Research And Implementation Of Multimodal Signal Acquisition And Recognition Systems For Lower Limbs

Human lower extremity behavior recognition is an important part of the lower extremity exoskeleton robot system.In order for the lower extremity exoskeleton to operate reliably,realtime and effectively,it is necessary to identify human behaviors,provide it to exoskeleton motion selection strategies,and increase exoskeleton and coordination of movements between wearers.In this paper,multiple motion gaits of the lower limb of the human body are identified and classified based on a variety of sensor signals.The gait of the human body walking up and down stairs and walking is divided into a swing phase and a support phase.For these six phases,an acquisition and recognition hardware system is designed.The surface electromyography(sEMG)and acceleration signal are classified and identified during human movement.The average recognition rate of six phases is 96.92%,and the recognition time of the motion algorithm is within 300ms.The main tasks in this paper include the following:(1)Design and implementation of the overall hardware of the gait acquisition and recognition system.According to the characteristics of the sEMG,an sEMG acquisition circuit is designed,and a bipolar power system for the sEMG acquisition is designed;the type of IMU is selected to collect the acceleration signal,and the IMU interface is designed;The plantarground contact force system made of flexible materials is used to collect plantar information to facilitate the division of active segments.The ADC for analog-to-digital conversion of sEMG is designed,and the AD7607 circuit is designed.Based on the complexity of data acquisition and data processing,the main control STM32F746ZGT6 is selected.On a PCB,the layered structure integrates the overall hardware system of the system.(2)Data acquisition and signal preprocessing.Different frequency signals are asynchronously collected,sEMG are collected through the AD7607,human motion acceleration signals are collected through the IMU,and plantar information is collected through the plantar-ground contact system.The three types of signals are synchronized to time window of 10ms using secondary cache and synchronous aggregation.Perform real-time filtering on the collected signals to filter out interference and retain main characteristic information.The sEMG was selected as the basis for dividing the active segment.The active segment was divided using the AMA method based on the sEMG.Moreover,a fast loop queue detection active segment method was designed based on the AMA method.(3)Implementation of classification recognition algorithm.Train the convolutional neural network(CNN)network for motion recognition classification based on the extracted active segment data.The CNN modle based on Keras is designed to classify and recognize the swing phase and the support phase of the staircase,downstairs,and flat walk for the sEMG and acceleration signal,and the average recognition rate reaches a high level.(4)Algorithm transplantation and hardware implementation.The trained CNN modle is transplanted to the main control via Cube.AI.After the transplantation,the algorithm usage resources are compressed,the algorithm recognition rate is slightly reduced,and the recognition algorithm is shorter in the main control,which can meet the requirements of real-time recognition.Then the entire system program task part was transplanted,including peripherals,timers,task priorities,etc.,to improve the overall part of the system.