Metode Modified Weighted Mean Filter dan Continuous Capsule Network untuk Pengenalan Emosi Berdasar Sinyal Electroencephalogram pada Peserta Pembelajaran

I MADE AGUS WIRAWAN, Retantyo Wardoyo; Danang Lelono; Sri Kusrohmaniah

2022 | Disertasi | DOKTOR ILMU KOMPUTER

Reaksi emosi memiliki peran yang sangat penting dalam keterlibatan, daya ingat dan penalaran peserta pembelajaran. Dari beberapa pendekatan, pengenalan emosi berdasar sinyal Electroencephalogram memiliki beberapa keunggulan dibandingkan dengan pendekatan eksternal maupun internal lainnya. Namun, beberapa permasalahan pada pengenalan emosi berdasar sinyal Electroencephalogram, diantaranya: (1) Sinyal baseline Electroencephalogram mengandung artefact, sehingga tidak dapat merepresentasikan kondisi netral dari partisipan, (2) Proses convolution yang terdapat pada metode Capsule Network mengakibatkan hilangnya informasi spasial antar channels dan frekuensi band dari sinyal eksperimen Electroencephalogram. Untuk mengatasi masalah tersebut, penelitian ini mengembangkan model pengenalan emosi peserta pembelajaran berdasar sinyal Electroencephalogram yang terdiri dari metode Modified Weighted Mean Filter untuk menghilangkan artefact pada sinyal baseline Electroencephalogram metode Differential Entropy untuk ekstraksi fitur, menggunakan satu metode yang tepat untuk proses baseline reduction, metode 3D Cube untuk representasi fitur, serta metode Continuous Capsule Network untuk klasifikasi. Berdasarkan keseluruhan eksperimen pada dataset DEAP, DREAMER, dan AMIGOS, dua metode yang diusulkan, yaitu metode Modified Weighted Mean Filter dan metode Continuous Capsule Network terbukti dapat mengatasi permasalahan dalam penelitian ini. Selain itu dari tiga metode baseline reduction yang telah dikaji, metode Relative Difference lebih tepat digunakan untuk proses baseline reduction. Proses baseline reduction dapat menggunakan sinyal baseline maupun sinyal eksperimen Electroencephalogram. Metode yang diusulkan pada penelitian ini dapat menghasilkan model pengenalan emosi berdasarkan sinyal Electroencephalogram yang lebih akurat daripada model pengenalan emosi berdasarkan sinyal Electroencephalogram pada penelitian sebelumnya. Model yang telah teruji ini, selanjutnya digunakan untuk mengenali emosi berdasarkan sinyal Electroencephalogram pada peserta pembelajaran.

Emotional reactions have a vital role in learning participants' involvement, memory, and reasoning. From several approaches, emotion recognition based on Electroencephalogram signals has several advantages compared to other external and internal procedures. However, there are several problems in emotion recognition based on Electroencephalogram signals, including: (1) The baseline Electroencephalogram signal contains artifacts, so it cannot represent the neutral condition of the participants, (2) The convolution process found in the Capsule Network method results in loss of spatial information between channels and frequencies bands of experimental Electroencephalogram signals. To overcome this problem, this study developed a model for recognizing the emotions of learning participants based on the Electroencephalogram signal, which consists of the Modified Weighted Mean Filter method to remove artifacts in the baseline Electroencephalogram signal, the Differential Entropy method for feature extraction, using one appropriate baseline reduction method, the 3D Cube for feature representation, as well as the Continuous Capsule Network method for classification. Based on all experiments on the DEAP, DREAMER, and AMIGOS datasets, the two proposed methods, namely the Modified Weighted Mean Filter method and the Continuous Capsule Network method, are proven to be able to overcome the problems in this study. In addition, the Relative Difference method is more appropriate for the baseline reduction process of the three baseline reduction methods studied. The baseline reduction process can use the baseline signal or the Electroencephalogram experimental signal. The method proposed in this study can produce an emotion recognition model based on Electroencephalogram signals that is more accurate than the emotion recognition model based on Electroencephalogram signals in previous studies. This tested model is then used to recognize emotions based on Electroencephalogram signals in learning participants.

Kata Kunci : Electroencephalogram, Emotion, Baseline Reduction, Continuous Capsule Network, Modified Weighted Mean Filter