Conference Agenda

Bridge managers and operators often oversee assets distributed in different geographical areas; thus, they are responsible for multiple similar structures exposed to different environmental conditions. Population-based SHM (PBSHM) provides a framework to share knowledge within a population of related structures, supporting the detection of anomalies on target structures for which less data might be available. However, the variability in environmental exposure introduces a significant challenge to the success of PBSHM approaches, as it can compromise the ability of algorithms to generalise across structures and maintain reliable performance under previously unseen environmental conditions.

To investigate strategies for a reliable knowledge transfer within a PBSHM framework despite differences in environmental conditions, a series of experiments were conducted on a modular laboratory-scale bridge. This paper presents the experimental setups and resulting datasets, which encompass multiple healthy and damaged configurations under controlled temperature variations. Three bridge configurations have been tested, differing in span length and in the number of beams of the deck. Damage was introduced by cutting a lateral beam in a reproducible way. The model was installed on a multi-axis shaking table inside a climatic chamber. Data acquired consists of accelerations and temperature measurements.

The resulting dataset forms a benchmark that can support studies for the development of more accurate and reliable data-driven anomaly detection tools to aid in the monitoring and condition assessment of real bridges in operation.

The methodology of Population-based Structural Health Monitoring (PBSHM) provides the framework for enhancing the overall knowledge of a structure. Structures are determined to be similar - thus a population - through a variety of graph-based methods. Once established, knowledge is then taken from one structure in the population to another.
Individual structures are often viewed within an industrial context as a part of a portfolio of assets within a company. These assets are regularly managed within an Infrastructure Management System. Adopting PBSHM technologies within industry will inevitably require substantial resources; hence, there must be a clear business case before any integration can begin.
This paper explores the initial steps towards maturing PBSHM into an industry-ready technology. If industry are already using Infrastructure Management Systems, how does one create a 'PBSHM-enhanced IMS'. A tiered hierarchy is introduced to breakdown how ready an organisations asset pools is for PBSHM. This hierarchy is then utilised to provide industry focussed examples to where PBSHM could add insight to existing workflows.

A recent development in Structural Health Monitoring (SHM) has been the introduction of a population-based variant (PBSHM), which aims to leverage information across populations of structures in order to enhance diagnostics on those with few (or no) data. The machine-learning discipline of transfer learning provides the mechanism for this capability. Recent work by the authors has developed the mathematical machinery for defining ‘interpolating structures’, which allow transfer-learning problems to be broken down into multiple transfers, each between structures more closely matched to each other. The current paper proposes a mathematical/geometrical framework for the ideas, which allows – among other things – a means of showing that increasing similarity scores between structures leads to lower distances in the feature spaces and thus, more effective transfer.

This study presents the energy distance non-parametric approach for condition monitoring of multiple wind turbines within the same wind farm. The main objective of this work is to monitor the performance of wind turbines by comparing them against each other within the same farm. Traditional methods often rely on parametric assumptions, including normality, which are rarely satisfied in real-world applications. In contrast, the energy distance approach is fully non-parametric, does not assume normality, and is effective for distributions of any shape or location. It enables comparison at the distributional level without requiring explicit assumptions. To validate the proposed approach, SCADA data from a wind farm in Portugal was analyzed. Four wind turbines, including the unit operating in a fault condition, were comparatively analyzed using multiple parameters through the application of the Energy Distance method. The results demonstrate that this method provides a robust means of detecting faults earlier than traditional approaches, as faulty turbines show clear deviations from healthy turbine states. The analysis considered multiple parameters, including generated power, wind speed, hydraulic oil temperature, gearbox oil temperature, generator bearing temperature 2, gearbox bearing temperature, and others. Overall, the results of the study clearly indicate that when comparing the four wind turbines using the energy distance method, the turbine with a fault exhibits a significantly larger energy distance relative to all other turbines. This distinct difference effectively distinguishes the faulty turbine from the healthy ones. These findings confirm that the energy distance approach, based on distributional comparisons, is a robust, efficient, and easily interpretable method for the condition monitoring of wind turbines.

With migration towards fleetwide monitoring, wind farm operators are installing tri-axial MEMS accelerometers in the nacelle of wind turbines. The nacelle installation allows measurements of accelerations in three orthogonal directions: parallel to wind direction (fore-aft), perpendicular to wind direction (side-side), and the gravitational component. The primary use of those sensors are to measure vibrations and resonance frequency monitoring utilizing Operational Model Analysis (OMA). However, recent research has shown that the acceleration data can be utilized to infer the fatigue rates acting on the substructure via fleetwide machine learning techniques (de N Santos, D’Antuono, Robbelein, Noppe, & Weijtjens, 2023). This enables the same sensors to serve multiple purposes, providing a population-based overview of key parameters across the farm. Extending on the application of the same MEMS accelerometers, inclinations can be inferred by leveraging the tri-axial properties of the sensors. In theory, at rest those accelerometers only measure gravitational force. Once tilt is introduced on the sensor due to a combination of sensor installation offsets, bending moment induced angles on the nacelle and static tilt of the structure, the gravitational component redistributes across the axes of the sensors. Linking the final sensor measurement with the rotational components of the different coordinate systems (tower, nacelle and sensor), the contribution of the different angles can be inferred.

This paper proposes using direct measurements from turbines with extended set up of strain gauges and inclinometers (fleet leaders), to establish a link between direct measurement of the bending moments and inclinations to the accelerometers of the nacelle, and to propagate this mapping to turbines equipped only with tri-axial MEMS accelerometers. As presented in Figure 1, the bending moment and the measured acceleration in the Z-direction of the fleet leaders are plotted versus windspeed. The acceleration measurements mimics the behavior of the loading, with the curves being almost identical. The differences between those curves are due to the different rotational angles induced on the sensors. Studying those variations on fleet leaders, the angles can be inferred by curve fitting techniques. This enables turbines equipped solely with MEMS accelerometers to obtain estimates of their load characteristics, as well as their static structural inclination.