The driving environment of heavy commercial cargo vehicles is complex and the working conditions are variable. Drivers are prone to fatigue and distraction during long-term operation, which poses risks. This paper proposes an intelligent cockpit driving state recognition and intervention mechanism based on multi-source perception fusion. By integrating multi-source perception data such as vehicle cameras, physiological sensors, and vehicle buses, a multi-source fusion recognition model based on CNN-LSTM is used to monitor the driver's attention, fatigue level, and dangerous driving behaviors in real time. The overall accuracy rate reaches 92.3%. On this basis, a dynamic adjustment mechanism for the intervention threshold at different levels is designed. According to the risk level (divided into four grades: 0-1.0), differentiated intervention measures such as audio-visual alerts, seat vibration, and active braking are taken. The false alarm rate is reduced to 5%. The research shows that multi-source data fusion technology can overcome the limitations of a single sensor, improve the recognition accuracy and robustness, provide technical support for the intelligent cockpit of heavy commercial cargo vehicles, and is of great significance for reducing the accident rate and ensuring transportation safety.