Abstract: The indication of b-value change and earthquake preparation has always been an important reference index for earthquake situation research and judgment. Based on the advantage that deep learning technology can mine the implicit characteristics of data, taking into account the natural phenomenon of frequent earthquakes in Sichuan and Yunnan in recent years and the high attention paid to the seismic activity of the Tanlu fault, this study uses the self-made data set of seismic events in Sichuan and Yunnan region from the earthquake catalogue of the China Earthquake Networks Center, and medium and strong earthquakes above M_\mathrmL 4.5 are labeled as 1, weak earthquakes below M_\mathrmL 3.0 are labeled as 0. The grid b-value in Sichuan-Yunnan region is calculated using the time sliding window method, and the b-value changes of each earthquake event in the five years before the earthquake are mapped to the labels. By using convolutional neural network models for training and classification, the optimized model is applied to the retrospective testing of medium and strong earthquakes in the central and southern section of the Tanlu fault zone. The verification accuracy can reach about 90%. Although the Sichuan-Yunnan region and the central and southern section of the Tanlu fault zone and their neighboring areas have different geographical and structural backgrounds, data-driven methods, reasonable generalization ideas, training datasets production, and deep learning model construction still have reference significance for mining strong earthquake laws.