Evaluation of Interval Type-2 Fuzzy for IMU-Based Sitting Posture Classification on Wearable Devices

Abstract

Poor sitting posture is a significant contributor to musculoskeletal disorders in sedentary environments such as office work and remote learning. This study evaluates Interval Type-2 Fuzzy (IT2 Fuzzy) as a lightweight standalone classifier for IMU-based sitting posture classification on embedded wearable devices, comparing its performance against Support Vector Machine (SVM) under controlled experimental conditions. A dataset of 900 samples was collected using an MPU-6500 sensor with angle-simulation methods, representing three posture classes: upright, slightly slouched, and slouched. The dataset exhibits high inter-class separation (pitch angle differences approximately 10.8 degrees, 53 times the within-class standard deviations), enabling assessment of classifier behavior under well-defined conditions. Both IT2 Fuzzy and SVM achieved 100% classification accuracy on the test set, demonstrating feasibility for the controlled experimental scenario. Direct firmware deployment on an ESP32 microcontroller revealed comparable median inference times (IT2 Fuzzy: 29 microseconds, SVM: 26 microseconds), but IT2 Fuzzy exhibited 2.4 times lower timing variance (1.08 microseconds versus 2.61 microseconds) and superior timing predictability. IT2 Fuzzy requires only 6 parameters (48 bytes) compared to SVM's 141 parameters (192 bytes) and operates without training, normalization, or support vector storage. Additionally, IT2 Fuzzy provides native confidence scores and certainty indices without additional computation. These findings demonstrate that IT2 Fuzzy offers a viable alternative for embedded posture classification under controlled conditions, with advantages in model simplicity, timing consistency, and interpretability. Future work should validate performance on naturalistic multi-subject datasets with overlapping posture distributions and evaluate robustness to sensor drift and placement variability.