Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Optimization of Accuracy Improvement through Modified ShuffleNet Architecture in Rice Classification
No full textAccurate rice classification is essential to determine the quality and market value of rice. Traditional methods of rice classification are often time-consuming and error-prone, so a more efficient and accurate solution is needed. This study aims to optimize rice classification using Convolutional Neural Networks (CNN) combined with the ShuffleNet architecture, which offers high computational efficiency without sacrificing accuracy. The dataset used comes from Kaggle, containing 8750 rice grain images divided into five classes: Arborio, Basmati, Ipsala, Jasmine, and Karacadag. The uniqueness of this study is the application of ShuffleNet Proposed in rice classification, which provides improved performance compared to basic CNN models such as MobileNet, ShuffleNet, and RestNet. The results showed that the MobileNet model achieved 80% accuracy, RestNet 94%, and ShuffleNet achieved 100% accuracy with precision, recall, and F1 values also 100%. However, the ShuffleNet model experienced overfitting when tested with new data, resulting in an accuracy of only 20%. To overcome this, further optimization was carried out on the model. The results of statistical tests (paired t-test and Wilcoxon test) show significant differences between ShuffleNet Proposed and other models, which proves that the improvements applied to this model provide significant improvements. The implications of this study can improve the efficiency and accuracy of rice classification, which has the potential to improve the quality and market value of rice in the agricultural industry
Stacking Ensemble Learning Model for Intrusion Detection in Electrical Substation
No full textElectrical substations are crucial infrastructure in power transmission and distribution but are increasingly vulnerable to cyber threats. However, existing intrusion detection systems (IDS) face challenges such as high false positive rates, limited adaptability to emerging attack patterns, and imbalanced detection across different intrusion types. This study proposes a Stacking Ensemble Learning model to enhance intrusion detection accuracy in electrical substations. The proposed model integrates Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and XGBoost (XGB) as base models with XGB acting as the meta-model. A real-world electrical substation IEC 60870-5-104 network traffic dataset comprising 319,949 instances with multiple attacks, such as DoS, Port Scan, NTP DdoS, IEC 104 Starvation, Fuzzy Attack, Flood Attack, and MITM, was used for this study. The results showed that the stacking model had the best accuracy (0.99990), precision (0.99990), recall (0.99990), and F1-score (0.99990), beating out the base, Bagging, and Boosting models. T-test results further confirmed statistical significance, with p-values of 0.00428 (LR), 0.04237 (SVM), 0.00000 (XGB), 0.00057 (KNN), 0.00549 (Boosting), and 0.00000 (Bagging) reinforcing the superiority of the Stacking Ensemble Learning approach. These findings highlight the effectiveness of Stacking Ensemble Learning in enhancing the detection accuracy of IDS for electrical substations and outperforming traditional models and other ensemble learning methods by minimizing false positives and false negatives
Improving the Accuracy of Tourism Recommendation System Based on Neural Collaborative Filtering
No full textThis study proposes a Neural Collaborative Filtering (NCF) model for tourism recommendation systems by integrating user ratings and review data. This model was developed to overcome the limitations of conventional recommendation systems that rely solely on numerical data, by adding contextual information from user reviews to improve the accuracy of preference prediction. The development process includes data preprocessing, conversion of text reviews into numerical representations using embedding techniques, and the application of NCF models with various parameter configurations. Experimental results show that the NCF model that combines rating and review data produces the best performance with Root mean Square Error (RMSE) values of 0.892, Hit Ratio at 10( HR@10) of 0.735, and Normalized Discounted Cumulative Gain at 10 (NDCG@10) of 0.629, outperforming models that only use one type of data. These results demonstrate that combining numerical and textual information can improve the model's understanding of user preferences, resulting in more relevant tourist destination recommendations. These findings contribute to the development of artificial intelligence-based recommendation systems in the tourism sector
Spatial-Temporal Analysis of Earthquakes in Indonesia with Deep Learning Models: Performance Evaluation of CNN, LSTM, and Hybrid CNN-GRU
No full textIndonesia, located along the Pacific Ring of Fire, experiences high seismic activity with over 6,000 earthquakes annually. Accurate earthquake prediction remains a major challenge because of the complexity of geological dynamics and limitations of traditional methods in capturing nonlinear seismic patterns. Although deep learning approaches have shown promise, previous studies have often treated spatial and temporal analyses separately, limiting holistic predictive performance. This study proposes a novel hybrid CNN-GRU deep learning model that integrates spatial feature extraction CNN and temporal sequence modeling GRU, and compares its performance with of that CNN, LSTM, GRU, and Bidirectional LSTM using a dataset of 117,251 earthquake events in Indonesia (2008–2024). The results show that Bidirectional LSTM achieved the best temporal accuracy (R² 0.653, RMSE 0.592), while the hybrid CNN-GRU provided balanced spatial-temporal performance (R² 0.587). Notably, the performance gap between Bidirectional LSTM and other models (e.g., Hybrid CNN-GRU) was statistically validated via paired t-test (p < 0.05). The proposed models generalize well to unseen regions such as Maluku-Papua. The key contribution is the hybridization of spatial-temporal learning in a single-model architecture - where CNN processes latitude-longitude coordinates via 1D convolutions while GRU handles temporal sequences - an approach lacking in earlier works. This directly improves early warning systems in seismically active areas by providing 32% higher accuracy than conventional methods
Sonified Cryptography: Secure Text Encoding with DNA and Non-Speech Audio
No full textThe increasing demand for data security in digital communication, particularly in high-risk sectors like defense, has led to the exploration of innovative encryption approaches. This study presents a dual-layer encryption model that enhances information concealment by integrating DNA-based cryptography with audio signal encoding. Initially, plaintext is converted into binary and obfuscated using XOR operations with randomly generated DNA sequences. The resulting DNA nucleotide sequences (A, G, C, T) form the first layer of encryption. In the second layer, these sequences are audified by mapping each nucleotide to a specific frequency, thereby transforming the encrypted data into non-speech audio signals. To evaluate the integrity and uniqueness of the encryption-decryption process, Fast Fourier Transform (FFT)-based cross-correlation is applied, comparing the original and recovered audio signals. The proposed method is implemented in MATLAB and tested on various input strings. Results demonstrate significant improvements in encryption speed and security, with the added benefit of imperceptibility in audio form. The method outperforms existing DNA-based techniques in terms of computational efficiency and resistance to brute-force attacks. This hybrid cryptographic technique offers a promising solution for secure, covert data transmission in sensitive applications
Comparing Optimization Algorithms in ANN Models for House Price Prediction in Pekanbaru
No full textThis study evaluates the performance of five optimization algorithms in Artificial Neural Network (ANN) models for predicting house prices in Pekanbaru. The optimizers tested include Adam, AdaDelta, Stochastic Gradient Descent (SGD), Nadam, and Adaptive Sharpness-Aware Minimization (ASAM). A total of 3,149 house sales records were collected from rumah123.com between January and December 2024. After cleaning 148 incomplete entries, 3,001 valid records remained. The dataset included seven features: price, location, number of bedrooms, number of bathrooms, land area, building area, and garage capacity, with the location encoded using one-hot encoding. The research involved a literature review, problem formulation, data acquisition, preprocessing, model development, and evaluation. Model performance was assessed using the Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE). The results show that SGD consistently achieved the best performance, particularly at a 90:10 train-test split, with the lowest MAPE (1.74%) and MSE (0.3279). Adam and Nadam also performed well, while ASAM had the highest error (MAPE 6.14%). These findings indicate that SGD was the most effective optimizer for this dataset. Future research should explore larger datasets and advanced hyperparameter tuning to improve the generalizability of this model
Optimizing Menu Planning for Children with Autism Using Improved Multi-Goal Programming Model
No full textMenu planning for every individual is essential to obtain a balanced and healthy food intake for growth and development. Children with Autism Spectrum Disorders (ASD) face more feeding difficulties than their peers due to neurodevelopmental disorders such as social skills problems and repetitive behaviors. They also tended to refuse or be selective for certain food intakes. Proper menu planning for children with ASD is important to maintain their overall well-being and mitigate autism-related dietary issues. The manual menu planning for children with ASD does not consider proper nutritional intake, food variation, or total cost minimization. Currently, the application of mathematical modelling for menu planning in children with ASD is limited. Thus, this study aims to explore the extent to which the optimization approach can solve the menu planning problem with more than one objective. Finally, this research constructed daily menu planning for children with ASD, which indicates the feasibility of utilizing the Improved GP (IGP) model compared to the Goal Programming model (GP) in terms of the value for the deviational variables for the unachieved goals. The unachieved deviational variables by IGP model for Day-2 had decreased by 17.69% and by 34.43 % on Day-3. The total cost of the IGP model is also less than RM 0.50 of the GP model
Improving Low-Light Face Recognition using DeepFace Embedding and Multi-Layer Perceptron
No full textFacial recognition systems often struggle under extreme lighting conditions, which distort facial features and reduce recognition accuracy. This study introduces a novel integration of DeepFace embeddings with a lightweight Multi-Layer Perceptron (MLP) classifier tailored to improve facial recognition under extreme lighting conditions. This combination has not been explored in previous studies and offers a compact alternative to conventional CNN-based methods. The Labeled Faces in the Wild (LFW) dataset was augmented using rotation, flipping, and lighting variations, and further enhanced with CLAHE for improved contrast under poor illumination. The resulting 128-dimensional DeepFace embeddings were classified using a four-layer MLP with LeakyReLU activation, Batch Normalization, and Dropout. The model was evaluated across three data-splitting schemes (70:30, 80:20, and 90:10), with the 80:20 configuration achieving the highest accuracy of 95.16%. Compared to the baseline CNN, the proposed method demonstrated superior robustness to illumination variations. This makes the proposed model suitable for real-time applications such as biometric authentication and AI-based surveillance systems
Question Answering through Transfer Learning on Closed-Domain Educational Websites
No full textNavigating complex educational websites poses challenges for users looking for specific information. This research discusses the problem of efficient information search on closed-domain educational platforms, focusing on the Universitas Indonesia website. Leveraging Natural Language Processing (NLP), we explore the effectiveness of transfer learning models in Closed Domain Question Answering (QA). The performance of three BERT-based models, including IndoBERT, RoBERTa, and XLM-RoBERTa, are compared in transfer and non-transfer learning scenarios. Our result reveals that transfer learning significantly improves QA model performance. The models using transfer learning scenario showed up to 4.91\% improvement in the F-1 score against those using non-transfer learning scenario. XLM-RoBERTa base outperforms all other models, achieving the F-1 score of 61.72\%. This study provides valuable insights into Indonesian-language NLP tasks, emphasizing the efficacy of transfer learning in improving closed-domain QA on educational websites. This research advances our understanding of effective information retrieval strategies, with implications for improving user experience and efficiency in accessing information from educational websites
NLP-Based Intent Classification Model for Academic Curriculum Chatbots in Universities Study Programs
No full textChatbots are increasingly prevalent in various fields, including academic fields. Universities often rely on lecturers and staff for information access, which can lead to delays, limited availability outside working hours, and the risk of missed questions. This study aims to develop a chatbot model capable of addressing questions about the curriculum through intent classification, reducing reliance on manual responses, and providing a solution that ensures quick, accurate information retrieval. The research focuses on optimizing the IndoBERT model for intent classification and addresses challenges that arose due to imbalance data, which could have impacted model performance. Data was collected through an open poll on common curriculum-related questions asked by students. To address data imbalance, we tried oversampling techniques, such as SMOTE, B-SMOTE, ADASYN, and Data Augmentation. Data augmentation was chosen and successfully addressed the imbalance problem while maintaining data semantics effectively. We achieved the best model with hyperparameters batch size of 8, learning rate of 0.00001, 15 epochs, and 64 neurons in the hidden layer, resulting in 98.7% accuracy on the test data. Evaluation metrics further demonstrate the model's robustness across multiple intents. This research demonstrates the advantages of the IndoBERT model in intent classification for academic chatbots, achieving excellent performance