Abstract: Taking Jinfoshan Lightning Special Test Field in Chongqing as the target area, 337 thunderstorm and non-thunderstorm samples from 2020 to 2022 were collected and analyzed. Based on the analysis of the correlation among sample variables, a fixed-point lightning warning model integrating multi-source observation data was constructed using principal component analysis and binary Logistic regression. Under the condition of imbalanced sample classification, the F1 score was introduced as a measure of model performance, and the classification effect of the early warning model under different boundary values and different lead times was evaluated. The results show that the model exhibits a good fit in terms of classification outcomes and various validation metrics. When the boundary value was set to 0.3, the classification effect of the model was the best. Taking the 18 min advance time as an example, the effective alarm rate, miss rate and false alarm rate of the model were 95.83%, 22.03% and 4.17% respectively. The established fixed-point lightning early warning model can provide references for lightning warning in sensitive places such as flammable and explosive places, scenic spots, mines, industrial parks and so on.