Abstract: When fire breaks out in a ship cabin， the adjacent cabins may also suffer from chain effects and catch fire. In that case， if the fire alarm system does not provide accurate and reliable information concerning the afflicted cabins timely， the entire ship could be seriously damaged. In order to improve the early warning and alert correlation function of the ship fire alarm system， this paper analyzes the diagnostic features of typical cabin fires quantitatively and establishes a BP neural network evaluation model regarding the fire alarm priority. The model is then trained via the LM algorithm. In order to validate the proposed evaluation model， several testing samples have been employed. The results show that the method significantly improves the accuracy of the fire alarm system and lowers the chance of catastrophic losses due to cabin chain fires.