Prototipe Fire Suppression System Berbasis Machine Learning dengan Raspberry Pi Pico dan IR Flame Sensor untuk Mendeteksi Api Nyata

Kebakaran merupakan insiden berisiko tinggi di berbagai sektor operasional, dipengaruhi oleh faktor teknis, lingkungan dan perilaku manusia, sehingga membutuhkan sistem deteksi yang andal dan adaptif. Pendekatan berbasis Internet of Things dan machine learning mampu meningkatkan akurasi deteksi melalui pemrosesan data multi-sensor. Pada mesin operasional dengan durasi kerja panjang, penggunaan IR flame sensor sebagai alat deteksi dini menghadapi tantangan false positive akibat interferensi radiasi inframerah lingkungan. Kondisi ini menegaskan kebutuhan akan metode klasifikasi machine learning yang mampu membedakan nyala api nyata sekaligus efisien untuk platform mikrokontroler berdaya rendah.
Penelitian ini mengembangkan fire suppression system berbasis machine learning menggunakan raspberry pi pico dan IR flame sensor untuk mendeteksi api nyata. Metode yang digunakan yaitu integrasi rules-based system sebagai mekanisme keputusan awal dengan reinforcement learning untuk menyesuaikan bobot keputusan secara adaptif berdasarkan data pembelajaran (training). Data mentah berupa data ADC dikumpulkan dari berbagai skenario kondisi api dan non-api, termasuk sinar matahari yang di kondisikan sebagai non-api. Kemudian dianalisis secara temporal untuk membedakan karakteristik radiasi api, non-api. Hasil klasifikasi digunakan sebagai sinyal kontrol untuk mengaktifkan relay dan solenoid valve pada sistem pemadaman.
Penelitian ini di uji melalui akuisisi 20.400 data ADC dari 17 skenario api dan non-api serta perumusan aturan evidance dan pola temporal cahaya matahari, yang menunjukan kestabilan ADC dan efektifitas pengurangan false positive. Pada tahap deployment, akurasi non-api mencapai 95-100% dengan false positive rate 14,58%, sementara recall deteksi api sebesar 38,54%. Hasil penelitian menegaskan trade-off antara minimalisasi alarm palsu dan sensitivitas deteksi api, sehingga sistem belum layak untuk diaplikasian di lingkungan nyata. Namun, penelitian ini bernilai sebagai prototipe yang menunjukan secara empiris bahwa integrasi rule-based system dan reinforcement learning mampu meningkatkan keandalan keputusan dalam menekan alarm palsu.
Kata kunci: Fire Suppression System, machine learning, rule-based system, reinforcement learning, IR flame sensor, Raspberry Pi Pico, ADC (Analog to Digital Converter).

===================================================================================================

Fires are high-risk incidents in various operational sectors, influenced by technical, environmental, and human behavioral factors, requiring a reliable and adaptive detection system. An Internet of Things and machine learning based approach can improve detection accuracy through multi-sensor data processing. In operational machines with long operating hours, the use of IR flame sensors as an early detection tool faces the challenge of false positives due to interference from environmental infrared radiation. This situation emphasizes the need for a machine learning classification method capable of distinguishing real flames while being efficient for low-power microcontroller platforms.
The method used integrates a rules-based system as an initial decision mechanism with reinforcement learning to adaptively adjust decision weights based on training data. Raw data, in the form of ADC data, was collected from various fire and non-fire scenarios, including sunlight conditioned as non-fire. Then, it was analyzed temporally to differentiate the radiation characteristics of fire and non-fire. The classification results were used as control signals to activate relays and solenoid valves in the fire suppression system.
Testing was conducted by acquiring 20,400 ADC data sets from 17 fire and non-fire scenarios, as well as formulating evidence rules and temporal patterns of sunlight. This demonstrated the stability of the ADC and its effectiveness in reducing false positives. During the deployment phase, the non-fire accuracy reached 95-100% with a false positive rate of 14.58%, while the fire detection recall was 38.54%. The results confirmed the trade-off between minimizing false alarms and fire detection sensitivity, requiring further development for the system to be applied in real-world environments. However, this research serves as a valuable prototype, empirically demonstrating that the integration of a rule-based system and reinforcement learning can improve decision reliability in suppressing false alarms.
Keywords: Fire Suppression System, machine learning, rule-based system, reinforcement learning, IR flame sensor, Raspberry Pi Pico, ADC (Analog to Digital Converter).