Data-driven and hybrid models for the underwater radiated noise of cavitating marine propellers

Student thesis: Doctoral Thesis

Abstract

The sustainability of anthropogenic activities is a fundamental problem requiring a multidisciplinary approach in order to be properly addressed. Recently, underwater radiated noise has been categorized as a form of pollution, due to the substantial increase of underwater noise levels on oceans worldwide, with severe effects on the marine ecosystem. For propeller-driven vessels, cavitation is the most dominant noise source, producing both structure-borne and radiated noise. As such accurate predictions of the noise signature are fundamental for the design of silent, yet efficient, propellers.In this respect, this work investigates a novel hybrid (combined physics-based and data driven) model for the prediction of underwater radiated noise of marine propellers. By relying on both the engineering knowledge (through the physics-based model), and advanced statistical inference procedures (through the data-driven model), the hybrid model will be capable of providing an accurate, yet computationally cheap, assessment of the noise levels emitted by a cavitating marine propeller. The proposed model relies on a novel hybridization strategy that is able to truly blend the knowledge of the underlying physical phenomena with information contained in historical data. This strategy allows the development of that models able to properly, i.e., physically plausibly, extrapolate as physics-based models, while being extremely accurate and computationally inexpensive as data-driven models. In particular, knowledge of the underlying physical phenomena is leveraged during model structure, model building, and in model enrichment: a dedicated feature engineering process is considered to extract meaningful information from available experimental data, as well as the noise estimates of a computationally efficient physics based model. Everything is empowered by state-of-the-art learning algorithms from the field of Machine Learning that take advantage of all information sources.The proposed model is tested on a series of complex extrapolation scenarios, in which the numerical predictions are compared with measurements collected in an extensive experimental campaign conducted at the Emerson Cavitation Tunnel of Newcastle University. The results support the feasibility of the proposed approach in all scenarios considered. The proposed model shows enhanced capabilities in predicting the underwater radiated noise levels: It commits low errors that are certainly acceptable during the early stage design process, and delivers predictions that are in agreement with state-of-the-art engineering knowledge of the underlying physical phenomena.
Date of Award20 Mar 2023
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorGerasimos Theotokatos (Supervisor) & Andrea Coraddu (Supervisor)

Cite this

'