Adaptive Denoising of RTD Fluxgate Sensor-Induced Signals Using Attention Residual Autoencoder and Generative Adversarial Network

The residence time difference (RTD) fluxgate sensor exhibits high sensitivity and enables precise detection of weak magnetic fields, and has been widely applied in geophysical exploration and environmental monitoring. However, in practical measurements, the sensor is susceptible to environmental and intrinsic noise, resulting in degraded accuracy. To address this issue, a multi-stage denoising algorithm integrating wavelet thresholding, attention-based residual learning, and multi-scale adversarial optimization is proposed. First, high-frequency disturbances are suppressed via multi-scale wavelet soft-thresholding. Then, an attention residual autoencoder is introduced, where residual connections preserve signal characteristics, and the attention mechanism captures both local details and global temporal features, improving denoising performance and fidelity. Finally, a generative adversarial network (GAN) with multi-scale constraints is employed to further attenuate residual noise and enhance signal consistency. Experimental results show that, compared with wavelet transform, Channel Attention Transformer, and EMD methods, the VRI is improved by 0.86, 3.01, and 3.44, respectively, while the time-difference fluctuation is reduced by 12.9%, 61.4%, and 62.5%. The proposed method improves signal accuracy and stability, providing effective algorithmic support for high-precision RTD fluxgate measurements.