Computer Engineering and Applications Journal
Not a member yet
101 research outputs found
Sort by
Implementation of Feature Selection for Optimizing Voice Detection Based on Gender using Random Forest
No full textGender-based voice detection is one of the machine learning applications that has various benefits in technology and services, such as virtual assistants, human-machine interaction systems, and voice data analysis. However, the use of too many features, including irrelevant features, can cause a decrease in accuracy and model performance. This research aims to optimize voice-based gender detection by applying a feature selection method to select significant features based on their correlation value to the target. Experimental results show that by using only the significant features selected through correlation analysis, the accuracy of the model is significantly improved compared to using all available features. This research confirms the importance of feature optimization to support the development of more efficient and accurate gender-based speech detection models
LinkedIn User Interface Optimization with Design Thinking Methods: Improve user experience and engagement
No full textLinkedIn is one of the world\u27s biggest social networking platforms, used to build connections, exploring career opportunities, and sharing professional insights. Users are increasingly prioritizing simple, effective, and user-friendly design, so improving LinkedIn\u27s user interface is exceptionally imperative. LinkedIn is often perceived as confusing by its users, with many stating that features within the application are difficult to find and navigation is inefficient. This problem can hinder LinkedIn\u27s effectiveness in meeting the professional needs of its users. This research aims to optimize the LinkedIn interface using a design thinking approach. The method involves five stages: empathize, define, ideate, prototype, and test. Design thinking prioritizes user needs, wants, and challenges. The advantage of the design thinking approach is that it becomes a solution, so what is produced is more relevant and effective in solving the real problems of the users. Data was collected through questionnaires to identify the key issues with navigation and ease of use of features on LinkedIn. The results show that the design thinking approach allows the creation of a simpler user interface with more effective navigation and features that are easy to find. The conclusion of this research is that the design thinking method is able to solve interface design problems in the LinkedIn application
Anxiety Detection for Autism Children through Vital Signs Monitoring using a Socially Assistive Robot
No full textSocially Assistive Robot (SAR) to detect anxiety levels in children with Autism Spectrum Disorder (ASD), a condition often accompanied by difficulties in recognising and expressing emotions, including anxiety. Early recognition of anxiety in children with Autism Spectrum Disorder (ASD) is crucial as it can affect their behaviour and social interactions. This SAR monitors vital signs namely blood pressure, heart rate and body temperature. This study involved children with Autism Spectrum Disorder (ASD) with two conditions, namely Asperger Syndrome and Classical Autism who interacted with a Socially Assistive Robot (SAR) equipped with a tensimeter (MPS20N0040D sensor) for blood pressure, MAX30100 sensor for heart rate, and MLX90614 sensor to measure body temperature. Results show that the Socially Assistive Robot (SAR) is able to measure vital signs with high accuracy and provide an indication of anxiety levels effectively, as vital signs correlate with anxiety levels. These findings demonstrate the potential of the Socially Assistive Robot (SAR) as a reliable tool in anxiety monitoring in children with ASD, with important implications for the development of future therapeutic interventions
Emotion Classification in Indonesian Text Using IndoBERT
No full textMental health issues have become a challenge that affects many individuals around the world. A 2018 WHO report noted an increase in deaths by suicide, with a frequency of one case every 40 seconds. The Ipsos Global 2023 survey showed that 44% of respondents in 31 countries are concerned about mental health, while 30% identified stress as a major issue. In Indonesia, the mental health situation is also a serious concern. The 2022 I-NAMHS survey found that 34.9% of adolescents face mental health problems, but only 2.6% of them utilize counseling services. Emotion detection in text is challenging due to the absence of facial expressions or voice modulation. This study aims to classify emotions in Indonesian text using the IndoBERT model. The dataset used consists of 5079 tweets with five emotion labels: Angry, Fear, Joy, Love, and Sad. Parameter variations include the composition of training, validation, and test data split (80:10:10, 75:15:15, and 60:20:20), as well as the combination of learning rate (1e-2 to 1e-7) and batch size (8, 16, and 32). The model was trained for 25 epochs with the application of early stop and patience for 5 epochs. The experimental results showed that the composition of data split 80:10:10, learning rate 1e-6, and batch size 8 resulted in optimal classification. Although some experiments showed indications of overfitting, this research has important implications in the early detection of emotions and can help in mental health treatment efforts
Improving Low-Cost Single-Phase Inverter Performance using DRL-Based Control System: Experimental Validation
No full textThis paper presents the improvement of a low-cost, single-phase pure sine wave inverter controlled by a deep reinforcement learning (DRL) agent. The study addresses the challenge of lacking performance of low-cost inverter, which is primarily due to the stability requirements of conventional control strategies. A DRL- based control approach is proposed to enhance voltage and frequency stability while reducing the need for extensive manual tuning. The system is validated through both simulation and experimental verification in a microgrid islanded configuration. The results demonstrate that the DRL-based inverter effectively maintains 220 VRMS at 50 Hz, achieving a stable root mean square voltage of 219.8 V, and a total harmonic distortion (THD) below 8%. The use of DRL making it an attractive solution for renewable energy systems, off-grid applications, and rural electrification. This study highlights the feasibility of DRL in power electronics and suggests that further optimization of training generalization and computational efficiency could enhance real-time and grid-tied deployment. The findings contribute to the advancement of intelligent inverter control, offering an alternative for next-generation microgrid and distributed energy systems
Implementation of Weightless Neural Network in Embedded Face Recognition for Eye and Nose Pattern Mobile Identification
No full textThe pattern of the human face is a form of self-identity and also a form of originality for each individual. The development of facial recognition technology impacts its application in various computing devices, both in computer vision and on single-chip processors. One of the continuously developed implementations is in the form of robot vision by identifying facial features. This research aims to develop a facial recognition system focusing on the identification of the eye and nose areas. This research utilizes the Weightless Neural Network (WNN) method with the Immediate Scan technique. The combination of methods allows for rapid and accurate pattern recognition, even when the face changes position. The detection process is carried out using the Haar Cascade Classifier algorithm, which functions to recognize faces and divides the area into nine different zones to ensure accurate identification. The hardware implementation was carried out on a Raspberry Pi for face detection and facial pattern recognition, as well as the data processor for the robot vision sensor and actuator on the microcontroller. The results of the robot\u27s movement testing have worked well according to the calculation of GPS data values to determine the robot\u27s last position. Then, in the face pattern recognition process, it shows that the proposed method can achieve a maximum accuracy level of up to 98.87% in testing with the internal data set, while testing under different conditions experiences a slight decrease in accuracy to 91.38%. The highest similarity percentage to the faces of other individuals reached 75.69%, indicating that this method is quite adaptive to various facial variations. The execution time of the identification process ranges from 11 ms to 17 ms, depending on the amount of data compared during the scanning. This research is expected to serve as a foundation for further development in robotics systems and embedded system-based facial recognition
Implementation Of The Eco Cycle Classifier Deep Neural Network (EECDN-Net) Model For Image-Based Waste Classification
No full textWaste management is a global challenge that demands effective solutions, especially in classification and recycling processes. This study presents the development of an Eco Cycle Classifier Deep Neural Network (ECCDN-Net) model based on deep learning for image-based waste classification. The model integrates the DenseNet201 and ResNet18 architectures to improve visual feature extraction and reduce the vanishing gradient problem. The dataset used is TrashNet, which contains 2,527 images across six waste categories. Training was conducted over 50 epochs, utilizing data augmentation and class balancing to address the imbalanced data. Results show that ECCDN-Net achieved a validation accuracy of 87.75% and an average F1-score of 0.88. The confusion matrix reveals that the model performs well in recognizing most classes, although it faces difficulty distinguishing categories with high visual similarity, such as plastic and glass. This research demonstrates that ECCDN-Net effectively provides accurate waste classification and could serve as a promising solution for more adaptive and sustainable automatic waste sorting
Deep Neural Networks for Intelligent Voice Authentication Systems in Large-Scale Electronic Voting
No full textThe authentication of eligible voters is an area of concern that needs further exploration of the prospects of electronic voting systems. The integration of voice authentication in electronic voting systems for varying numbers of disabled and prospective voters should be secure, scalable and suitable in both federal and state elections. Machine learning (ML) is an evolving field of computing that presents prospects in electronic voting. Applying ML algorithms to electronic voting provides optimal solutions to a wide range of biometric authentication challenges. This paper presents the design of an effective voice classification algorithm from a narrower perspective that can be used in developing prototype electronic voting systems in large-scale voting scenarios, particularly for disabled voters. Applying the knowledge of deep neural networks, a three hidden layer network using a feed-forward architecture is designed for classifying voice data acquired from prospective voters. The proposed design is tested on two different datasets and is adapted to handle small and vast amounts of voters’ voice information. Results indicated average training and average validation accuracies of 92% and 97% respectively for both deep learning models for inclusivity and accountability of disabled voters in secure electronic voting systems
Enhanced Short-Term Residential Load Forecasting Using K- means Clustering and Iterative Residual LSTM Networks
No full textAccurate short-term load forecasting (STLF) is essential for optimizing energy management systems, ensuring operational efficiency, and balancing supply and demand in power grids. This study introduces a hybrid model, K-RNLSTM, which integrates K-means clustering with iterative Residual Long Short-Term Memory (LSTM) networks to improve prediction accuracy. The K-means clustering algorithm categorizes similar load patterns, allowing the model to handle seasonal and hourly variations more effectively. Iterative ResBlocks are incorporated within the LSTM framework to capture complex non-linear dependencies and improve the learning process without suffering from degradation. The model was evaluated using real- world residential electricity consumption data across four seasons: winter, spring, summer, and autumn. The K-RNLSTM model consistently outperformed traditional methods such as Extreme Learning Machines (ELM), Seasonal-Trend Loess (STL), Gated Recurrent Units (GRU), and standard LSTM in terms of Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The results demonstrated that K-RNLSTM achieved an average RMSE of 0.71, MAE of 0.43, and MAPE of 1.31%, surpassing benchmark models across all seasonal variations. Furthermore, the integration of ResBlocks significantly improved the model\u27s ability to minimize large forecasting errors, particularly during peak demand periods. This research demonstrates the effectiveness of combining clustering techniques with deep learning models for short-term load forecasting, offering a robust solution for power system operators to optimize energy distribution and reduce operational costs
Efficient Hierarchical Temporal Audio-Video Cross-Attention Fusion Network For Audio-Enhanced Text-To-Video Retrieval
No full textWith video and audio being integral to modern multimedia content, accurately retrieving relevant segments based on textual queries is crucial for enhancing user experience and information accessibility. However, contextual misalignment across video segments presents significant challenges, particularly when different segments exhibit varying degrees of relevance to specific portions of a text query. To address this issue, a novel Hierarchical Temporal Audio-Video Cross-Attention Fusion Network has been developed. This model utilizes a Video Swim Feature Pyramid video encoder to enhance the extraction of multi-scale spatial features and capture intricate details within videos. Additionally, a Temporal RoBERTa Graph Network serves as the text encoder, enabling a deep understanding of relationships within the text and allowing for minute interpretations of queries that encompass multiple themes. To effectively align video and audio representations with textual queries, the model employs a Hierarchical multiscale spatial-temporal attention mechanism. Furthermore, an Audio Spectrogram Short-Term Memory Transformer is utilized to capture the temporal dynamics of complex audio streams. To refine audio-text alignment, the model incorporates a Threshold-Based audio-text Dynamic Time cross-attention block, which selectively filters irrelevant audio components and dynamically adjusts for temporal misalignments. The experimental results demonstrate that the proposed model significantly enhances retrieval accuracy by effectively aligning video and audio representations with textual queries, resolving multi-scene transitions, and isolating relevant audio cues among complex soundscapes.