Abstract
In this article, an Autonomous Fuzzy Echo State Network (AFESN) algorithm is being proposed. This algorithm combines Autonomous Learning Multi-Model Systems and Echo State Network (ESN) to perform Human Activity Recognition on the Heterogeneity Human Activity Recognition HHAR dataset. We incorporate the structural information of the dataset into the ESN by using the Takagi-Sugeno (TS) models/rules concept. The process of obtaining the structural information in the dataset was fully automated by the unimodal density based membership function, which then acts as the antecedent part. Subsequently, a consequent part composed by ESN was used to replace the affine models commonly used in TS models. The final model predictions are inferred by the fuzzy-weighed mechanism in a new fuzzy rule composed of multiple ESNs. Since our algorithm is fully automatic and does not require any prior knowledge, its use in online learning scenarios is adequately applicable. Lastly, the effectiveness of the algorithm is tested using the HHAR dataset and an accuracy of 42.9% was obtained. Compared with the performance of recent Autonomous Learning Multi-Model Systems (19.6%), our algorithm has gained a significant improvement.