International Journal of Sensor Networks (9 papers in press)

Regular Issues

• Improving Fault Diagnosis in Elevator Systems with GAN-Based Synthetic Data  
  by Xiaomei Lv, Zhibin Lu, Zhihao Huang, Zhanhao Wei 
  Abstract: Elevator maintenance and fault diagnosis are critical in ensuring reliable and safe operation. Elevator systems are complex electromechanical systems prone to various faults, such as sensor failures, motor malfunctions, and mechanical wear and tear. Detecting these faults promptly and accurately ensures elevators' safe and reliable operation. However, there is a lack of labelled data that may be used to train machine learning models, making it difficult to diagnose problems with elevators. This paper presents a novel approach for elevator fault diagnosis based on optimised generative adversarial networks (GANs). The proposed method employs a GAN model that generates synthetic data to augment the limited amount of labelled data and then trains a classifier on the augmented dataset. To improve the performance of the GAN, the authors introduce an optimisation algorithm that combines gradient ascent and descent, resulting in better-quality synthetic data. The efficiency of the system is evaluated using real-world elevator sensor data and compared its performance to traditional fault diagnosis methods. The results show that the proposed system can accurately diagnose faults with high accuracy and can potentially reduce maintenance costs and downtime. The proposed system provides a promising solution for elevator fault diagnosis, especially when labelled data is limited.
  Keywords: fault diagnosis; optimised generative adversarial networks; GANs; elevators; augmented dataset; and maintenance costs.
  DOI: 10.1504/IJSNET.2024.10066136
   
  
• A Road Traffic Sign Recognition Method Based on Improved YOLOv5  
  by Lu Shi, Haifei Zhang 
  Abstract: With the rapid development of artificial intelligence technology, the automatic driving of intelligent vehicles has gradually entered people's lives. The traditional vision will fail in many scenarios, such as snow, lane line wear, occlusion, or haze weather. In addition, there are still errors in the identification of traffic signs, leading to missed and erroneous detection in the intricate road network of the city. This study aims to provide an accurate and efficient method for recognising traffic signs in their natural surroundings. This paper thoroughly explores the network architecture of YOLOv5 (You Only Look Once version 5) and the ideas underpinning its loss function, considering the limitations of the existing YOLOv5-based traffic sign recognition technology. It then modifies the YOLOv5 network model to enhance its performance. According to the most recent experimental data, the enhanced YOLOv5 model performs exceptionally well at recognising traffic signs in various natural settings.
  Keywords: deep learning; object detection; traffic sign recognition; YOLOv5.
  DOI: 10.1504/IJSNET.2024.10066756
   
  
• Temperature and Humidity Monitoring and Communication System for Coal Mine Working based on LoRa  
  by Baofeng Zhao, Kaiyuan Zhu 
  Abstract: Aiming at the complexity of the underground coal mine environment and the safety of workers facing geothermal disasters, a LoRa-based communication system for real-time monitoring and control of temperature and humidity in the working environment of underground coal mines is studied and designed. Based on the introduction of the fair-access criterion of MAC protocol, the time division multiplexing protocol is adopted and improved, and the nodes complete the self-clock synchronisation after sending the information packets by redesigning the LoRa data frame structure. While adopting linear topology networking, LoRa relay nodes are adopted and designed for packet sensing, transmission and forwarding functions, and the distance arrangement of LoRa relay nodes is experimentally investigated. The experimental test proves that the system realises the expansion of the transmission distance and the improvement of the success rate of packet reception, which meets the wireless communication requirements in underground coal mines.
  Keywords: LoRa technology; wireless communication technology; coal mine environmental monitoring; Low power consumption.
  DOI: 10.1504/IJSNET.2024.10066757
   
  
• A Beamspace Channel Estimation based on Deep Convolutional Reconstruction Networks  
  by Teng Fei, Zhengyu Zhu, Jingyu Zhang, Lanxue Liu, Xinzong Yang 
  Abstract: One major challenge in millimetre-wave massive multiple-input multiple-output (MIMO) systems is achieving precise channel estimation, which still faces low accuracy and reliance on prior channel information. This paper proposes a novel beamspace channel estimation algorithm using a deep convolutional reconstruction network called DeRePixNet without requiring prior channel information. The multi-scale fusion module (MSFM) is designed to form a rich feature mapping in this network. MSFM and residual block (RB) are organically combined to prevent gradient vanishing while the network depth increases, to identify efficient local sparse structures in a convolutional visual network and replicate it spatially. The inverse transformation process from measurement vectors to the original channel is solved directly using DeRePixNet in a data-driven manner. We conducted theoretical derivations and system simulations based on the Saleh-Valenzuela channel model. The proposed DeRePixNet demonstrates superior performance compared to most existing methods. Compared to the orthogonal matching pursuit, approximate message passing learned approximate message passing, and Gaussian mixture learned approximate message passing algorithms, DeRePixNet reduces the average normalised mean squared error by approximately 11.14 dB, 8.95 dB, 1.98 dB, and 1.19 dB, respectively.
  Keywords: deep convolutional reconfiguration networks; millimetre wave; massive MIMO; channel estimation; multi-scale fusion module; MSFM.
  DOI: 10.1504/IJSNET.2024.10066759
   
  
• CNN-Based Lane-Level Positioning with Only On-Board Camera  
  by Li Chen, Liu Zhengqiong, Zhou Momiao, Sun Yanshi, Zhizhong Ding 
  Abstract: Lane-level positioning is a vital prerequisite for realising autonomous driving in complex scenarios. Existing methods for lane-level positioning mostly rely on the global positioning system (GPS) and vision-based approaches. Although the positioning accuracy of civil GPS can reach up to metre-level in an environment with good signal, it is hardly to meet the precision requirement for the lane-level vehicle's positioning in which centimeter-level precision is desired. Some traditional vision-based methods can achieve decimeter-level accuracy, they usually suffer from the weakness of low detection speed and the difficulty in handling multi-lane detection tasks. This paper proposes a one-camera low-cost approach that utilises convolutional neural network (CNN)-based segmentation for lane detection and traditional image processing techniques for lane determination. The effectiveness and robustness of the proposed approach have been tested and verified on the widely-used dataset TuSimple. It is shown that our method can achieve high detection speed while maintaining a certain detection accuracy.
  Keywords: lane-level vehicle’s positioning; lane detection; CNN; instance segmentation.
  DOI: 10.1504/IJSNET.2024.10067172
   
  
• Channel Estimation in OFDM Systems based on the Mamdani Fuzzy Genetic Algorithm  
  by Lanxue Liu, Teng Fei, Jingyu Zhang, Zhengyu Zhu, Xiaolin Wang 
  Abstract: Utilising compressive sensing technology for channel estimation can effectively enhance the spectrum efficiency of orthogonal frequency division multiplexing (OFDM) systems. However, the computational efficiency of conventional sparse channel estimation algorithms is a concern, and their performance is highly dependent on the quality of the measurement matrix and the sparsity level of the channel. Metaheuristic algorithms, currently, are among the commonly used methods for solving optimisation and search problems. Based on the principles of compressive sensing theory, this paper introduces a novel algorithm, the Mamdani fuzzy genetic algorithm (MGA), for sparse channel estimation by incorporating metaheuristic algorithms. Under basic testing conditions, the MGA algorithm can overcome the drawbacks of excessive reliance on measurement matrices, performing well, particularly in low sparsity scenarios. Experimental results indicate that, compared to classical channel estimation algorithms, the proposed algorithm is more suitable for achieving estimation accuracy with lower pilot overhead.
  Keywords: Compressed sensing; Channel estimation; Metaheuristics; Orthogonal frequency division multiplexing.
  DOI: 10.1504/IJSNET.2024.10067585
   
  
• GM-YOLOV8-Based Safety Hazard Detection Method in Power Construction  
  by Entie Qi, Jialong Ge, Liying Zhao, Hongxia Ni, Cheng Li, Dianzhi Chen, Sinan Shi 
  Abstract: In power construction settings, the operation of heavy machinery and the risk of fire present substantial hazards to the safety of transmission lines. There is an urgent need for real-time surveillance of potential safety threats during the construction process. This paper proposes an generalised multi-scale-YOLOv8 (GM-YOLOv8) hazard detection algorithm. This algorithm introduces the reparameterised generalised feature pyramid network (RepGFPN) that improving the model’s capacity to capture overarching patterns and fine-grained details. It also introduces a multi-scale cross-axis attention module (MCA). This module enhancing the network's representational capabilities and improving the detection of distant hazards. Additionally, the adoption of the Powerful-IOU loss function, which includes a non-monotonic focus mechanism, enhances the model by adaptively penalizing object size and modulating gradients based on anchor box quality. Compared to a lightweight YOLOv8 model (YOLOv8n) algorithm, GM-YOLOv8 enhances detection precision by 5.3%, accuracy by 6.8%, and recall by 6.4%, ensuring improved safety in construction environments.
  Keywords: safety hazard detection; power construction; YOLOv8; reparameterised generalised feature pyramid network; RepGFPN; multi-scale cross-axis attention module; MCA; PIOU.
  DOI: 10.1504/IJSNET.2024.10067786
   
  
• A Student Performance Prediction Model Based on Multimodal Generative Adversarial Networks  
  by Junjie Liu, Yong Yang 
  Abstract: In recent years, blended learning has been widely applied in universities, introducing complex and diverse learning data. This study aims to use machine learning algorithms to extract useful information from this data for early student performance prediction. There are still some problems in current related research, including the neglect of short text data for online learning, data imbalance, and insufficient utilisation of multimodal data. To address the mentioned issues, this study proposes an innovative solution. Firstly, adjusting the generative adversarial network generators objective function solves data imbalance in student performance prediction, and the prediction ability for minority-category students is improved. Secondly, Using short text data from online learning to map the emotions of student learning states and enhance the models accuracy and generalisation ability. Finally, this study introduces a multimodal generative adversarial network performance prediction model, which achieves the fusion of multimodal data, improves the accuracy and comprehensibility of prediction.
  Keywords: hybrid teaching; performance prediction; multimodal; Generative Adversarial Network (GAN); short text sentiment.
  DOI: 10.1504/IJSNET.2024.10067900
   
  
• A Reinforcement Learning Algorithm for Mobile Robot Path Planning with Dynamic Q-value Adjustment  
  by Chang Hua, Hao Zheng, Bao YIqin 
  Abstract: Path planning is essential for mobile robots to execute various tasks across different fields, including intelligent systems. It primarily focuses on the interaction between the agent and its environment, allowing the agent to maximise total reward by an optimal strategy. Many path-planning algorithms that are not agent-based struggle with effectively exploring entirely unknown environments. To address these issues, we propose the Adam deep Q-learning network (ADQN) to solve such problems. ADQN introduces an innovative approach to choosing action and reward functions, optimising Q-value updates dynamically based on temporal-difference error changes for enhanced model convergence and stability. Evaluated across four simulations in two maze environments of varying complexities, ADQN shows significant improvements: reduced steps, increased rewards, faster and stable loss convergence, and notably higher success rates compared to Munchausen reinforcement learning, prioritised experience replay-double duelling deep Q-networks, max-mean loss in deep Q-network algorithms in grid-based experiments.
  Keywords: Adam deep Q-learning network; ADQN; path planning; agent; reward; selection strategy; Q-value.
  DOI: 10.1504/IJSNET.2024.10068014