The reliable and efficient operation of rotating machines depends on fault diagnosis through the analysis of vibration signals. The long short-term memory (LSTM) and its variants have demonstrated potential in analyzing time-series vibration data. However, the lack of interpretability in existing LSTM-based methods hinders their practical applicability. To address this limitation, we propose a parallel hidden layers’ bidirectional LSTM framework with hierarchical attention mechanisms for diagnosing the faults of rotating machines using time-series vibration data. The framework incorporates four improvements: (1) hierarchical attention mechanisms after each bidirectional LSTM layer to focus on informative features in vibration signals, (2) parallel bidirectional LSTM layers to capture complex temporal dependencies, (3) grid search with k-fold cross-validation method to determine optimal network parameters, and (4) an improved activation function. Experimental validation using motor bearing vibration datasets demonstrates the effectiveness and superiority of the framework compared with LSTM-based methods. The proposed model achieves accuracy rates of 99.78%, 99.22%, 99.31%, and 99.27%, for four loading conditions, and 98.66% for the combined load condition. Additionally, the use of t-distributed neighbor embedding visualization effectively represents the high-dimensional hierarchical attention-based bidirectional LSTM outputs in lower-dimensional spaces.