Showing 47 results for Neural Network
Noor Fazliana Fadzail, Samila Mat Zali, Ernie Che Mid,
Volume 21, Issue 2 (6-2025)
Abstract
The activation function has gained popularity in the research community since it is the most crucial component of the artificial neural network (ANN) algorithm. However, the existing activation function is unable to accurately capture the value of several parameters that are affected by the fault, especially in wind turbines (WT). Therefore, a new activation function is suggested in this paper, which is called the double sigmoid activation function to capture the value of certain parameters that are affected by the fault. The fault detection in WT with a doubly fed induction generator (DFIG) is the basis for the ANN algorithm model that is presented in this study. The ANN model was developed in different activation functions, namely linear and double sigmoid activation functions to evaluate the effectiveness of the proposed activation function. The findings indicate that the model with a double sigmoid activation function has greater accuracy than the model with a linear activation function. Moreover, the double sigmoid activation function provides an accuracy of more than 82% in the ANN algorithm. In conclusion, the simulated response demonstrates that the proposed double sigmoid activation function in the ANN model can effectively be applied in fault detection for DFIG based WT model.
Elahe Rezaee Ahvanooii, Sheis Abolmaali,
Volume 21, Issue 3 (8-2025)
Abstract
Touch, one of the fundamental human senses, is essential for understanding the environment by enabling object identification and stable movements. This ability has inspired significant advancements in artificial neural networks for object recognition, texture identification, and slip detection applications. However, despite their remarkable capacity to simulate tactile perception, artificial neural networks consume considerable energy, limiting their broader adoption. Recent developments in electronic skin technology have brought robots closer to achieving human-like tactile perception by enabling asynchronous responses to temperature and pressure changes, thereby enhancing robotic precision in tasks like object manipulation and grasping. This research presents a Spiking Graph Convolutional Network (SGCN) designed for processing tactile data in object recognition tasks. The model addresses the redundancy in spiking-format input data by employing two key techniques: (1) data compression to reduce the input size and (2) batch normalization to standardize the data. Experimental results demonstrated a 93.75% accuracy on the EvTouch-Objects dataset, reflecting a 4.31% improvement, and a 78.33% accuracy on the EvTouch-Containers dataset, representing an 18% improvement. These results underscore the SGCN's effectiveness in reducing data redundancy, decreasing required time steps, and optimizing tactile data processing to enhance robotic performance in object recognition.
Seyyedeh Ensiyeh Hashemi, Hamid Behnam,
Volume 21, Issue 3 (8-2025)
Abstract
Increasing the frame rate of ultrasound imaging while keeping image quality is important for following fast movements, especially the heart. There are different modalities for B-mode image recording, including line-by-line scanning with linear, phased, convex array, synthetic aperture imaging (STA), plane waves (PWI), then the combination of plane waves (CPWI), and so on. Researchers have tried to increase the frame rate in each case using different methods. Three approaches for this aim are data acquisition, post-processing, and beamforming. This article reviews these approaches and their solutions for compensating image quality reduction. Ultrafast ultrasound imaging, which provides exceptional temporal resolution (high frame rate), is promising in diagnosing heart diseases due to its ability to capture rapid heart movements. It can record images faster than conventional imaging, usually exceeding 1000 frames per second. This can be achieved through plane wave imaging (PWI). However, high frame rate data acquisition can lead to a decrease in image quality. Transmitting at different angles and then combining plane wave imaging is a popular method to enhance PWI quality but reduces the frame rate by the number of angles. As a result, researchers have aimed to increase the temporal resolution while compensating for the loss of quality.
Tara Sistani, Seyed Javad Kazemitabar,
Volume 21, Issue 4 (11-2025)
Abstract
Forests play several vital roles in our lives and provide various resources. However, in recent years, the increasing frequency of wildfires has led to the widespread burning and destruction of many forests and wildlands. Therefore, detecting forest fires and finding suitable solutions to address this issue has become one of the critical challenges for researchers. Today, with the advancement of artificial intelligence, forest fire detection using deep learning is an important method with the aim of increasing the efficiency of forest fire detection and monitoring systems. In this article, a method based on a type of convolutional neural network called Xception is proposed for classifying forest fire images. In this method, transfer learning technique is used on the proposed neural network and a new classifier is designed for the problem. Also, various hyperparameters have been used to optimize the performance of the proposed model. The proposed method is performed on the DeepFire dataset, which contains 1900 images equally divided between fire and no-fire classes. The results obtained from the implementation of the proposed method show that this method with an accuracy of 99.47% has achieved a favorable performance in classifying forest fire images.
Zead Mohammed, Basil Mahmood, Saad Saeed, Aws Hazim,
Volume 22, Issue 0 (3-2026)
Abstract
A mobile robot must be autonomous to avoid obstacles while traveling towards the target. The problem of dynamic obstacle avoidance is still a significant challenge. Reactive mobile robot navigations handled this problem, but using a single-stage module leads to a deficiency and a limitation in performance. This paper proposes combining an adaptive neuro-fuzzy inference system and a neural network. The data for obstacle severity classification were used to train the Bayesian regularization Back-Propagation Neural Network. The relative velocity and distance between the mobile robot and obstacles determine the zone. Zone 1 is dangerous, and Zone 5 is safe. This paper uses an Adaptive Neuro-Fuzzy Inference System (ANFIS) to avoid obstacles during the mobile robot's motion and to avoid collision. Based on our empirical study, we consider three essential features in this paper: the relative speed, distance, and angle between the robot and the obstacle as inputs to the obstacle avoidance system ANFIS. The output was a suggested steering angle and speed for the mobile robot. The simulation results for the tested cases show the capability of the proposed controller to avoid static and dynamic obstacles in a fully known environment. Our results show that the Adaptive Neuro-Fuzzy Inference System enhances the proposed controller's performance, reducing path length, processing time, and the number of iterations compared.
Balamanikandan A, Venkataramanaiah N, Sukanya M, Sudhakar Reddy N, Gomathy G, Venkatachalam K,
Volume 22, Issue 0 (3-2026)
Abstract
Physics-informed neural networks (PINNs) offer a promising route to bridge device-level simulations and compact circuit models. In this work, we present a hybrid modeling framework that integrates TCAD datasets with a baseline compact model and applies a PINN correction to capture stress-condition effects with high fidelity. The proposed approach achieves ≤ 2% route mean square error (RMSE) across more than 2,000 bias points, maintaining stable predictions under temperature (273–373 K) and radiation (0–100 krad) variations. Extracted Berkeley Short-channel IGFET Model (BSIM) parameters enable direct SPICE simulation, ensuring compatibility with standard circuit design workflows. For deployment, the trained PINN is exported as a quantized ONNX model, achieving sub-millisecond inference and ultra-low energy consumption (0.25 pJ/op) on a Cortex-M55 platform. This dual pathway supports both high-accuracy circuit simulation and real-time edge inference, making it suitable for embedded applications under constrained conditions. Comparative analysis with recent ANN-based models confirms that our physics-informed approach offers superior interpretability, SPICE readiness, and deployment efficiency. All datasets, code, and models are released to support reproducibility, benchmarking, and further research in compact modeling and edge-AI integration.
S. Tohidi, M. R. Mosavi,
Volume 22, Issue 1 (3-2026)
Abstract
A vital part of people's daily life is the position, navigation, and time service provided by the Global Positioning System (GPS), which is always accessible globally. Consequently, the security of the GPS receivers is crucial. Occasionally, intentional and unintentional interferences cause GPS location issues. Spoofing attack is the most severe interference to the GPS receivers, which results in positional mistakes. This paper's goal is to defend against the carry-off spoofing attacks. In a carry-off spoofing attempt, the spoofer transmits signals whose code phase and carrier frequency parameters are strikingly close to the actual signal in order to change the correlation values generated in the tracking stage. Discriminator output values alter as correlation values change. As a result, the Pseudo Random Noise (PRN) code generator unit creates a local replica, which forces the tracking loop to follow the fake signal instead of the real one. It is proposed in this paper that when spoofing attacks occur, discriminator output values be generated independently of correlation values. Specifically, when a spoofing signal is detected, the conventional discriminator is replaced by a Non-linear Autoregressive Exogenous Neural Network (NARX NN)-based predictor. This strategy protects the tracking loop from the effects of the spoofing signal. The efficiency of the provided strategy was evaluated using three spoofing data sets. The results of the suggested mitigation method, based on NARAX NN, show that it mitigates spoofing attacks by an average of 95.82%.
