A Multilayer Interval Type-2 Fuzzy Extreme Learning Machine for the recognition of walking activities and gait events using wearable sensors

Abstract

In this paper, a novel Multilayer Interval Type-2 Fuzzy Extreme Learning Machine (ML-IT2-FELM) for the recognition of walking activities and Gait events is presented. The ML-IT2-FELM uses a hierarchical learning scheme that consists of multiple layers of IT2 Fuzzy Autoencoders (FAEs), followed by a final classification layer based on an IT2-FELM architecture. The core building block in the ML-IT2-FELM is the IT2-FELM, which is a generalised model of the Interval Type-2 Radial Basis Function Neural Network (IT2-RBFNN) and that is functionally equivalent to a class of simplified IT2 Fuzzy Logic Systems (FLSs). Each FAE in the ML-IT2-FELM employs an output layer with a direct-defuzzification process based on the Nie-Tan algorithm, while the IT2-FELM classifier includes a Karnik-Mendel type-reduction method (KM). Real data was collected using three inertial measurements units attached to the thigh, shank and foot of twelve healthy participants. The validation of the ML-IT2-FELM method is performed with two different experiments. The first experiment involves the recognition of three different walking activities: Level-Ground Walking (LGW), Ramp Ascent (RA) and Ramp Descent (RD). The second experiment consists of the recognition of stance and swing phases during the gait cycle. In addition, to compare the efficiency of the ML-IT2-FELM with other ML fuzzy methodologies, a kernel-based ML-IT2-FELM that is inspired by kernel learning and called KML-IT2-FELM is also implemented. The results from the recognition of walking activities and gait events achieved an average accuracy of 99.98% and 99.84% with a decision time of 290.4ms and 105ms, respectively, by the ML-IT2-FELM, while the KML-IT2-FELM achieved an average accuracy of 99.98% and 99.93% with a decision time of 191.9ms and 94ms. The experiments demonstrate that the ML-IT2-FELM is not only an effective Fuzzy Logic-based approach in the presence of sensor noise, but also a fast extreme learning machine for the recognition of different walking activities.

Deep Learning in Human Activity Recognition with Wearable Sensors: A Review on Advances

“…Autoencoders have been widely adopted as an unsupervised method for learning features. As such, the outputs of autoencoders are often used as inputs to other networks and algorithms to improve performance [120][121][122][123][124]. An autoencoder is generally composed of an encoder module and a decoder module.…”

“…For handling missing input data, a compelling strategy is to train an autoencoder with artificially corrupted input x, which acts as an implicit regularization. Usually, the considered corruption includes isotropic Gaussian As such, autoencoders are most commonly used for feature extraction and dimensionality reduction [120,[122][123][124][125][126][133][134][135][136][137][138][139][140][141]. Autoencoders are generally used individually or in a stacked architecture with multiple autoencoders.…”

Deep Learning in Human Activity Recognition with Wearable Sensors: A Review on Advances

“…Autoencoders have been widely adopted as an unsupervised method for learning features. As such, the outputs of autoencoders are often used as inputs to other networks and algorithms to improve performance [70,94,130,163,201]. An autoencoder is generally composed of an encoder module and a decoder module.…”

“…As such, autoencoders are most commonly used for feature extraction and dimensionality reduction [13,20,38,49,50,56,70,94,120,121,139,163,201,205,210,212]. Autoencoders are generally used as is, or in a stacked architecture with multiple autoencoders.…”

Locomotion Mode Recognition Algorithm Based on Gaussian Mixture Model Using IMU Sensors

“…Recently, fusion sensor type researches using various sensors such as inertial measurement unit (IMU) sensors have been actively conducted in relation to ADL, researches are being conducted to determine the terrains [ 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 ], which are the walking environments or to determine the gait phase related to human intention [ 13 , 14 , 15 ]. In this study, a pilot study on hip exoskeleton robots as well as locomotion mode recognition (LMR) algorithm for five terrains is considered, and aims to help walking in ADL.…”