Machine learning approach to construct global phase-averaged flow field based on local flow features

Abstract

This paper reports a new approach toward constructing a full-domain phase-averaged flow field; the approach applies particle image velocimetry (PIV) measurement without referring to time-resolved signals. This approach is a departure from the conventional phase-averaging method based on proper orthogonal decomposition (POD). The POD-based method requires the full flow field to be covered during the measurement, which can result in a low spatial resolution. The proposed method combines multiple local flow fields in different subdomains with high spatial resolution to construct a full phase-averaged flow field. The local flow fields are phase-identified using a machine learning approach, namely, k-nearest neighbor (KNN) classification. Prior to the classification, the full flow fields are first measured with low spatial resolution and then phase labeled by POD analysis. Then, the full flow fields are divided into multiple local flow fields. The major flow features of the phase-labeled local flow fields are extracted by POD again. And, the produced POD coefficients serve as training samples. Subsequently, the new local flow fields are measured with high spatial resolution and then phase-identified by comparing to the training samples using KNN classification. Finally, a full-domain flow field is constructed by combining the phase-averaged local flow fields with high spatial resolution. In the application of KNN, the training samples can be better differentiated in a high-dimensional space defined by the POD modes.