To address the challenges of high deployment costs and limited edge computing power in intelligent retrofitting of small and medium-sized vessels, this study proposes a lightweight intelligent operation and maintenance (O&M) system architecture based on the digital twin five-dimensional model framework. The architecture features a five-layer collaborative architecture comprising “physical entity—virtual entity—service system—twin data—connection.” At the physical entity layer, a low-cost multi-source data acquisition platform is implemented using an Arduino Uno R3 integrated with a DS18B20 temperature sensor and SW-420 vibration sensor. The virtual entity layer employs the Three.js engine to achieve real-time 3D scene rendering and virtual-real mapping, reducing model loading time from 3.2 seconds to 1.8 seconds while maintaining an average rendering frame rate of 24 FPS. The service system layer integrates functions for status monitoring, threshold alerts, and fault classification inference. The twin data layer combines 500 samples from public datasets and laboratory-collected data to construct a four-dimensional feature space. The connection layer enables end-to-end data synchronization via USB serial port and WebSocket protocol. For fault diagnosis, a random forest classifier performs ternary classification of equipment status (normal/overtemperature/anomalies), achieving an overall accuracy of 71% and a normal state recall rate of 78% after hyperparameter optimization. Analysis of the physical mechanism underlying the confusion matrix reveals that the cross-false judgment between vibration anomalies and excessive temperature stems from the frictional heat generation coupling effect induced by bearing wear, highlighting the limitations of feature separability when relying on a single sensor dimension in multi-fault-mode coupling scenarios. From the perspective of the Nyquist sampling theorem, this study demonstrates the fundamental constraints of the current 1 Hz sampling rate on vibration feature capture. The prototype system’s hardware cost is kept below ¥200, with power consumption under 5W, verifying the technical feasibility of implementing core digital twin functionalities under constrained edge computing resources.