Method of echo recognition of target in smoke environment based on residual convolutional neural network for pulsed laser detection

A recognition algorithm based on full waveform decomposition and residual convolutional neural network is proposed for the anti-interference problem of pulsed laser detection in smoky environments. As the original waveform data lack of features, a deep network -Resnet18 was used for the initial training. Considering the complexity of the algorithm, the structure, number of convolutional kernels, and sampling rate of Resnet18 were optimized, resulting in a lightweight version of Resnet6 with optimal parameters. Then, a pulsed laser detection platform under smoke environment was built. Considering the generalization of experimental data, five different scenes were designed to evaluate the algorithm. Compared to other algorithms, the proposed algorithm demonstrates superior performance in terms of low computational complexity (0.0441 MB parameters) and high recognition accuracy (94.59% +/- 0.46). Finally, additional experiments were conducted on the detection platform. Frosted material was used as the detection target with different intensity of noise signals to verify the robustness.