Abstract: Aiming at the problem of rock physical facies identification, a identification method based on the interpretable one-dimensional convolutional neural network is proposed. By introducing the global average pooling layer, the dynamic variation part of the logging waveform is highlighted. The interpretability of the method is enhanced by the classification activation mapping. By introducing dilated convolution and batch normalization, the performance degradation caused by the global average pooling layer is compensated. According to the experimental results, the average F1 score of the four petrophysical facies in the test set is 0. 97, which is about 0. 15 higher than that of other similar methods. The results show that the proposed method can be used to identify petrophysical facies and improve the interpretability in the classification process, providing a new deep learning method for the prediction of high quality tight sandstone reservoirs.

Key words: petrophysical facies, interpretable one-dimensional convolutional neural networks, global average pooling layer, dilated convolution, batch normalization