Gait perception is an important basis for rehabilitation training and effect evaluation of hemiplegic patients.However

existing methods rely on multiple sensors

leading to cumbersome wearing and limited recognition accuracy.Based on the signals of ankle inertial measurement unit(IMU)

gait data of 10 hemiplegic patients and 5 healthy controls are collected under clinical conditions

and a hemiplegic gait dataset is constructed.In the gait classification experiment

a support vector machine(SVM)realizes distingetion of healthy gait unaffected side gait

and affected side gait

with an average accuracy of 97.90%.For gait phase recognition

a model combining one-dimensional convolutional neural network(1D-CNN)

long short-term memory(LSTM)network

and attention mechanism is proposed

which achieves accuracy of 96.56% and F1—score of 95.63%.Its accuracy is 4.44% and 2.16% higher than that of the LSTM and 1D-CNN-LSTM models respectively.Experimental results demonstrate that the method can achieve high-accuracy recognition of gait classification and phase while ensuring simple wearing

which verifies the effectiveness of the fusion neural network method.