Denoising method for TEM data with weak IP signals using an attention-based residual gated recurrent unit denoising autoencoder

The time-domain electromagnetic (TEM) method is widely used in resource exploration. Recently, measured data often show symbol inversion, generally attributed to the induced polarization (IP) effect. For some deposits, the IP effect is weak and easily masked by random spike, instrument, and other noises, reducing interpretation accuracy and hindering reliable target identification. Effective denoising of TEM signals under weak IP conditions is therefore crucial. Traditional filtering methods can improve signal-to-noise ratio (SNR) but suffer from parameter selection issues. Deep learning denoising has shown promise, yet few approaches specifically address weak IP interference. We present an attention-augmented residual gated recurrent unit denoising autoencoder (ARGRU-DAE). GRU layers are used to capture the temporal decay of TEM responses, skip connections are added to improve gradient flow during training, and the decoder applies an attention mechanism to highlight weak polarization features in late-time data. In numerical experiments, ARGRU-DAE reduced noise while retaining polarization responses, with mean errors under 3% and higher SNR than baseline methods. The model improved the consistency of TEM inversion results under weak-IP conditions.