IAES International Journal of Artificial Intelligence (IJ-AI)
Not a member yet
1769 research outputs found
Sort by
Performance evaluation of pre-trained deep learning model on garbage classification with data augmentation approach
Get PDFWaste classification is one of the interesting topics for classifications in which data can be very varied and complex. This data diversity is a challenge to develop a model that is able to classify well. The purpose of this study is to analyze the performance of the pre-trained deep learning model using a data augmentation approach. There are three pre-training models used in this study, namely residual networks 50 (ResNet50), visual geometric group with 16 layers (VGG-16), and MobileNetV2. The results showed that the MobileNetV2 model received the highest accuracy value, reaching 84.45% for data without augmentation. With data augmentation there is a decrease of 2.73%. Conversely, VGG-16 shows performance stability with an increase in accuracy with augmentation data, reaching 75.84%. While ResNet50 gets the lowest results compared to both models. The application of data augmentation techniques with the aim of increasing data variations does not always have an impact on increasing the generalization of the model
Improving multilayer perceptron on rainfall data using modified genetics algorithm
Get PDFRainfall prediction is essential for managing water resources, agriculture, and disaster response, particularly in regions affected by climate variability. This study introduces a modified genetic algorithm (MGA) to optimize hyperparameters of a multilayer perceptron (MLP) for rainfall forecasting. The MGA incorporates elitism to retain top-performing solutions and adaptive selection based on model accuracy. The proposed MGA–MLP model was tested on rainfall datasets from Australia and Indonesia (BMKG). Experimental results show that configurations with two hidden layers, rectified linear unit (ReLU) activation and limited-memory Broyden Fletcher Goldfarb Shannon (LBFGS) optimizer, a learning rate of 0.001 and 1000 epochs consistently delivered strong performance. The model achieved accuracies of 86.02% and 79.05%, respectively. These findings indicate that MGA significantly improves MLP performance and provides a reliable, generalizable method for rainfall prediction across diverse climatic conditions
Comparative evaluation of machine learning models for intrusion detection in WSNs using the IDSAI dataset
Get PDFThis paper provides comparative assessment of three lightweight machine learning (ML) models (logistic regression (LR), random forest (RF), and gradient boosting (GB)), which are employed to detect intrusions in wireless sensor networks (WSNs) using the IDSAI dataset. The goal is to determine the most effective and deployable classifier within the constraints of WSN resources. In order to prevent data leakage and report accuracy, precision, recall, F1-score, and receiver operating characteristic-area under the curve (ROC-AUC) with mean±SD, we implement stratified 5-fold cross validation with in fold preprocessing. The results indicate that RF provides the most optimal generalization and overall performance (accuracy 0.9994 ± 0.0001, precision 0.9995±0.0001, recall 0.9994±0.0001, F1-score 0.9994±0.0001, ROC–AUC 0.9998 ± 0.0000). RF is closely followed by GB (accuracy 0.9990±0.0001, precision 0.9995±0.0001, recall 0.9985±0.0001, F1-score 0.9990 ± 0.0001, ROC-AUC ≈ 1.0000). LR demonstrates limitations in linearly overlapping classes, as evidenced by its high precision but reduced recall (accuracy 0.9167±0.0010, precision 0.9829±0.0002, recall 0.8481±0.0018, F1-score 0.9105 ± 0.0011, ROC–AUC 0.9707 ± 0.0001). In order to evaluate deployability, we characterize the inference throughput on a modest PC: LR ∼ 6.5 × 105 samples/s, GB ∼ 2.2 × 105 samples/s, and RF ∼ 1.3 × 105 samples/s, indicating a tiered intrusion detection system (IDS) (LR at sensors, RF at cluster-heads, and GB at the gateway). We also address the potential dangers of overfitting that may arise from the cleanliness of the dataset and provide a roadmap for future validation on a more diverse set of traffic. The research establishes a baseline for lightweight IDS in actual WSNs that is deployable and reproducible
Classification of regional language dialects using convolutional neural network and multilayer perceptron
Get PDFRegional languages are vital for communication and preserving cultural identity, safeguarding local heritage. However, globalization and modernization endanger their existence as they are increasingly replaced by national or global languages. Despite progress in dialect recognition research, particularly for certain languages, further studies are needed to improve model performance and address less-represented dialects, including those in Indonesia. This study enhances a custom-built dataset for dialect recognition through the application of data augmentation techniques, specifically adding noise, time stretching, and pitch shifting. Using Mel-frequency cepstral coefficients (MFCC) for feature extraction, it evaluates the performance of convolutional neural network (CNN) and multilayer perceptron (MLP) in classifying six Indonesian dialects. Results indicate that CNN outperformed, achieving 97.92% accuracy, 97.90% recall, 97.97% precision, 97.92% F1-score, and a kappa score of 97.49% with combined augmentation techniques, setting a foundation for further research
Hybrid AI framework for anomaly detection and root cause analysis in multi-agent systems
Get PDFAnomaly detection and root cause analysis (RCA) are critical for securing intelligent systems against evolving threats. Traditional models often suffer from high false alarms, weak adaptability to streaming contexts, and limited interpretability. This work proposes a hybrid artificial intelligence (AI) framework that integrates machine learning (ML) with prior knowledge, semantic rules, and bio-inspired modeling. The approach strengthens detection of diverse attacks, including DoS/DDoS, Probe, U2R, and R2L, while reducing human intervention. Experiments on the NSL-KDD dataset demonstrate that our method decreases spurious alerts by up to 90%, improves accuracy by 2–4%, and reduces false positives/negatives by about 4%. Beyond statistical gains, the framework ensures robustness in real-time environments, offering interpretable and scalable anomaly detection for heterogeneous systems. These results highlight the potential of hybrid symbolic–subsymbolic AI to enhance reliability in next-generation security infrastructures
Advanced risk assessment using machine learning and sentiment analysis on log data
Get PDFStandard risk assessment approaches are sometimes time-consuming and subjective. In order to overcome these challenges an innovative method will be presented in this article by mixing sentiment analysis and machine learning (ML). The suggested technique improves the effectiveness, precision, and scope of risk insights when it comes to the detection of feelings in logs via the use of automated data collection. The research examines several different ML classifiers and makes use of a deep learning model that has been pre-trained to evaluate risks in logs that are multi-linguistic. This proves the adaptability and scalability of our technique when used in a multilanguage setting. This combination of sentiment analysis and ML are a significant advancement in comparison to traditional approaches since it enables real-time processing and delivers important insights into the management of organizational risks
Comparison of HSV-color and ANN-HSV-color segmentation for detecting soybean adulteration
Get PDFSoybeans are an important food crop, but their quality is often compromised by contamination with other materials, a process known as adulteration. Conventional methods for detecting adulteration are slow; therefore, there is a need for rapid and non-invasive alternatives. This study aimed to assess the capability of hue-saturation-value (HSV) color segmentation and its combination with artificial neural networks (ANN) to identify adulteration in soybean samples. This research employed image processing and machine learning to segment soybeans mixed with adulterants at concentrations of 5%, 10%, 15%, 20%, and 25%. The HSV method successfully distinguished soybeans and other materials, but some challenges were observed in shadow regions and areas with similar colors. The HSV-ANN model with six hidden layers performed well with a calibration accuracy of R² value of 0.97 and root-mean-square error (RMSE) of 2.16%, which provided more detailed segmentation, although it still had some problems in shadow regions and undetected corn embryo parts. The validation results indicated that the HSV model had an R² value of 0.98 and RMSE of 4.48%, while the HSV-ANN model had an R² value of 0.96 and RMSE of 1.3%. Both models were capable of predicting the levels of adulteration, and the HSV-ANN model proved to be more accurate. It is concluded that both methods are efficient; however, there is a need for more work on modeling and sampling to increase the segmentation precision and decrease the biases, especially in the shadow and overlapped color
Improving firewall performance using hybrid of optimization algorithms and decision trees classifier
Get PDFOne of the primary concerns of governments, corporations, and even individual users is their level of online protection. This is because a large number of attacks target their primary assets. A firewall is a critical tool that almost every organization uses to protect its assets. However, firewalls become less reliable when they deal with large amounts of data. One method for reducing the amount of data and enhancing firewall performance is feature selection. The main aim of this study is to enhance the firewall's performance by proposing a new feature selection method. The proposed feature selection method combines the strengths of Harris Hawks optimization (HHO) and whale optimization algorithm (WOA). Experiments were performed utilizing the NSL-KDD dataset to measure the effectiveness of the proposed method. The experiments employed the decision trees (DTs) as a machine classifier. The experimental results show that the achieved accuracy is 98.46% when using HHO/WOA for feature selection and DT for classification, outperforming the HHO and WOA when used separately for feature selection. The study's findings offer insightful information for researchers and practitioners looking to improve firewall effectiveness and efficiency in defending internet connections against changing threats
Enhancing energy efficiency and accuracy in IoT-based wireless sensor networks using machine learning
Get PDFThis study presents a novel sensor data fusion framework designed to improve accuracy and energy efficiency in internet of things (IoT)-driven wireless sensor networks (WSNs). The proposed approach combines machine learning techniques with the Kalman filter, addressing the limitations of traditional methods, such as high computational overhead and limited precision. By utilizing machine learning algorithms for pattern recognition and the Kalman filter for precise state estimation, the framework optimizes data processing while minimizing energy consumption. MATLAB-based simulations validate the model’s effectiveness, demonstrating a significant improvement in key performance metrics, including F1-score, recall, and precision, with an overall accuracy of 98.36%. The results highlight the framework’s ability to enhance fault tolerance, accelerate convergence rates, extend network lifespan, and optimize energy utilization, making it highly suitable for real-time data fusion applications in complex sensor environments. Furthermore, the proposed hybrid model is scalable and adaptable, allowing it to be implemented across various fields, including environmental surveillance, industrial automation, and healthcare monitoring. With integration of intelligent data processing techniques, this research contributes to the development of sustainable and efficient IoT-based monitoring systems capable of handling dynamic and resource-constrained environments
Optimization control design and simulation of furnace-fired boiler exit pressure: leveraging disruptive technology
Get PDFThe efficient operation of furnace-fired drum boilers is critically dependent on the precise control of downstream exit pressure, especially in the presence of stochastic heat fluctuations. This paper presents a stochastic control approach for regulating the downstream exit pressure in a furnace-fired boiler subject to random heat fluctuations. A stochastic model of the boiler dynamics is developed, incorporating heat transfer and combustion uncertainties. By leveraging disruptive technology, such as the model predictive control (MPC), strategies were designed to optimize the downstream exit pressure in real-time, and minimizing deviations from the set point. Simulation studies demonstrated the effectiveness of the proposed approach in maintaining a stable exit pressure despite random heat fluctuations. Results show significant improvements in boiler performance and efficiency compared to traditional proportional integral derivative (PID) control. The proposed stochastic control strategy offers a promising solution for reliable and efficient operation of furnace-fired boilers under uncertain conditions