Deployment of a damage diagnosis model for offshore jacket platforms in embedded systems

Deep learning-based structural health monitoring for offshore jacket platforms remains difficult to implement on small-scale maritime edge devices due to limited energy supply, constrained communication bandwidth, and restricted computational resources. This study presents an engineering deployment pathway that enables the implementation of a stacked sparse autoencoder with a Softmax classifier on an Android-based embedded platform for damage diagnosis. To improve deployment efficiency under resource constraints, asymmetric quantisation and structured architectural reduction are employed to compress the trained model while preserving diagnostic performance. Experimental validation on the ATK-DLRK3568 development board demonstrates a nine-class damage classification accuracy of 91.67%, with only a 0.03% deviation from the PC-based implementation. Meanwhile, average power consumption is reduced to 3.67 W, memory usage is substantially lowered, and processor utilisation remains stable during operation. A sliding-window alarm mechanism and model health self-assessment strategy are further incorporated as operational design components intended to support long-term unattended operation. The results demonstrate the feasibility of deploying deep learning-based damage diagnosis on resource-constrained offshore edge-computing platforms and provide a practical pathway for embedded structural health monitoring applications. • We constructed a general-purpose model migration pipeline to form a full-process embedded deployment solution. • A dedicated model retraining and update mechanism was established to support long-term autonomous offshore operations. • The power consumption decreased from 85.32 W to 3.67 W, and the peak memory usage was reduced from 1745.6 MB on a PC to 131.5 MB. • The model deployed on Android devices achieved an accuracy of 91.67%, merely 0.03% lower than the PC baseline. • The model size was compressed by 87.1%, facilitating efficient edge-side deployment.