Abstract: With the development of society, high-quality magnetotelluric signal acquisition is becoming more and more difficult because of various types of human interference are increasingly intensified. Scholars have proposed many corresponding denoising methods for different types of noise to improve data quality. It is impossible to manually interpret each data before denoising due to the huge amount of data. So an efficient noise recognition and classification method is urgently needed. Based on this, artificial neural network is applied in the classification of magnetotelluric time series in this paper. Four types of time series, namely, simulated square wave, power frequency, impulse noise, and measured noiseless data, were used to conduct noise classification training and measured data classification on LSTM、FCN、ResNet、LSTM-FCN and LSTM-ResNet models. The results show that FCN and LSTM-FCN has a relatively good effect on the classification of magnetotelluric time series. Among them, the highest classification accuracy of FCN measured data can reach 99.84%, and the average time for each epoch is 9.6 s. LSTM-FCN has higher classification accuracy than FCN, the highest classification accuracy of measured data sets is nearly 100%, but the average time for each epoch is 24.6 s, and it is easier to overfit than FCN. Overall, LSTM-FCN can achieve higher classification accuracy when the amount of data is relatively small , if the amount of data is large, it is necessary to consider the time cost, using FCN is more appropriate. Finally, the magnetotelluric data containing different types of noise was successfully processed using the magnetotelluric noise processing system which constructed by the LSTM-FCN classification model and the LSTM denoising model.