Activity recognition in rehabilitation training based on ensemble stochastic configuration networks

Rehabilitation training for patients with limb activity dysfunction and sub-healthy state has gradually shifted from therapies to strategies with remote assistance. Stochastic configuration networks (SCNs) are characterized by a structure that varies with task complexity, making them ideal for use as the lightweight AI activity recognition model in a remote rehabilitation training system. Given an imbalanced data classification and large-scale data analytics task, the original SCN classifiers may fail to provide satisfied performance. In this paper, we propose two solution that are Bagging SCNs and Boosting SCNs for HAR based on SCNs. Bagging SCNs use the bootstrap method to generate balanced subsets to reduce the influence caused by imbalance dataset. Then, multiple SCNs models are trained in parallel, followed by the identification of the best ensemble model through validation sets. Boosting SCNs employ forward stagewise additive modeling and utilize the SAMME algorithm to minimize the multi-class exponential loss for multi-class classification. This algorithm progressively enhances the base learner’s focus on previously misclassified instances from previous rounds, ultimately lowering the misclassification rate. The activity datasets of three groups of tests are collected by using a self-built experimental platform. Our experiments compare the performance of two Ensemble SCNs with original SCNs, Convolutional Neural Networks, Long Short-Term Memory, Gradient Boosting Decision Tree(GBDT) and Support Vector Classifier. Results in the performance of two Ensemble SCNs demonstrate that our proposed algorithm has good potential to be applied for HAR algorithm.