Human Lower Limb Movement Intention Recognition And Prediction Based On Time Series Deep Learning

Recognition of human lower limb movement intention refers to the use of computer technology to identify the intention of human lower limb movement.The application of technology for recognizing the intention of human lower limb movement is very extensive,and the most promising field is rehabilitation medicine.By identifying the patient’s movement intention,doctors can develop personalized rehabilitation plans for patients based on their condition and physical condition.In addition,technology for recognizing the intention of human lower limb movement can also be used for intelligent fitness,helping fitness enthusiasts better master exercise techniques and fitness effects.Therefore,technology for recognizing the intention of human lower limb movement is a very promising technology,and its widespread application will help promote the development of rehabilitation medicine and intelligent fitness and other fields.Currently,models constructed using machine learning and deep learning techniques have shown good performance in human lower limb movement intention recognition.In data processing,most methods extract statistical features from time series motion data and then learn the features using machine learning or deep learning neural networks to achieve recognition of human lower limb movement intention.However,the above methods still have some limitations: Firstly,there is a high recognition delay rate.The method based on statistical features can only extract the statistical information of the time window after the motion pattern has occurred.Secondly,the recognition effect of short-term data is poor.When the feature extraction time window length is short,the extracted features cannot effectively express the differences between each feature in the gait cycle of the motion pattern and other motion patterns.Finally,the internal correlation of time series data is ignored.Only extracting statistical features from the entire time window can eliminate the internal correlation of time series data.To overcome or at least significantly alleviate these limitations and achieve human lower limb movement intention recognition,this paper carries out the following work:(1)Construction of a time series motion database.Firstly,this paper defined the concept of motion patterns,and based on the categories of daily human lower limb activities,this paper obtained six steady-state motion patterns and ten transitional motion patterns.Then,this paper used the sliding window method to segment human lower limb motion data,preserving the internal correlation of the motion time series to reflect the continuity and variation rules of human lower limb motion patterns,and obtain more accurate motion pattern features.In order to overcome the limitation that biomechanical signal acquisition lags behind human lower limb movement,this paper offset the transitional motion pattern in time window segmentation to achieve the prediction and recognition of the intention of transitional motion patterns.(2)A wireless sensor data acquisition system for human lower limb movement was designed,which consists of six 9-axis sensors.The system can collect three-axis acceleration and three-axis angular velocity data of the wearer’s thigh,calf,and foot positions,and obtain three-axis Euler angle data through attitude calculation using quaternion method.The hardware module is connected to the data processing module through Bluetooth,and can process,store and visualize the collected data in real time,in order to perform subsequent motion intention recognition and analysis.The purpose of this system is to collect biomechanical signals of human lower limb movements with minimum impact on the wearer’s normal movements.(3)A human lower limb movement intention recognition model based on deep learning time series model was designed.The model uses an Encoder-Decoder structure,in which a feature selection network layer and LSTM are introduced in the encoder.The feature selection network layer is used to select important features in multi-feature variables for learning human lower limb movements,while LSTM can learn the relationship between time series data and retain long and short-term information,effectively improving the model’s generalization ability.In the decoder,this paper added self-attention mechanism to allocate weights to the temporal features learned by the encoder,improve the model’s attention to key information,and thus improve the recognition accuracy of motion patterns.