Klasifikasi Tingkat Roasting Biji Kopi Melalui Fusi Data Sensor Spektral dan Gas Berbasis Deep Learning

Rahman Maulana Jundika, Dr. Danang Lelono, S.Si., M.T.;Nur Achmad Sulistyo Putro, S.Si., M.Cs., Ph.D.

2026 | Tesis | MAGISTER ELEKTRONIKA DAN INSTRUMENTASI

Proses roasting merupakan tahapan paling krusial dalam menentukan karakteristik rasa dan mutu biji kopi, namun evaluasinya hingga kini masih banyak bergantung pada penilaian subjektif manusia yang rentan terhadap inkonsistensi. Metode analisis berbasis sensor tunggal, baik sensor spektral maupun electronic nose, memiliki keterbatasan yang bersifat komplementer sehingga tidak mampu menangkap karakteristik roasting secara menyeluruh. Penelitian ini merancang dan mengimplementasikan sistem klasifikasi tingkat roasting biji kopi secara otomatis berbasis fusi multimodal yang mengintegrasikan larik sensor gas MOS (TGS2600, MQ136, MQ4, MQ7, MQ3, MQ138) dengan sensor spektral AS7265x 18-kanal pada platform embedded system Raspberry Pi 5.

Data diperoleh melalui protokol akuisisi tiga fase (flushing, collecting, purging) pada closed chamber dengan sampel biji kopi Arabika dari empat single origin yang mewakili empat kelas: Green Bean (Toraja), Light Roast (Kerinci), Medium Roast (Bali Kintamani), dan Dark Roast (Temanggung). Fitur temporal dari fase collecting dan purging dengan total 480 timestep dan 24 kanal diproses menggunakan tiga arsitektur deep learning yang dibandingkan, yaitu 1D-CNN, LSTM, dan CNN+LSTM, dengan evaluasi stratified shuffle split 5-fold cross-validation.

Arsitektur 1D-CNN mencapai akurasi tertinggi sebesar 97,75 persen plus minus 1,05 persen, mengungguli LSTM (89,25 persen plus minus 4,01 persen) dan CNN+LSTM (90,00 persen plus minus 5,00 persen). Studi ablasi modalitas membuktikan bahwa fusi multimodal meningkatkan akurasi sekitar 10 persen dibandingkan penggunaan sensor gas saja (87,25 persen) maupun sensor spektral saja (88,00 persen), dengan visualisasi t-SNE yang mengonfirmasi sifat komplementer kedua modalitas secara empiris. Model 1D-CNN terbaik yang diimplementasikan pada Raspberry Pi 5 mencapai akurasi inferensi real-world sebesar 95,00 persen pada 20 sesi pengujian independen dengan latensi 4,19 milidetik, konsumsi CPU 33,5 persen, peak RAM 417,5 MB, dan energi per inferensi sebesar 3,84 mikrowatt-jam.

The roasting process is the most critical stage in determining the flavor characteristics and quality of coffee beans; however, its evaluation still largely relies on subjective human assessment, which is prone to inconsistency. Single-modality sensor approaches, whether spectral sensors or electronic nose systems, have complementary limitations that prevent them from capturing roasting characteristics comprehensively. This study designed and implemented an automated coffee roasting level classification system based on multimodal data fusion, integrating an array of MOS gas sensors (TGS2600, MQ136, MQ4, MQ7, MQ3, MQ138) with an AS7265x 18-channel spectral sensor on a Raspberry Pi 5 embedded system platform.

Data were acquired using a three-phase protocol (flushing, collecting, purging) inside a closed chamber with Arabica coffee bean samples from four single origins representing four classes: Green Bean (Toraja), Light Roast (Kerinci), Medium Roast (Bali Kintamani), and Dark Roast (Temanggung). Temporal features from the collecting and purging phases, comprising 480 timesteps across 24 channels, were processed using three deep learning architectures: 1D-CNN, LSTM, and CNN+LSTM, evaluated via stratified shuffle split 5-fold cross-validation.

The 1D-CNN architecture achieved the highest accuracy of 97.75 percent plus minus 1.05 percent, outperforming LSTM (89.25 percent plus minus 4.01 percent) and CNN+LSTM (90.00 percent plus minus 5.00 percent). A modality ablation study confirmed that multimodal fusion improved accuracy by approximately 10 percent compared to gas-only (87.25 percent) and spectral-only (88.00 percent) configurations, with t-SNE visualization empirically validating the complementary nature of both modalities. The best-performing 1D-CNN model deployed on Raspberry Pi 5 achieved a real-world inference accuracy of 95.00 percent across 20 independent measurement sessions, with an inference latency of 4.19 milliseconds, CPU utilization of 33.5 percent, peak RAM of 417.5 MB, and energy consumption of 3.84 microwatt-hours per inference.

Kata Kunci : klasifikasi roasting kopi, fusi multimodal, electronic nose, sensor spektral AS7265x, 1D-CNN, deep learning