With the global aging population and an increase in respiratory system diseases, the usage scenarios of non-invasive ventilators are gradually shifting from hospitals to everyday homes. Therefore, there is an urgent need for ventilators to possess stronger autonomous adaptive capabilities to provide personalized treatment for different conditions. However, the current level of intelligence in ventilator products is relatively low, mainly constrained by the ventilator's ability to recognize patient respiratory states and its adaptive capacity to adjust ventilator parameters accordingly. In response to this situation, this paper combines technologies such as sensor fusion and deep learning to design and implement a sensor fusion-based adaptive algorithm for non-invasive ventilators. This algorithm consists of two parts: a deep learning-based parameter initialization algorithm and a deep learning-based stepwise titration algorithm. The parameter initialization algorithm initializes the ventilation mode and parameters of the ventilator based on the patient's historical respiratory data. Building upon the parameter initialization, the stepwise titration algorithm monitors real-time changes in patient state parameters through various sensors and adjusts the ventilator based on these state parameters until the entire treatment process is completed. Finally, we conduct simulation experiments on the proposed adaptive algorithm on a simulation platform, simulating different respiratory symptoms and leakage scenarios. The experimental results demonstrate that the adaptive algorithm outperforms existing works of the same type in terms of classification accuracy, precision, recall, and $F_1$ score, which suggests the potential to accelerate the intelligence of ventilators and provide the possibility of personalized treatment for patients.