Construction of a Marine Diesel Engine Fault Sample Database and Deep Learning–Based Diagnostic Framework

Marine diesel engines often operate under long-term high-load conditions, where hidden and weakly expressed faults are difficult to identify due to the scarcity of labeled fault data and the lack of high-quality samples for intelligent diagnostic models. To address this problem, this paper constructs a structured and scalable operational fault sample database for marine diesel engines and develops deep-learning-based diagnostic models tailored to multi-source non-stationary signals. First, a high-frequency data acquisition workflow is designed to synchronously record multi-channel parameters, allowing comprehensive reconstruction of engine operating conditions. A date-partitioned SQLite storage architecture is further established to support efficient long-term data management and fault traceability. Second, four typical engine fault patterns are introduced into the raw dataset through controlled fault injection, forming a balanced and diverse fault sample repository. Based on this database, a one-dimensional residual convolutional network (1D-ResNet) and a frequency-domain-enhanced two-dimensional residual network (2D-ResNet) are developed to achieve end-to-end temporal and time–frequency feature extraction. Experimental results demonstrate that the proposed deep learning models significantly outperform traditional time-domain and frequency-domain statistical feature approaches, with the 2D-ResNet achieving an accuracy of 99.16% and an average improvement of more than 10% in F1 score across several complex fault types. The constructed dataset and proposed diagnostic framework provide a solid foundation for intelligent monitoring and predictive health management of marine diesel engines and support further research on physics-informed and hybrid data–model diagnostic methods.