Combining deep learning algorithm and a-star algorithm to increase the accuracy of tracking lost gamma source

This paper proposes an automatic search algorithm, A-star CNN-RNN or ACR, which is a hybrid of two deep learning algorithms: recurrent neural network (RNN) and convolutional neural network (CNN) combined with the A-star algorithm to locate and route lost gamma sources. The radiation dose rate of 5 consecutive steps 50 cm apart in 81 meshes simulated by the Monte Carlo N-Particle (MCNP) code was used as the input for deep learning models to predict the location of the lost gamma source, and performance of ACR was evaluated based on the average number of steps and failure rates of prediction in a room with different obstacles. In a room without obstacles, the average tracking steps of the ACR algorithm were approximately twice as low as that of the gradient search algorithm (GS). In a room with a single wall as an obstacle, the search accuracy of the GS algorithm decreases as the length of wall increases (failure rate of 44 % with an 8-m-long wall), while it was improved significantly in the ACR algorithm with failure rates smaller than 5 %. In the rooms with two parallel walls and an L-shaped wall as obstacles, the ACR algorithm still maintained its prediction accuracy of over 94 % without needing to re-training the deep learning models, and demonstrated its adaptability which can be a useful search algorithm in detecting a lost gamma source.