To address the challenge in detecting low-frequency signals in ZPW-2000 frequency shift signals under unbalanced traction current interference,lightweight multi-scale neural network approach based on a convolutional attention mechanism is proposed for low-frequency signal detection.First,multi-scale layers with different convolutional kernel sizes are employed to extract frequency-shift signal features across various carrier frequency modulations,based on the carrier frequency range of ZPW-2000 signals.Second,an inverse residual linear bottleneck is introduced to optimize the network's efficiency,reducing parameters and shortening detection time while maintaining high detection accuracy.Finally,the Convolutional Block Attention Module is introduced to calibrate channel and spatial feature weights,thereby enhancing network performance.Classification is performed through a fully connected layer,producing the probability distribution of 18 types of low-frequency signals.Experimental results indicate that when frequency shift signals containing five types of noise,including power-frequency harmonic interference,are input into the low-frequency detection network,the average accuracy reaches 99.22％,the recall rate 99.21％,and the F1 score 0.992,with detection time not exceeding 0.249 seconds.Compared to traditional detection methods,the proposed method excels in detection performance and robust anti-interference capabilities,providing a valuable reference for detecting low-frequency information of ZPW-2000 frequency-shift signals under in-band noise interference conditions.
关键词
轻量化卷积神经网络/谐波干扰/多尺度神经网络/信号检测/ZPW-2000移频信号
Key words
Lightweight convolutional neural network/Harmonic interference/Multi-scale parallel computing Neural network/Signal detection/ZPW-2000 frequency shift signal
维普
万方数据