Identification of lithology from drilling cuttings is a key step for reservoir characterization. At present, the traditional method is to collect and analyze the cuttings by manual interpretation, which is subjective and time-consuming. In order to improve the accuracy, timeliness, and automation of identification of cuttings lithology, this paper completed lithology identification and classification through batch iterative training based on Resnet-34 network.
Automated rock type identification from cuttings images captured by a given camera is the goal of this work. The main challenge in cuttings recognition is the similarity of color and grain size in two or different cuttings. Another important factor is that the light source of the camera. In order to capture the images of cuttings more comprehensively for identification, a single light source with 8 channels (ultraviolet, blue, green, yellow, red, infrared, far infrared, and white light) was used to irradiate the cuttings during the imaging process, and the imaging information of cuttings in the full spectral segment can be obtained as the input of the deep learning network. In order to solve the overfitting problem caused by data duplication under different light sources, the input images are first preprocessed by brightness equalization and random cropping, so that each grayscale image can have the same brightness and do not have too much repetition under different light sources. In terms of the efficiency and accuracy, Resnet-34 is selected as the model framework of cuttings recognition and classification. The overall accuracy of the model is increased to 98.6% combined with the actual hardware support in the field, which promotes the more efficient identification of logging cuttings and provides a theoretical basis for improving the online logging of logging in the field.