“…Stroke and brain injury often leave sequelae such as hand and foot hemiplegia, greatly impairs their rehabilitation confidence, and induces negative, depression and other mental states. Early rehabilitation training is of great significance for the recovery of hand motor function [1], can also shorten the rehabilitation cycle [2]. However, in the traditional clinical practice of rehabilitation training, most patients rely on the guidance and help of rehabilitation doctors, the rehabilitation cycle is long, the rehabilitation evaluation depends on the judgment of doctors' experience, and there are few rehabilitation resources, expensive and difficult to see a doctor, resulting in many patients unable to complete rehabilitation training; At present, the rehabilitation devices on the market are relatively simple, which can drive the affected limb to do passive motion, lack of active motion prediction and participation, and cannot provide patients with personalized and more effective training process [3][4], (For example, ZHAGN et al [5] used BP neural network to realize real-time mapping of surface EMG signal and hip joint, knee joint and ankle joint angles; DAI [6] uses GRNN neural network, uses surface EMG signal to predict ankle joint angle, and uses golden section algorithm to determine the best smoothing parameters σ; CHEN et al used a limited Boltzmann mechanism to build a depth automatic encoder to extract features, and used BP neural network to estimate the hip and knee joint motion angles in lower limb motion [7]); Because there are many finger joints and complex movements, the finger diameter is relatively small, the angle sensor is large, difficult to install, and to collect the finger joint movement angle, there is less prediction of the hand finger joint movement angle (for example, literature [8] proposed to use optical motion capture camera to collect the hand movement angle, which has a high recognition rate, but the decoding time is long, and it is not suitable for real-time prediction).…”