Workload is found to be a critical factor driving human behavior in human-machine interactions in modern complex high-risk domains. This paper presents a dynamic workload adjustment feedback loop with a dynamic cognitive load (CL) adaptation model to control workload adjustment during human-machine interaction. In this model, physiological signals such as Galvanic Skin Response (GSR) are employed to obtain passive human sensing data. By analyzing the obtained sensing data in real-time, the task difficulty levels are adaptively adjusted to better fit the user during working time. The experimental results showed that SVM outperformed other methods in offline CL classifications, while Naïve Bayes outperformed other methods in online CL level classifications. The CL adaptation model 1 (average performance is 87.5 %) outperformed the adaptation model 2 during the dynamic workload adjustment.