Research on Laser Warning Angle Prediction Based on Deep Neural Networks

To address the challenges of high computational resource consumption and low accuracy in determining the incident laser angles by laser warning device systems, this study proposes a prediction method for the azimuth and elevation angles of incoming lasers based on deep neural network models. By constructing a feature dataset of zero-order and first-order diffraction spot center coordinates, a fully connected neural network model was established. The Mini-Batch gradient descent method (batch_size =32, learning rate <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\eta=0.01$</tex>) was employed for 1000 iterations of training. Through comparative analysis of four activation functions (ReLU, Sigmoid, Tanh, and Softmax), numerical examples demonstrated that the ReLU function achieved the highest prediction accuracy. Further optimizer comparisons revealed that the Adam algorithm achieved optimal performance on 50 independent validation sets (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{RMSE}=1.41^{\circ}, \text{MAE}=0.94^{\circ}$</tex>), outperforming SGD and RMSprop by 88.9 % and 99.13 %, respectively. The maximum absolute error measured in practical tests was 1.05°, meeting the angular resolution requirements of laser warning equipment. This research identifies the optimal parameter combination through multi-algorithm cross-validation, providing a novel solution for rapid computation of incoming laser angles.