The growing demand for reliable fault diagnosis in wind turbines motivates the exploration of advanced deep learning models capable of capturing complex multivariate time series patterns of SCADA systems. This study develops deep learning classifiers for binary fault detection using time series data from operational wind turbines. The paper introduces and investigates two architectures tailored for this purpose: a modified TSMixer model for classification and a proposed hybrid CNN–LSTM–attention model. The hybrid model integrates convolutional layers for local feature extraction, LSTM networks for temporal dependency modeling, and an attention mechanism to emphasize critical time steps. After an extensive preprocessing pipeline, both models were trained and rigorously evaluated on a real SCADA dataset. Experimental results demonstrate that both architectures deliver robust and well-balanced performance. TSMixer achieved test accuracy of 81.2%, while the proposed hybrid model exhibited slightly superior performance with an accuracy of 82.0%. To enhance interpretability and trust, SHAP analysis was conducted. The interpretability study revealed the most influential features contributing to model decisions Both TSMixer and the hybrid CNN–LSTM–attention fusion model offer effective solutions for wind turbine fault classification. The choice between them presents a practical trade-off: TSMixer is faster and more computationally efficient, while the hybrid model provides more dependable predictions for high-stakes applications. This work offers a foundation for selecting suitable architectures and highlights the critical role of explainable AI in developing not only accurate but also trustworthy predictive maintenance systems.