Digital Circuit Correlation Optimization of Neural Networks for Apnea Detection during Transplantation

Sleep apnea syndrome (SAS) is a dangerous and high incidence sleep disorder. As more and more people are affected by SAS, monitoring SAS in family life becomes increasingly important. It is very meaningful to design an automatic SAS monitoring device. We designed a SAS detection model based on neural network, and transplanted the model to an application specific integrated circuit (ASIC). There are many problems in the process of transplanting neural network to ASIC. One of the most serious problems is the transplantation of nonlinear activation functions. We proposed a software-hardware joint optimization method to solve the problem of activation function in the SAS model transplantation. In the process of building the neural network model, our research modified the activation function in the traditional LSTM model and the attention mechanism, and adopted hard-sigmoid and LeakyReLU activation function for digital circuits. In the hardware transplantation, the activation function was constructed using the binary shift-based division, three-segment and two-segment function. A digital circuit without migration errors can be obtained with a small area and time consumption. We jointly optimized the structure of the model design and the digital circuit implementation. The model was more suitable for the digital circuit structure, so that the neural network can be transplanted more smoothly.