Intrusion Target Classification Based on Seismic Signals Using Parallel Neural Network With Time/Frequency Dimension Reduction

Ground intrusion target classification based on seismic signals is an important task in unattended ground sensor (UGS) systems. In recent years, many deep learning methods using seismic signals to identify intrusion targets have been proposed. However, it is challenging for these algorithms to take into account the three indicators of recognition accuracy, computational speed, and anti-noise ability at the same time. In this article, a target classification method based on seismic signals [time/frequency domain dimension reduction-parallel neural network (TFDR-PNN)] is proposed to solve the above problem. TFDR-PNN first reduces the dimension of both the time and frequency domains of the signal by using an averaging pooling layer and spectrum interception. Subsequently, TFDR-PNN uses a parallel network to extract features from the dimension-reduced data and fuses these features for classification. The proposed method and the benchmark methods are tested on the dataset based on field acquisition. Experimental results indicate that the proposed TFDR-PNN can achieve comparable accuracy compared with the state-of-the-art methods, and its computational time is only 66.30% of the fastest benchmark algorithm. Meanwhile, the accuracy of the TFDR-PNN method is 10.15% higher than that of the optimal benchmark method under additional noise.