Research & scientific contributions

SHM is an interdisciplinary subject that encompasses the fields of physics (structural mechanics), statistics and data science (system identification and machine learning), and electrical engineering (sensors and actuators, sensor networks, internet of things).

My role in this research area relates to the development of system identification and damage diagnosis techniques able to capture the variation in the dynamic characteristics of structures, for instance, resonance frequencies and damping ratios.
These variations may have either a benign origin, due to changing environmental and operational conditions or arise from unwanted conditions, like deterioration, damage, or failures.

While developing these techniques, I aim at helping us understand how structures operate through their lifespan and how natural deterioration processes lead to damage and failures, while improving our capability to detect, locate and assess the severity of damage in a structure.

Long-term analysis of the vibration response of structures

When observed over long analysis intervals, structures will display large variations in their dynamic properties due to the constantly evolving environmental and operational conditions. This may have dramatic effects on their lifetime performance (fatigue accumulation) and damage diagnosis. Back from my PhD years, I’ve been working on different topics associated with this type of slowly-evolving non-stationary behavior. Some of these topics are discussed further below.

Significant changes on the vibration response characteristics are observed when monitoring the vibration response of these structures over long analysis intervals. In this research topic, I’m interested on the development of time-series models capable of tracking the long-term variations on the vibration response of structures. These time-series models take a hierarchical structure, in which short-term dynamics are represented via conventional stationary time-series models, while long-term variations are tracked by changes on the basic model parameters. The obtained models take the form of Linear-Parameter Varying (LPV) models, where the parameters depend on measurable environmental and operational parameters. In this regard, I have developed different approaches based on functional basis expansions and Gaussian Process regressions aimed at tracking the variations in the parameters of AutoRegressive Moving Average time-series models. The obtained models provide a global representation of the vibration dynamics which facilitates the analysis of the underlying dynamics (in frequency and modal domain) and the development of SHM algorithms.

Damage detection methods in SHM are based upon the calculation of Damage Sensitive Features (DSFs), which are features carrying information on the condition of the structure. However, DSFs are also sensitive to Environmental and Operational Variability (EOV), in the same way as they are sensitive to damage. As a result, damage is often masked by the effects of EOV, resulting in a reduced performance of SHM algorithms if not treated properly. Together with some colleagues, we have focused on developing techniques for (implicitly or explicitly) compensate the effects of EOV in DSFs. Implicit approaches attempt at capturing or rejecting benign sources of variation in the normal state of the structure, based solely on the characteristics of DSFs themselves. These include linear transformation methods (PCA or factor analysis), time-series modeling and cointegration analysis, and manifold learning methods. On the other hand, explicit approaches attempt at compensating the variations in DSFs by correlating with measurable Environmental and Operational Parameters (EOPs). Various regression methods are used for this purpose. Best results are often paired to the specific problem at hand, while a proper combination of techniques can outperform a single methodology.

Vibrations are of high interest in the design of offshore wind turbines (OWTs), as fatigue in all components must be limited to extend the operational life of these structures. Therefore, a lot of research has aimed at determining the vibration damping properties of OWTs, as it ultimately determines the levels of vibration. However, there is a large uncertainty in damping calculations in analytical models and from vibration signatures. It is equally unclear the extent in which complex vibration phenomena, such as Vortex Induced Vibration (VIV), occur in actual operating conditions. By developing methods and automatically analyzing vibration and environmental data in a database from more than 20,000 wind turbines in Siemens Gamesa Renewable Energy (SGRE) (more than 3400 OWTs) this project will obtain validated damping values to use in future design of OWTs, and provide data on actual vibrations in OWTs. This will enable more optimized design to reduce cost of OWTs and will also allow assessing the remaining lifetime of existing OWTs, achieving reduced levelized cost of energy (LCOE).

Anomaly detection and useful lifetime assessment based on operational data

Complex structures are often instrumented with several sensors monitoring hundreds of variables related to the operation of structural components and systems. The obtained data is characterized by large volumes, but also is heterogeneous and complex. However, important information on the structural condition and remaining lifetime of components could be extracted with the application of proper data analysis algorithms. The selected topics below are examples of such methods, where we use different machine learning algorithms to make sense of the operational data acquired in wind turbines and marine engines.

Load analysis is essential for the assessment of the fatigue life of wind turbines. Ideally, wind parks should have dense sensor networks capable of measuring loads at several structural components. However, this is far from practical, as the cost of installation and maintenance, as well as the associated computational resources, would take incredible dimensions. Likewise, installation of sensors at some locations, like portions of the foundation under the waterline, is just not feasible. Otherwise, sparse sensor networks, where sensors are located at a limited number of wind turbines and components, are more viable. Moreover, fatigue life analysis could be still performed in non-instrumented components, with the help of virtual sensor techniques, where loads are predicted/estimated based on measurements performed elsewhere. While loads are mostly determined by the incoming wind characteristics, other effects related to the interaction with the wakes of other wind turbines, increase the complexity on the prediction and estimation of loads. In this research line, we develop machine learning methods that make use of environmental and operational parameters, as well as measured loads to provide load estimates at different components of wind turbines located in a wind park.

Marine engines are subject to incredible operational demands, which increase the fatigue accumulation and deterioration of different components and sub-components. At the same time, operators must guarantee continuous availability to ensure revenue and avoid unnecessary costs. Condition-based maintenance has emerged as a promising paradigm, where Operation and Maintenance (O&M) activities are scheduled on the basis of the actual condition of components. In practice, condition-based maintenance requires an estimate of the unknown component condition and its remaining useful lifetime. In this PhD project, we attempt at developing statistical models that can correlate different operational variables (internal engine temperatures and pressures, RPM, etcetera) with the actual degradation level of engine components, as measured by actual condition indicators.

Machine Learning for Condition Monitoring of Wind Turbine Drivetrains

Vibration-based Condition Monitoring (CM) is based on the analysis of the vibration responses measured on rotating machinery. Such vibration responses are quite complex, and as a result, modern CM algorithms make use of advanced machine learning methods based on deep learning to obtain reliable damage assessment. The construction and training of ML architectures is not straightforward, particularly for wind turbine drivetrains, where the highly variable operational conditions and the lack of information on damages are main challenges. The topics below correspond to recent project ideas, where we attempt to overcome these hurdles.

Condition Monitoring (CM) systems are essential for the optimization of operation and maintenance activities and lifetime extension of Wind Turbine (WT) drivetrains, and as a result can have a significant impact in the cost of energy production. Current CM algorithms are challenged by the complexity of vibration signatures and the amount of potential failures, and poorly deal with the continuously changing operational conditions of WTs. Moreover, as real WT vibration data is scarce, transferability of algorithms from lab-scale to real-life structures is critical, but still remains unresolved. In AdaptCM, we aim towards CM algorithms for WT drivetrains, which through adaptability to operational and inter-machine uncertainty, will provide high performance, robustness and transferability. To this end, smart integration of vibration and operational variables and explicit recognition of non-stationarity in deep learning architectures are our primary tools.

Deep Learning (DL) provides the best diagnostic performance in CM so far but is limited by the quality of the datasets used for training of the algorithms: while data from the healthy state may be abundant, data from damaged conditions is extremely scarce in comparison. This, in addition to the large variability on the characteristics of the vibration response on the healthy condition, arising from the highly uncertain operational conditions in which WTs operate, makes the training of effective and generally applicable ML algorithms for CM a very challenging task. The main aim of this project is to develop and implement physics-based ML-enabled algorithms for simulation of faulty WT drivetrain vibration data that can be used in tandem with real-life healthy condition data for training of highly effective and generally applicable CM algorithms. The novelty of our approach lies on the combination of high-fidelity physical modelling of isolated WT drivetrain components with ML techniques to couple simulations with the features of real-life data.

Want to

collaborate?

Let’s build something great together

Get in touch