Back to search

National and Kapodistrian University of Athens

Predictive maintenance for vessel machinery using federated learning

Abstract

dc:description

Maritime predictive maintenance (PdM) is increasingly supported by data-driven models trained on high-frequency shipboard sensor streams. However, most existing approaches are evaluated under centralized assumptions that are difficult to satisfy in operational maritime environments. Vessel data are distributed across ships, operators, routes, and machinery states; communication links are often intermittent or costly; onboard resources vary across clients; raw telemetry may be commercially sensitive; and model outputs must be technically defensible when used to support maintenance decisions. This thesis studies maritime PdM as a deployment-constrained distributed-learning problem rather than as an isolated supervised forecasting task. The thesis first reviews maritime maintenance strategies, PdM methodologies, and federated learning (FL) approaches, identifying five recurring barriers to practical adoption: data quality and availability, scalability and onboard resource constraints, security risks, model interpretability, and communication efficiency. It then addresses these issues through four complementary technical contributions. First, it introduces a resource-aware federated distillation framework for heavy-fuel-oil purifier remaining-useful-life prediction. Clients are assigned to resource tiers, full or pruned model variants are dispatched accordingly, and collaboration is performed through prediction-space aggregation over a server-available reference set. This enables heterogeneous model capacities while reducing repeated full parameter exchange. Second, it develops a Data Space-native hierarchical FL framework in which stewards act as aggregation and audit entities. The framework combines client-level differential privacy, steward-local secure aggregation, signed hash-chained manifests, CKKS protected steward-to-server payloads, and clipped adaptive server fusion, thereby quantifying the privacy–robustness–overhead trade-off of governed collaborative training. Third, it studies interpretability for cylinder-level exhaust gas temperature forecasting. It combines a BiLSTM encoder, deterministic attention pooling, a Kolmogorov–Arnold Network based nonlinear head, and residual forecasting, and evaluates not only predictive error but also the stability of SHAP-based feature rankings across federated clients. This analysis shows that models with similar accuracy can differ substantially in explanation coherence. Fourth, it addresses decentralized maritime learning under intermittent connectivity by introducing a carbon-aware orchestration layer over gossip-style training. The control plane schedules vessel participation, communication links, compression modes, mixingweights, and resynchronization actions, enabling accuracy–communication–carbon trade-off analysis over AIS-inspired topology regimes. Across these contributions, the thesis shows that maritime PdM cannot be assessed only by the predictive accuracy of a centralized model. The learning protocol, aggregation mechanism, privacy layer, resource profile, communication pattern, carbon footprint, and explanation behavior all affect whether a model is suitable for distributed maritime use. The main conclusion is that deployment constraints should be formulated, measured, and incorporated into the design of maritime PdM systems from the outset, supporting learning methods that are not only accurate, but also more communication-aware, privacy-conscious, interpretable, and operationally accountable.

Degree

thesis:*
Grantor dc:publisher
National and Kapodistrian University of Athens
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Kalafatelis, Alexandros
  • Καλαφατέλης, Αλέξανδρος

Subjects

dc:subject × 18

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/62275

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Kalafatelis, Alexandros; Καλαφατέλης, Αλέξανδρος. Predictive maintenance for vessel machinery using federated learning. National and Kapodistrian University of Athens, 2026. http://hdl.handle.net/10442/hedi/62275