Recent developments in wind energy extraction methods from vortex-induced vibration (VIV) have fueled the research into vortex shedding behaviour. The vortex shedding map is vital for the consistent use of normalized amplitude and wavelength to validate the predicting power of forced vibration experiments. However, there is a lack of demonstrated methods of generating this map at Reynolds numbers feasible for energy generation due to the high computational cost and complex dynamics.

Leveraging data-driven methods addresses the limitations of the traditional experimental vortex shedding map generation, which requires large amounts of data and intensive supervision that is unsuitable for many applications and Reynolds numbers. This thesis presents a data-driven approach for generating vortex shedding maps of a cylinder undergoing forced vibration that requires less data and supervision while accurately extracting the underlying vortex structure patterns.

The quantitative analysis in this dissertation requires the univariate time series signatures of local fluid flow measurements in the wake of an oscillating cylinder experiencing forced vibration. The datasets were extracted from a 2-dimensional computational fluid dynamic (CFD) simulation of a cylinder oscillating at various normalized amplitude and wavelength parameters conducted at two discrete Reynolds numbers of 4000 and 10,000. First, the validity of clustering local flow measurements was demonstrated by proposing a vortex shedding mode classification strategy using supervised machine learning models of random forest and 𝑘-nearest neighbour models, which achieved 99.3% and 99.8% classification accuracy using the velocity sensors orientated transverse to the pre-dominant flow (𝑢𝑦), respectively. Next, the dataset of local flow measurement of the 𝑦-component of velocity was used to develop the procedure of generating vortex shedding maps using unsupervised clustering techniques. The clustering task was conducted on subsequences of repeated patterns from the whole time series extracted using the novel matrix profile method. The vortex shedding map was validated by reproducing a benchmark map produced at a low Reynolds number. The method was extended to a higher Reynolds number case of vortex shedding and demonstrated the insight gained into the underlying dynamical regimes of the physical system. The proposed multi-step clustering methods denoted Hybrid Method B, combining Density-Based Clustering Based on Connected Regions with High Density (DBSCAN) and Agglomerative algorithms, and Hybrid Method C, combining 𝑘-Means and Agglomerative algorithms demonstrated the ability to extract meaningful clusters from more complex vortex structures that become increasingly indistinguishable. The data-driven methods yield exceptional performance and versatility, which significantly improves the map generation method while reducing the data input and supervision required.