This study proposed a ship fire identification approach that integrates an attention mechanism, a convolutional neural network (CNN) and a bidirectional long short term memory network (BiLSTM) to address the low accuracy of existing methods. A three-deck ferry fire simulation model was constructed using the fire dynamics simulator (FDS), and sensors were used to collect temperature, carbon monoxide, and visibility data from the simulated ship fire process. A CNN was employed to extract longitudinal features from fire data, while dimensionality reduction was used to compress data length and to reduce the number of model training parameters. A cascaded deep learning neural network based on BiLSTM was established to extract transverse features from fire data, where an attention mechanism was incorporated at the output layer. Furthermore, to accelerate convergence, an improved grey wolf optimization algorithm was developed by integrating the chaotic game algorithm. The improved algorithm was applied to optimize the CNN-BiLSTM-Attention model, which was subsequently utilized to perform ship fire identification experiments under two scenarios. The experimental results indicated that, despite the imbalance in ship fire data samples, the proposed approach outperformed other fire classification methods, achieving 100% fire identification accuracy and satisfying practical engineering requirements.