<p>The pitch system is one of the most failure-prone subsystems in offshore wind turbines due to frequent actuation under variable wind loads. Conventional SCADA alarm systems detect faults only after significant degradation, limiting opportunities for proactive maintenance. This study proposes an integrated early fault warning (EFW) framework addressing three limitations of existing methods: (1) linear correlation-based feature selection fails to capture non-linear relationships between environmental drivers and system degradation; (2) deep learning models suffer from suboptimal hyperparameter configurations; and (3) fixed alarm thresholds perform poorly under variable operating conditions. The proposed framework employs the Relief algorithm to identify health-sensitive features that reflect physical loading mechanisms, constructs a self-attention enhanced bidirectional LSTM (SABiLSTM) network with hyperparameters optimized via Sparrow Search Algorithm (SSA), and establishes adaptive warning thresholds using kernel density estimation (KDE). Validation using SCADA data from six 5.2 MW offshore wind turbines in the Taiwan Strait demonstrates that the proposed model achieves superior prediction accuracy. In two documented pitch motor encoder faults, the framework issued warnings 109–132 hours (4.5–5.5 days) before SCADA fault records, significantly exceeding the 24–72 hours reported in recent literature. Analysis of fault progression patterns revealed that early-stage anomalies manifested predominantly during high-pitch operation periods, with prediction errors amplified 30–50 times above baseline levels. The framework successfully tracked complete degradation trajectories from incipient warnings to catastrophic failures, providing preliminary evidence of the framework’s potential for condition-based maintenance scheduling in offshore wind operations, subject to further validation on larger and more diverse fault datasets.</p>