This study examined the effectiveness of online English learning among college students by integrating data from multiple learning behavior sources. The research revealed a bidirectional regulatory relationship between device environment and cognitive strategies, clarified the critical threshold for the regulation of metacognitive strategies, and improved language learning theory systems. A hybrid modeling technology was developed based on Dynamic Time Warping -- Long Short Term Memory (DTW-LSTM), a 42-dimensional feature matrix was constructed, and a federated learning framework designed that enhanced data processing and analysis capabilities. Regarding teaching practice, an online diagnostic system was developed based on the behavior pattern recognition algorithm. Additionally, an adaptive feedback mechanism to support personalized teaching was designed according to the prediction model.