Evaluasi Perbandingan Arsitektur MLP dan 1DCNN Pada Sistem e-nose

Muhammad Fauzan, Galih Setyawan, S.Si., M.Sc., Ph.D.

2026 | Tugas Akhir | D4 Teknologi Rekayasa Instrumentasi dan Kontrol

Electronic nose (e-nose) banyak dikembangkan untuk memantau kualitas produk pertanian, termasuk kematangan pisang. Namun, sebagian besar penelitian sebelumnya hanya melaporkan akurasi tanpa mempertimbangkan ukuran model, memori (RAM), dan latensi inferensi pada mikrokontroler. Penelitian ini membandingkan arsitektur Multilayer Perceptron (MLP) dan 1-Dimensional Convolutional Neural Network (1DCNN) untuk mengklasifikasikan tiga kelas aroma, yaitu angin (kontrol), pisang muda, dan pisang busuk, menggunakan larik empat sensor gas MEMS (CO, VOC, CH, dan HCHO) berbasis ESP32-S3. Dataset terdiri atas 46 capture. Setiap capture diringkas menjadi 16 fitur, yaitu respons maksimum, steady state, slope naik, dan slope turun setiap sensor pada skala logaritma natural yang dinormalisasi dengan z-score. Model diuji dengan menahan satu capture secara bergiliran sebanyak 46 putaran dengan tiga seed. Pengujian meliputi variasi dropout rate, jumlah neuron dan lapisan tersembunyi, normalisasi, jumlah epoch, serta kuantisasi INT8 TensorFlow Lite. Hasilnya, MLP 40–20–10 dengan dropout 0,25 (MLP-P1) mencapai akurasi 97,10% ± 1,02, lebih tinggi daripada 1DCNN (94,93% ± 1,02), sedangkan model tanpa normalisasi hanya mencapai 71,74%. Kuantisasi INT8 memperkecil ukuran model 22,5–41,7?n estimasi memori aktivasi 74,3–75,0% tanpa menurunkan fidelitas model yang dinormalisasi. Model MLP-P1 INT8 berukuran 6.728 byte dengan estimasi memori aktivasi 76 byte dan latensi inferensi 1,50 µs. Model telah dijalankan pada purwarupa berbasis ESP32-S3, tetapi karena nilai sumber daya tersebut diukur pada komputer, kinerja aktualnya di perangkat masih perlu diverifikasi melalui pengukuran on-device.

Electronic noses (e-noses) have been widely developed to monitor the quality of agricultural products, including the ripeness of bananas. However, most previous studies report only accuracy without considering model size, memory (RAM), and inference latency on a microcontroller. This study compares Multilayer Perceptron (MLP) and 1-Dimensional Convolutional Neural Network (1DCNN) architectures for classifying three odor classes, namely air (control), unripe banana, and rotten banana, using an array of four MEMS gas sensors (CO, VOC, CH, and HCHO) based on an ESP32-S3. Each of the 46 captures was summarized into 16 features, namely the maximum response, steady state, rising slope, and falling slope of each sensor, computed on a natural-logarithm scale and normalized with z-score. The models were tested by holding out one capture in turn over 46 rounds with three seeds. The experiments varied the dropout rate, number of neurons and hidden layers, normalization, number of epochs, and TensorFlow Lite INT8 quantization. The 40–20–10 MLP with a dropout of 0.25 (MLP-P1) achieved an accuracy of 97.10% ± 1.02, higher than the 1DCNN (94.93% ± 1.02), whereas the model without normalization reached only 71.74%. INT8 quantization reduced model size by 22.5–41.7% and the estimated activation memory by 74.3–75.0% without lowering the fidelity of the normalized models. The INT8 MLP-P1 model occupies 6,728 bytes with an estimated activation memory of 76 bytes and an inference latency of 1.50 µs. The model has been deployed on the ESP32-S3-based prototype; however, since these resource figures were measured on a computer rather than on the device, its actual on-device performance still needs to be verified through direct measurement.

Kata Kunci : Electronic Nose, Pisang, 1DCNN, MLP, Dropout rate, Normalization, TinyML, Kuantisasi INT8