Hardware Limitations Research Articles

YOLO_Bolt: a lightweight network model for bolt detection

Accurate, fast, and intelligent workpiece identification is of great significance to industrial production. To cope with the limited hardware resources of factory equipment, we have made lightweight improvements based on You Only Look Once v5 (YOLOv5) and proposed a lightweight YOLO named YOLO_Bolt. First, ghost bottleneck lightweight deep convolution is added to the backbone module and neck module of the YOLOv5 detection algorithm to reduce the model volume. Second, the asymptotic feature pyramid network is added to enhance the feature utilization ability, suppress interference information, and improve detection accuracy. Finally, the relationship between the loss function and the decoupling head structure was focused on, and the number of decoupling head layers was redesigned according to different tasks to further improve the detection accuracy of the workpiece detection model. We conducted experimental verification on the MSCOCO 2017 dataset and the homemade bolt dataset. The experimental results show that compared with YOLOv5s, the number of model parameters is only 6.8 M, which is half that of the original model. On the MSCOCO 2017 dataset, the mAP increased by 2.4%. FPS increased by 104 frames/s. On the homemade dataset, the mAP 0.5 increased by 4.2%, and our proposed method is 1.2% higher than the latest YOLOv8s. The improved network can provide effective auxiliary technical support for workpiece detection.

Optimized QoS Routing in Software-Defined In-Vehicle Networks

To address the problems of low network centralized management, weak interaction, limited hardware scalability, compatibility issues and challenges in expansion in the static deployment of traditional in-vehicle networks (IVNs), a new IVN architecture is designed. At the same time, to better meet the IVN Quality of Service (QoS) requirements and improve the real-time guarantee of data transmission, the QoS routing optimization mechanism under the new architecture is established. First, a new IVN architecture including a forwarding plane, control plane and application plane is designed by introducing software defined network (SDN) technology and combining it with the IVN itself. Second, the end-to-end delay optimization model of IVN is established by introducing network calculus theory, defining system parameters, creating a network model, calculating latency, considering queuing and congestion. the traditional routing algorithm is improved, and the DBROA algorithm has been proposed, which enhances the performance of QoS routing by introducing features such as distributed routing decisions, beacon mechanisms, optimization algorithms, and adaptability. This improvement allows it to better meet the QoS requirements of various applications and services, thereby enhancing existing QoS routing algorithms. Finally, an IVN routing optimization system is built and implemented, and the performance of different algorithms is compared and analyzed. The experimental results show that compared with the traditional Dijkstra and ECMP algorithms, the DBROA algorithm can effectively reduce the data forwarding delay and packet loss rate, improve the overall performance of IVN, and provide better QoS guarantees for IVN real-time data transmission.

Space-Specific Mixing Excitation for High-SNR Spatial Encoding in Magnetic Particle Imaging.

Magnetic Particle Imaging (MPI) is a radiation-free tracer-based imaging technology that visualizes the spatial distribution of superparamagnetic iron oxide nanoparticles. Conventional spatial encoding methods in MPI rely on a gradient magnetic field with a constant gradient strength to generate a field-free point or line for spatial scanning. However, increasing the gradient strength can enhance theoretical spatial resolution but also lead to a decrease in the Signal-to-Noise Ratio (SNR) and sensitivity of the imaging system. This poses a technical challenge in balancing spatial resolution and sensitivity, necessitating intricate hardware design. To address this, we present a Space-Specific Mixing Excitation (SSME) technique for achieving high-SNR spatial encoding in MPI. By utilizing a dual-frequency excitation magnetic field with a non-homogeneous field strength, magnetic particles at each position generate unique intermodulation responses. By performing multi-channel acquisitions across the entire field of view, high SNR MPI signals can be acquired. When combined with reconstruction techniques based on system matrix, multi-dimensional SSME-MPI can be achieved. The effectiveness of the proposed method was validated through phantom and in vivo imaging experiments. The results demonstrate significant improvements in sensitivity (3.6-fold improvement) and spatial resolution (better than 1 mm) without any hardware modifications. These findings demonstrate the capability of SSME to enhance both the spatial resolution and sensitivity of MPI. This method provides a solution to the ongoing challenge of balancing spatial resolution and sensitivity in MPI, potentially facilitating the implementation of MPI in a wider range of medical applications.

Exploring the Spectral Prior for Hyperspectral Image Super-Resolution.

In recent years, many single hyperspectral image super-resolution methods have emerged to enhance the spatial resolution of hyperspectral images without hardware modification. However, existing methods typically face two significant challenges. First, they struggle to handle the high-dimensional nature of hyperspectral data, which often results in high computational complexity and inefficient information utilization. Second, they have not fully leveraged the abundant spectral information in hyperspectral images. To address these challenges, we propose a novel hyperspectral super-resolution network named SNLSR, which transfers the super-resolution problem into the abundance domain. Our SNLSR leverages a spatial preserve decomposition network to estimate the abundance representations of the input hyperspectral image. Notably, the network acknowledges and utilizes the commonly overlooked spatial correlations of hyperspectral images, leading to better reconstruction performance. Then, the estimated low-resolution abundance is super-resolved through a spatial spectral attention network, where the informative features from both spatial and spectral domains are fully exploited. Considering that the hyperspectral image is highly spectrally correlated, we customize a spectral-wise non-local attention module to mine similar pixels along spectral dimension for high-frequency detail recovery. Extensive experiments demonstrate the superiority of our method over other state-of-the-art methods both visually and metrically. Our code is publicly available at https://github.com/HuQ1an/SNLSR.

Control system for suction valves within gas compressors skids located in potentially explosive zones

The paper presents the implementation of a feedback loop controller for the suction valves used within the gas compressors skids manufactured by Romanian Research and Development Institute for Gas Turbines COMOTI. These compressors are installed at beneficiaries, in potentially explosive environments (e.g., gas extraction and transportation applications). The proposed solution addresses the problems observed due to the high suction pressures of screw compressors, which may occur at start-up or during operation, leading to failed start-ups or emergency shutdowns. An automatic sequence implemented in the control software and displayed on the operating panel, could assist the operator during high suction pressures (above 1.2 bar). These pressure conditions normally lead to alarm limits and to performing an automatic shutdown sequence for protecting the compressor. The paper herein discusses the implementation of a start-up sequence in the programmable logic controller’s software, which can be applied without critical interventions on the compressor packages, only through minor control software modifications which can safely extend the inlet pressure range and the exploitation of the compressors. Thus, since no hardware modifications are required, we also ensure maintaining the ATEX certifications of equipment and assemblies.

Super resolution label-free dark-field microscopy by deep learning.

Dark-field microscopy (DFM) is a powerful label-free and high-contrast imaging technique due to its ability to reveal features of transparent specimens with inhomogeneities. However, owing to the Abbe's diffraction limit, fine structures at sub-wavelength scale are difficult to resolve. In this work, we report a single image super resolution DFM scheme using a convolutional neural network (CNN). A U-net based CNN is trained with a dataset which is numerically simulated based on the forward physical model of the DFM. The forward physical model described by the parameters of the imaging setup connects the object ground truths and dark field images. With the trained network, we demonstrate super resolution dark field imaging of various test samples with twice resolution improvement. Our technique illustrates a promising deep learning approach to double the resolution of DFM without any hardware modification.

Towards Autonomous Control System in Brazilian Navy's USV-Lab using MOOS-IvP framework

This article aims to describe the implementation of an autonomous control system for the Brazilian Navy's Unmanned Surface Vehicle Laboratory (USV-Lab) using the MOOS-IvP framework. The main goal of this project is to enable an existing vehicle to be a laboratory for research of maritime autonomous systems, to perform unmanned maritime missions, following desired routes without human intervention or being remotely piloted. The paper starts with the importance of autonomous maritime vehicles and presents MOOS-IvP framework as a validated tool for autonomous maritime systems development. In succession, the main objectives are set, the physical system and components are shown and the proposed software and communication are structured. The work comprises the development of control interface, the integration of sensors and actuators, the tuning PID controllers, and specific software settings, along with several hardware modifications. The methodology consisted in progressive testing, from simulations to local operations as made in Guanabara Bay in Rio de Janeiro and Bay of All Saints in Bahia. Results revealed the system's effectiveness and robustness in maintaining configured routes, enabling its use in a Brazilian Navy military reconnaissance exercise, as well as serving as a laboratory for future research.

Enhancing Cybersecurity Protocols in Financial Networks through Reinforcement Learning

Cybersecurity in financial networks is facing an unprecedented level of sophistication from cyber threats, necessitating the adoption of advanced technologies to safeguard sensitive financial data. This review paper explores the integration of Reinforcement Learning (RL), Quantum Computing (QC), and Data Science (DS) to enhance cybersecurity protocols in financial networks. RL offers promising solutions for automating threat detection, intrusion prevention, and response systems by leveraging adaptive learning techniques. QC introduces powerful computational capabilities to both strengthen encryption methods and challenge traditional cryptographic systems, while DS provides data-driven insights for predictive analytics and real-time anomaly detection. By examining the application of these technologies individually and in tandem, this paper highlights their potential to transform financial cybersecurity. We discuss existing case studies and research developments, focusing on their contributions to threat intelligence, encryption, and network defense. The paper also identifies the key challenges associated with implementing RL, QC, and DS, including scalability, hardware limitations, and integration complexities. In conclusion, we provide insights into future research directions aimed at addressing these challenges, presenting a roadmap for fully integrating RL, QC, and DS into financial cybersecurity frameworks. This comprehensive review underscores the critical role these technologies will play in safeguarding financial systems against emerging cyber threats.

Assessment of lattice Boltzmann method for low-rise building wind flow simulation with limited resources

The study of low-rise buildings' interaction with wind is gravitating toward large eddy simulation (LES) for simulating airflow. As it is straightforward to implement and also offers the option to simulate particles, sound, and heat flow, the lattice Boltzmann method (LBM) may be an interesting option among the LES methods but has so far barely been explored for this application. In this work, the LBM is investigated and assessed as a tool for the analysis of wind flow around low-rise buildings. The hardware resources used are limited to one GPU to mimic the limited resources an industrial setting or a preliminary study might allow for. We suggest an efficient LBM simulation setup and compare its results from the wind flow around nine exemplary gable roofed low-rise buildings to findings of previous studies, both in detail and in terms of found behaviors. All comparisons show good agreement between our findings and the previous results obtained via other simulation methods or wind tunnel measurements. These results confirm that the LBM may be used to investigate low-rise buildings, even with limited hardware. It is thus a further contender among the available LES methods and our setup may serve as starting point for its application.

Developing a constraint model for using artificial intelligence on existing, limited hardware in manufacturing machines

In the rapidly evolving field of Artificial Intelligence (AI), its application in industrial settings, particularly for anomaly detection in time series data, poses unique challenges. Current methods often lack a comprehensive understanding of the trade-offs involved in achieving optimal model performance, data preparation effort, and prediction quality. To bridge this gap, this paper presents an adaptive approach to address these challenges, focusing on making conscious decisions about mentioned trade-offs. Inspired by the principles of the Iron Triangle from product engineering, our methodology introduces a novel ”AI triangle” with dimensions of Speed, Effort, and Quality. We applied this methodology to a real-world case study involving anomaly detection in a constrained hardware environment in the context of a forming production process. The results highlight the effectiveness of our approach in achieving a practical balance between speed, effort, and quality constraints for implementing AI in an industrial setting. This paper provides valuable insights into the considerations and trade-offs necessary for the successful deployment of AI in manufacturing and other similar industrial applications.

Industrial collaborative environments integrating AI, Big Data and Robotics for smart manufacturing

In recent years, global events have posed significant challenges to manufacturing firms’ business models, necessitating adaptable production systems that integrate human and automated resources seamlessly. Fortunately, technological advancements over the past decade have facilitated flexible production solutions. This paper introduces the approach of the CONVERGING EU project for collaborative smart manufacturing systems aiming at increasing flexibility, efficiency, and operators’ satisfaction. The proposed approach is structured into three technical pillars: perception, adaptation, collaboration, enhanced by modules for fostering user experience. Perception involves identifying and recognizing features of parts, resources, and surroundings to infer and analyse their status, and formulate action plans. Automated adaptations through hardware and control system modifications allow for executing formulated plans while complying with user characteristics and needs. Collaboration focuses on the software and hardware interfaces to ensure safe and seamless interaction with collaborative robotic solutions. Training, ergonomics tracking, and adaptable interfaces are also proposed to pave the way toward social industrial environments. This approach is applied to four industrial scenarios from different manufacturing sectors: automotive, white goods, aeronautics and additive manufacturing. The design of systems incorporating AI and robotics is presented together with the expected impact including safer, more intuitive, and trustful human-robot interaction, but also more inclusive industrial environments.

Ultrasonic activated energy-efficient autonomous camera system for activity observation

This Paper presents a proof-of-concept system for autonomous object detection and tracking, employing a pan-tilt mechanism and ultrasonic sensors for energy-efficient operation. The system’s design aims to reduce unnecessary power consumption by activating energy-intensive components, such as the camera and LED lights, only when nearby objects are detected. Key functionalities, including distance sensing and object detection, are evaluated across multiple tests to assess system performance. While the system demonstrated consistent distance sensing and power savings, object detection results were inconsistent, primarily due to the limited training dataset and hardware constraints. Recommendations for future improvements include expanding the dataset, exploring advanced object detection models, and upgrading the camera to enhance detection consistency and overall performance. This prototype serves as a stepping stone toward developing robust, autonomous underwater photography systems for marine research.

Reconfigurable Manufacturing Systems through Intelligent Workpiece Handling and Artificial Intelligence

In recent years, businesses have strived to align their manufacturing goods with the unique needs and preferences of their customers. However, the mass customization paradigm introduces challenges mostly associated to the manipulation of small product batches and knowledge management. This paper proposes a framework for intelligent reconfigurable manufacturing systems, that leverages robotic automation supported by a collaborative dual-arm manipulator, cognitive mechatronic devices, and machine-learning algorithms. In more detail, pioneer end-effectors and machinery provide the capacity of grasping and handling a variety of products without time-consuming hardware modifications. The backbone of this manipulation scheme is a set of machine learning tools for: a) the planning and synchronization of handling agents, b) identification and localization of workpieces, c) the generation of optimal packaging configurations based on customer orders and identified workpieces, and d) the allocation of the respective handling operations to robots, operators, and machinery. The proposed framework is evaluated through its implementation on a use case deriving from the metal industry. The particularities of this use case, as expressed in product variety, workpiece complexity, customer orders and intralogistics, provide a fertile ground for the validation of the discussed technologies.

Investigation of alternative attention modules in transformer models for remaining useful life predictions: addressing challenges in high-frequency time-series data

Predicting the Remaining Useful Life of a machine components or systems is a pivotal technology in condition-based maintenance and essential for ensuring the reliability and safety in various production-engineering applications. The influx of extensive industrial data has notably enhanced the efficacy of data-driven Remaining Useful Life prediction models, especially deep learning models. One of the promising deep learning model architectures is the Transformer-based model with the Self-Attention mechanism at its core. However, inherent limitation arise when applying Self-Attention to high- frequency time-series data with large window sizes. Due to its high computational complexity, hardware limitations hinder the practical implementation of Transformer with Self-Attention in production-engineering applications. This study looks into the utilization of alternative Attention modules with reduced complexity, making it more applicable to high-frequency time-series data. In order to allow comparability, this study uses the well-known C-MPASS dataset for benchmarking Remaining Useful Life approaches. Although this dataset does not consist of high-frequency data, it demonstrates the usefulness of alternative Attention modules without noteworthy losses in model accuracy.

Sensing nature's alarm: SnO2/MXene gas sensor unveils methyl jasmonate signatures of plant insect stress.

The incorporation of artificial intelligence into agriculture presents challenges, particularly due to hardware limitations, especially in sensors. Currently, pest detection relies heavily on manual scouting by humans. Therefore, the objective of this study is to create a chemoresistive sensor that enables early identification of the characteristic volatile compound, viz., methyl jasmonate, released during pest infestations. Given the lower reactivity of esters, we have fine-tuned a composite consisting of SnO2 nanoparticles and 2D-MXene sheets to enhance adsorption and selective oxidation, resulting in heightened sensitivity. The optimized composite demonstrated a notable response even at concentrations as low as 120 ppb, successfully confirming pest infestations in tomato crops.

Job CPU Performance comparison based on MINIAOD reading options: Local versus remote

Job CPU Performance comparison based on MINIAOD reading options: Local versus remote

Evaluation of teaching effect of online education courses based on random forest

At present, online education platforms are very popular. However, these platforms also have some drawbacks, including that they cannot be completely based on the needs of students, many students ignore the needs of students in order to adapt to the platform, do not adopt a more personalized learning style, and do not fully realize the individualized teaching. There are many factors at play, such as shifts in learning and teaching patterns, online platforms, hardware limitations of the network, and more. However, whether this model can finally satisfy teachers and students, and how the teaching effect is, has become a problem of concern. We propose a random forest-based method to evaluate college online courses. This method not only focuses on a single influencing factor, but also takes into account the special needs of students and teachers, and formulates different evaluation criteria, so that the system has not only universal applicability. The experimental results show that this method can improve the proficiency and level of teachers' online teaching, and can effectively combine several courses taught by students to form a personalized teaching method.

Upgrading the neutronic version of the PSI cycle check-up methodology for full core PWR Monte Carlo simulations

Full core Monte Carlo (MC) simulations are of high interest and can be used to study the beginning of live conditions of reactors. However, current hardware limitations make it difficult to perform such studies for other cycles, as the whole irradiation history of the fuel assemblies (FAs) loaded must be considered on a time scale of several years using MC depletion capabilities. Therefore, at PSI, a cycle check-up methodology (CHUP) is currently being developed to address this challenge. By extracting operating conditions and isotopic compositions from validated reference deterministic core-follow models (using CASMO5/SIMULATE5), MC neutron transport models are generated for selected operating points. These models can then be used to estimate quantities beyond the capabilities of deterministic solvers. This article presents the verification performed on a hot zero power (HZP) operating condition of a Swiss PWR using the newly updated CHUP methodology. This update consists of extracting information generated automatically from reference core models, minimising as much human intervention as possible. The performance of the method was first evaluated against the reference deterministic lattice code. Then, the memory requirements of the full core simulation were studied under different approximations. Finally, verification was performed against the nodal reference scheme.

MCMC: Multi-Constrained Model Compression via One-Stage Envelope Reinforcement Learning.

Model compression methods are being developed to bridge the gap between the massive scale of neural networks and the limited hardware resources on edge devices. Since most real-world applications deployed on resource-limited hardware platforms typically have multiple hardware constraints simultaneously, most existing model compression approaches that only consider optimizing one single hardware objective are ineffective. In this article, we propose an automated pruning method called multi-constrained model compression (MCMC) that allows for the optimization of multiple hardware targets, such as latency, floating point operations (FLOPs), and memory usage, while minimizing the impact on accuracy. Specifically, we propose an improved multi-objective reinforcement learning (MORL) algorithm, the one-stage envelope deep deterministic policy gradient (DDPG) algorithm, to determine the pruning strategy for neural networks. Our improved one-stage envelope DDPG algorithm reduces exploration time and offers greater flexibility in adjusting target priorities, enhancing its suitability for pruning tasks. For instance, on the visual geometry group (VGG)-16 network, our method achieved an 80% reduction in FLOPs, a 2.31× reduction in memory usage, and a 1.92× acceleration, with an accuracy improvement of 0.09% compared with the baseline. For larger datasets, such as ImageNet, we reduced FLOPs by 50% for MobileNet-V1, resulting in a 4.7× faster speed and 1.48× memory compression, while maintaining the same accuracy. When applied to edge devices, such as JETSON XAVIER NX, our method resulted in a 71% reduction in FLOPs for MobileNet-V1, leading to a 1.63× faster speed, 1.64× memory compression, and an accuracy improvement.

CSD: An Online Spectral Sensing Method for Wastewater Quality Monitoring Based on Compressed Sensing and Incremental Learning

The outstanding spectral resolution of full-spectrum detection (FSD) ensures excellent performance for complex scenarios in which target signals are affected by particle interference. However, FSD is limited by the cost of its measurement devices and demanding install requirements. Therefore, it is not always practical to utilize FSD for online water quality monitoring in wastewater treatment plants (WWTPs). In this study, we developed a novel detection approach entitled “compressive spectrum detection (CSD)” to overcome the hardware restrictions posed by FSD. An intelligent source-encoding laser is first developed to perform spectrum encoding, which avoids the utilization of sophisticated equipment and reduces the detection cost. Subsequently, a dual-stage compressive detection framework was proposed based on incremental learning to accurately reconstruct spectral signals with full-spectrum resolution, which consisted of three innovative algorithms and could solve difficulties brought about by sample insufficiency and changing environments in practical situations. Based on our constructed online detection prototype, the superiority of CSD was evaluated through practical application in which the training samples were reduced to a quarter of their original size with an overall improvement in detection precision greater than 50%.