In manufacturing environments, maintenance logs are often manually recorded and inconsistently documented, resulting in inefficient fault resolution and limited access to historical knowledge. This paper proposes an intelligent recommendation system that predicts appropriate machine maintenance actions based on textual issue descriptions. The system leverages BERT-based embeddings for semantic representation of input text and employs a Feedforward Neural Network (FNN) for multi-class classification across 303 resolution categories. To address severe class imbalance in the dataset, a hybrid training strategy combining focal loss and random oversampling was applied. Experimental results demonstrate significant improvements in macro F1-score and recall using the hybrid approach. The model is integrated into a PyQt5 desktop application with real-time prediction, email-based verification via Flask and Mailgun, and a feedback mechanism to support continuous learning. This human-in-the-loop system enhances troubleshooting consistency, supports adaptive learning, and offers a scalable solution for intelligent maintenance decision-making in industrial settings.