Sistem Monitoring Performa Neuro-Fisiologis Terintegrasi IoT untuk Analisis Beban Latihan Fisik

Abstract

IHSAN HADI GUNAWAN. Sistem Monitoring Performa Neuro-Fisiologis Terintegrasi IoT untuk Analisis Beban Latihan Fisik. Dibimbing oleh BAYU WIDODO. Penelitian ini membahas sistem monitoring IoT untuk analisis beban latihan melalui respons neuromuskular dan kardiovaskular. ESP32 menerima dua kanal sEMG serta HR Polar H10 melalui BLE, kemudian mengirimkan data melalui Wi-Fi dan broker MQTT. Analisis menggunakan tiga rekaman sEMG dan tiga ekspor Polar Beat yang diolah terpisah. Segmentasi menghasilkan 1959 jendela berdurasi 2,56 detik dengan langkah 1,28 detik. Empat fitur RMS dan MAV kedua kanal digunakan pada dua model Random Forest dengan pembagian kronologis: 1368 jendela latih, 293 validasi, dan 274 uji. Model aktivitas mengenali label sesi Jalan, Jalan–lari, dan Sepeda statis dengan akurasi 90,15% dan Macro-F1 90,72%. Model beban menggunakan rata-rata empat persentil amplitudo dengan kalibrasi per sesi dari data latih, kemudian membaginya menurut tertil menjadi label sementara Light, Normal, dan Fatigue. Pada data uji, model memperoleh agreement 93,07% dan Macro-F1 88,43%; Light mendominasi jalan dan Fatigue mendominasi sepeda statis. Fatigue menandai kategori skor amplitudo relatif tertinggi dan belum memvalidasi kelelahan fisiologis. HR dianalisis tersendiri karena pasangan waktu dengan sEMG belum terkonfirmasi. Kinerja komunikasi streaming dan generalisasi lintas sesi belum terverifikasi. Kata kunci: IoT, MQTT, Polar H10, Random Forest, sEMG bilateral

IHSAN HADI GUNAWAN. IoT-Integrated Neuro-Physiological Performance Monitoring System for Exercise Load Analysis. Supervised by BAYU WIDODO. This study examines an IoT monitoring system for exercise load analysis using neuromuscular and cardiovascular responses. The ESP32 receives two sEMG channels and Polar H10 heart rate via BLE, then transmits data over Wi-Fi through an MQTT broker. Three sEMG recordings and three Polar Beat exports were analyzed separately. Segmentation produced 1959 windows lasting 2.56 seconds with a 1.28-second step. Four bilateral RMS and MAV features supported two Random Forest models with chronological partitions of 1368 training, 293 validation, and 274 test windows. The activity model classified Walking, Walking–running, and Static cycling session labels with 90.15% accuracy and 90.72% Macro-F1. The load model averaged four amplitude percentiles calibrated per known session using training data, then applied training-score tertiles to define provisional Light, Normal, and Fatigue labels. Test agreement was 93.07%, with 88.43% Macro-F1; Light dominated walking and Fatigue dominated static cycling predictions. Fatigue denotes the highest relative amplitude-score category and does not establish physiological fatigue. Heart rate was analyzed separately because temporal pairing with sEMG was unconfirmed. Streaming communication performance and cross-session generalization were not established. Keywords: bilateral sEMG, IoT, MQTT, Polar H10, Random Forest