The use of underwater vehicles, either remotely operated vehicles (ROVs) or autonomous underwater vehicles (AUVs), has become increasingly relevant in the operation and maintenance (O&M) routines of offshore wind farms. This article provides a critical review of how these platforms are being integrated into inspection and maintenance tasks, contributing not only to safer and more precise operations but also to greater autonomy in challenging marine environments. Beyond the technical and operational aspects, this review highlights their growing connection with artificial intelligence, digital twins, and multi-robot collaboration. The studies analyzed indicate a progressive shift away from conventional methods, traditionally dependent on crewed vessels and manual inspections, toward more automated, sustainable, and integrated approaches that align with the environmental and social commitments of the offshore wind sector. Finally, emerging trends and persisting obstacles, notably energy autonomy, are discussed, outlining the requirements for consolidating a robust, connected, and sustainability-oriented model for offshore maintenance.

This work proposes the classification of motor imagery signals for brain-machine interfaces with functional electrical stimulation in the quadriceps muscle. Five volunteers participated in the test, 3 healthy participants, aged 28 ± 3 years, and 2 paraplegic volunteers, aged 43 (ASIA-B, C7 level - 16 years) and 47 (ASIA-A, T7 level - 20 years) years respectively. In total, each participant performed 90 repetitions of motor imaging of the lower limb under electrical stimulation, with frequencies of 20Hz, 35Hz, and 50Hz and current amplitude of 20mA. The patterns were analyzed off-line and submitted to the classification architectures after application of spatial filtering to extract the characteristics. The classification of the patterns was performed using the architectures: (i) Linear Discriminant Analysis (LDA), (ii) Multilayer Perceptron (MLP), and (iii) Support Vector Machine (SVM). To validate the proposal, the performance was compared between the classifiers through the accuracy of cross validation, variance, precision, and sensitivity. With the SVM classifier, the best accuracy percentage was 86.5%. These results are promising and the trained architectures are feasible for implementation in neuroprostheses with lower computational resources.