Pengembangan Sistem Proteksi Aset Medis dengan Fitur Deteksi Anomali Suhu Berbasis Autoencoder dan Sumber Kebocoran Air Berbasis Decision Tree

Abstract

ZAHDAN FAQIH RAHBANI. Pengembangan Sistem Proteksi Aset Medis dengan Fitur Deteksi Anomali Suhu Berbasis Autoencoder dan Sumber Kebocoran Air Berbasis Decision Tree. Dibimbing oleh FALDIENA MARCELITA. Sistem peringatan dini pada gudang PT Synergy Dua Kawan Sejati sebelumnya hanya mengandalkan ambang batas suhu rata-rata statis dan deteksi genangan air secara biner, sehingga rentan menghasilkan alarm palsu akibat fluktuasi suhu ambient harian dan gagal mendeteksi anomali termal lokal berskala kecil seperti hotspot korsleting. Penelitian ini bertujuan mengembangkan sistem proteksi aset medis melalui integrasi kecerdasan buatan pada arsitektur Internet of Things berbasis ESP32-S3, mencakup deteksi anomali suhu menggunakan Convolutional Autoencoder dan klasifikasi sumber kebocoran air menggunakan Decision Tree. Pengujian dilaksanakan pada lingkungan simulasi Tingkat Kesiapterapan Teknologi Level 6 menggunakan tiga tumpukan kardus berdimensi 35×26,5×96 cm yang merepresentasikan produk medis baru dalam kemasan aslinya, sesuai kondisi penyimpanan aktual gudang. Data termal dari sensor MLX90640 berupa matriks 32×24 piksel dikumpulkan selama 4 hari 6 jam menghasilkan 24.179 frame, diproses menggunakan Zero-Mean Normalization, dan digunakan untuk melatih model Autoencoder berkapasitas ringan (3.217 parameter) selama 50 epoch dengan ambang batas ganda berbasis kaidah tiga-sigma. Klasifikasi sumber genangan air dilakukan dengan mengintegrasikan pembacaan sensor K-0135 dan data cuaca real-time dari OpenWeatherMap API untuk membedakan kebocoran atap saat hujan dan kebocoran pipa internal saat cuaca cerah. Hasil pengujian menunjukkan model Autoencoder berhasil mereduksi alarm palsu sebesar 65% dan meningkatkan sensitivitas deteksi anomali lokal sebesar 100% dibandingkan sistem berbasis ambang batas statis, sementara logika Decision Tree berhasil mengklasifikasikan sumber kebocoran air dengan akurasi 100% dari 10 skenario pengujian. Seluruh komponen telah terintegrasi ke dalam dashboard web dengan visualisasi heatmap termalreal-time dan notifikasi Telegram Bot disertai foto bukti, sehingga sistem ini terbukti layak diterapkan untuk mendukung manajemen risiko aset medis di lingkungan gudang.

ZAHDAN FAQIH RAHBANI. Development of a Medical Asset Protection System with an Autoencoder-Based Temperature Anomaly Detection Feature and Decision Tree-Based Water Leakage Source Classification. Supervised by FALDIENA MARCELITA. The early warning system previously deployed at PT Synergy Dua Kawan Sejati's warehouse relied solely on a static average temperature threshold and binary water detection, making it prone to false alarms from daily ambient temperature fluctuations and unable to detect small-scale local thermal anomalies such as shortcircuit hotspots. This research aims to develop a medical asset protection system through the integration of artificial intelligence into an ESP32-S3-based Internet of Things architecture, encompassing temperature anomaly detection using a Convolutional Autoencoder and water leakage source classification using a Decision Tree. Testing was conducted in a Technology Readiness Level 6 simulation environment using three stacked cardboard boxes with total dimensions of 35×26.5×96 cm, representing newly packaged medical products in their original packaging, consistent with the actual warehouse storage condition. Thermal data from the MLX90640 sensor in the form of a 32×24 pixel matrix were collected over 4 days and 6 hours, yielding 24,179 frames, which were processed using Zero-Mean Normalization and used to train a lightweight Autoencoder model (3,217 parameters) for 50 epochs with a dual threshold based on the three-sigma rule. Water leakage source classification was performed by integrating readings from the K-0135 sensor with real-time weather data from the OpenWeatherMap API to distinguish between roof leaks during rain and internal pipe leaks under clear weather. Testing results showed that the Autoencoder model successfully reduced false alarms by 65% and increased local anomaly detection sensitivity by 100% compared to the previous static threshold system, while the Decision Tree logic achieved 100% accuracy in classifying water leakage sources across 10 test scenarios. All components were successfully integrated into a web dashboard featuring real-time thermal heatmap visualization and Telegram Bot notifications with photographic evidence, demonstrating that the system is feasible for supporting medical asset risk management in warehouse environments.