Objective: As the aircraft cargo compartment is a closed and complex environment, challenges occur in fire detection due to diverse cargo and limitations in sensor placement. Traditional fire detection methods are unable to accurately address fire recognition in such complex environments, particularly during the early stages of a fire. To address the challenge of accurately identifying the fire source localization in aircraft cargo compartments, this paper proposed a method based on a Bayesian-optimized bi-directional long short-term memory network (BO-BiLSTM). Method: This paper involved establishing an experimental platform of a real aircraft cargo compartment to simulate various fire scenarios. Fusion sensors were installed at multiple locations on the cargo compartment ceiling to collect multidimensional feature parameters in real time, including smoke volume fraction, CO volume fraction, and temperature. These data were used to build a fire feature database to capture the complex and dynamic changes in fire scenes. To improve the accuracy of detecting the fire source location, a bidirectional long short-term memory (BiLSTM) network was utilized, using its bidirectional information transmission mechanism to capture the forward and backward dependencies in time-series data. Meanwhile, the BiLSTM network structure was optimized using a Bayesian optimization algorithm to find the optimal combination of hyperparameters, enhancing the model's generalizability and robustness. Results: The experimental results indicate the following. First, the model was validated using a sliding window approach. When the number of windows was set to 8, the accuracy of the BO-BiLSTM model reached 97.2%. Compared with traditional models, such as recurrent neural networks (RNN), gated recurrent units (GRU), long short-term memory (LSTM) networks, and unoptimized BiLSTM models, the accuracy increased by 22%, 21%, and 2.6%, respectively. Second, in robustness tests with missing features, the BO-BiLSTM model maintained good stability. When only temperature and CO volume fraction were used as inputs, the model achieved an accuracy of 80.5%, which increased to 82.2% when using temperature and smoke volume fraction. Meanwhile, With when temperature and CO volume fraction were considered as inputs, the accuracy was 75.6%. The combination of temperature and smoke volume fraction performed the best, showing a strong correlation between these two features, with smoke volume fraction being more accurate for fire source localization. Finally, in the analysis of the model under sensor failure conditions, even with the number of functioning sensors reduced to four, the BO-BiLSTM model maintained an accuracy of 59.4%, significantly outperforming other models and demonstrating its advantages in complex and dynamic fire environments. The accuracy of the BiLSTM and LSTM models was lower than that of BO-BiLSTM, but their accuracy declined more gradually as the number of sensors increased, indicating some degree of resistance to interference. The GRU model performed better than the RNN model; however, when the number of damaged sensors was three or four, the accuracy of the GRU model was significantly lower than that of BO-BiLSTM, LSTM, and BiLSTM. The RNN model performed the worst in all scenarios, with its accuracy rapidly declining as the number of damaged sensors increased, dropping to approximately 45.2%. Conclusions: By significantly enhancing the accuracy and efficiency of fire source localization, this study provides essential technical support for the early detection, rapid response, and effective management of aircraft cargo compartment fires, which can help reduce fire risks and ensure safe and reliable air transportation.