Objective To address the difficulty of directly using raw monitoring data for artificial intelligence (AI) training in prognostics and health management (PHM) of marine equipment, this study investigates a method for constructing an AI-ready time-series sample repository.
Method A time-series sample information model integrating multi-level annotation, complete data provenance, and multidimensional quality descriptions is proposed to provide a unified representation of heterogeneous monitoring data and their associated semantic and quality information. On this basis, a three-level quality assessment framework covering channels, samples, and datasets is established to quantitatively characterize data quality from different perspectives. Furthermore, targeted learning strategies are developed for several typical data-quality problems, including scarce annotations, label noise, temporal alignment deviations, and class imbalance, so as to enhance the robustness of AI models under imperfect data conditions. An edge–cloud collaborative service architecture and a standardized AI-ready workflow are also designed to support efficient data processing, sample management, model training, and online inference.
Results A fault diagnosis case study involving the thrust bearing of a marine main engine demonstrated that the proposed information model could effectively organize multi-source heterogeneous time-series data with sampling rates spanning five orders of magnitude. The Cohen’s kappa coefficient between the automated quality assessment results and expert review results reached 0.886, indicating a high level of consistency. With the proposed learning strategy for low-quality samples, the model performance retention rate on degraded datasets reached 92.1%, demonstrating improved robustness against data-quality degradation. In addition, the edge–cloud collaborative architecture reduced average daily bandwidth consumption by 98.7%, while the latency of a single inference task at the edge was only 78 ms, showing good potential for real-time deployment.
Conclusion This study establishes a methodological framework for constructing AI-ready time-series sample repositories for marine equipment PHM, providing a reference for data organization, quality governance, and service implementation in related applications.
DownLoad: