structures. Last but not least, is the class containing all methods that are used in SHM. This class is the most important one, as it uses components from every other class and combines them into methods that monitor a structure. Some of the methods here are trivial, such as monitoring the maximum value of a sensor signal; but there are also much more sophisticated methods, like monitoring the modal force error on a structure, that require data from the “data” class, finite element models from the corresponding class and also data processing methods. All methods included refer to a specific structure, use data from sensors, process them with a processing method and make inference about the current situation of the structure according to a model from the class “analysis models”. Many types of SHM methods are included in the current ontology. They are partitioned into subclasses according to the type of data they use and the types of algorithms/methods used to make inferences. For example, there are the “acoustic emission monitoring” [21] and “guided wave approaches” [22] whose data come from receivers trying to “hear” cracks in materials and ultrasound receivers correspondingly. Other classes of methods are the “novelty detection methods”, which are unsupervised ones [23], trying to identify changes in the behaviour of structures, having as inputs only normal condition/undamaged state data. Major classes of this superclass can be seen in Fig. 6.3. It should be noted that for the specific ontology, this list is exhaustive, but it can definitely be extended with more SHM methods.
RkJQdWJsaXNoZXIy MTMzNzEzMQ==