Lost circulation is a common complex condition in drilling operations that severely compromises safety and efficiency. Most existing early-warning models focus mainly on predicting when lost circulation occurs, with insufficient attention to its severity and risk assessment. To address this gap, an intelligent lost circulation diagnostic model integrating logging time series data and multi-scale decomposition is proposed, combining Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), Intrinsic Mode Function (IMF) screening criteria, and Support Vector Machine (SVM) optimized by Particle Swarm Optimization (PSO). First, from drilling and completion reports of Iran’s Azadegan Oilfield, logging parameters that effectively characterize different leakage degrees are selected. Second, the logging data undergo multi-scale decomposition via ICEEMDAN; IMF components strongly correlated with leakage characteristics are filtered using Pearson correlation coefficients and energy entropy, and the data are reconstructed to eliminate medium-to-high-frequency noise. Finally, a diagnostic model is built with the PSO-optimized SVM—where the PSO optimizes the penalty coefficient C and kernel parameter γ—to enhance accuracy and generalization. Results indicate the ICEEMDAN-PSO-SVM model’s test set accuracy exceeds 98%, outperforming PSO-SVM, Random Forest (RF) and SVM across key performance metrics. This study offers valuable support for mitigating drilling safety risks and enhancing engineering efficiency and economic benefits.