By installing sensors on the lift door system, metrics such as door panels’ 3-dimensional vibration accelerations, door motor current, and distance of door panels are continuously monitored in real time. By simulating various typical faults in lift door system, the monitored data is collected under continuous operational conditions and pre-processed into a multidimensional time series dataset with feature values and fault labels. By adjusting and optimizing the input and hidden layers of AI models, 4 different AI models are designed. All 4 models are trained and validated. The AI model yielded ideal training results when the input layer included typical feature values of both vibration acceleration and current, and the hidden layer combined LSTM and fully connected layers, which validation accuracy reaches 97%. The selected optimal model, which has relatively optimal fault recognition performance, was quantized, converting its weight parameters from 32-bit floating-point (float32) to 8-bit integers (int8). This resulted in an edge AI model capable of running on edge devices. The generated edge AI model was recompiled for × 86–64 and Arm64 hardware architectures and deployed on the respective hardware devices to test the fault recognition inference speed and calculate the recognition accuracy, which demonstrated minimal accuracy loss and high-speed inference performance. The results of this research can be applied to the development of on-site intelligent inspection devices for lift door systems.