Learning-based Solution Research Articles

Smart Markets for Real-Time Allocation of Multiproduct Resources: The Case of Shared Electric Vehicles

Motivated by the deep transformation of the transportation sector toward electric and shared mobility, this research develops a decision support system that enables real-time allocation of shared electric vehicles (SEVs) to either the rental market or aggregated as a virtual power plant to the electricity market. The multipurposing of SEVs beyond the traditional scope of mobility services can boost utilization and profitability for fleet operators while also facilitating the integration of variable rate renewable energy sources. This SEV setting embodies the problem of real-time multiproduct resource allocation via smart markets. We develop an approach that leverages highly digitalized real-time markets and intelligent decision making to enable dynamic multiproduct resource allocation. The tool continuously evaluates market attractiveness and strategically places iterative bids on the resource owner’s behalf until all resources have been committed for a specific period. Our key contribution is to provide a blueprint for similar multi-product resource allocation settings, and demonstrating its efficacy in the domain of SEV fleets by developing tailored analytics and machine learning-based solutions.

Learning Observation Model for Factor Graph Based-State Estimation Using Intrinsic Sensors

The navigation system of autonomous mobile robots has appeared challenging when using exteroceptive sensors such as cameras, LiDARs, and radars in textureless and structureless environments. This paper presents a robust state estimation system for holonomic mobile robots using intrinsic sensors based on adaptive factor graph optimization in the degradation scenarios. In particular, the neural networks are employed to learn the observation and noise model using only IMU sensor and wheel encoder data. Investigating the learning model for the holonomic mobile robot is discussed with various neural network architectures. We also explore the neural networks that are far more powerful and have cheaper computing power when using the inertial-wheel encoder sensors. Furthermore, we employ an industrial holonomic robot platform equipped with multiple LiDARs, cameras, IMU, and wheel encoders to conduct the experiments and create the ground truth without a bulky motion capture system. The collected datasets are then implemented to train the neural networks. Finally, the experimental evaluation presents that our solution provides better accuracy and real-time performance than other solutions. <i>Note to Practitioners</i>&#x2014;Autonomous mobile robots need to serve in challenging environments robustly that deny extrinsic sensors such as cameras, LiDARs, and radars. In order to operate in this situation, the navigation system shall rely on the intrinsic sensor as inertial sensor and wheel encoders. Existing conventional methods have combined the intrinsic sensors in the form of the recursive Bayesian filtering technique without adapting these models. Besides, deep learning-based solutions have adopted extensive networks like LSTM or CNN to handle the estimation problem. This work aims to develop a state estimation subsystem of the navigation system for the holonomic mobile robots that utilize the intrinsic sensors in adaptive factor graph optimization. In particular, we present how to join the factor graph efficiently with the learning observation model of IMU and wheel encoder factor. Moreover, the neural networks are introduced to learn the observation model with an IMU and wheel encoder data inputs. We recognize that lightweight neural networks can achieve more power than deep learning techniques using the IMU sensor and wheel encoders. Finally, the neural networks are embedded in a factor graph to handle the smoothing state estimation. The proposed system could operate with high accuracy in real time.

HardGBM: A Framework for Accurate and Hardware-Efficient Gradient Boosting Machines

Gradient boosting machine (GBM) is a powerful and widely used type of ensemble machine learning methods, among which the most famous one is XGBoost. However, the cost of running large GBMs on hardware could become prohibitive given stringent resources. Ensemble reduction on boosting ensemble is intrinsically hard because member models are constructed in a sequential order, where the training targets for latter ones depend on the performance of the former ones. In this work, a GBM reduction framework is proposed for the first time to tackle the problem. For the first time, the framework supports automatic hardware implementation of regression tree ensembles. Experiments on 24 datasets from various applications demonstrate that our method reduces overall area utilization by 81.60% (80.64%) and power consumption by 21.15% (19.06%), while exceeding or successfully maintaining the performance level comparing with the original XGBoost (LightGBM) ensembles. In comparative experiments, to attain approximately the same accuracy level as our framework or XGBoost, deep learning-based solutions require >52.7× footprints, 6.0× power consumption, and 1.4× training time. Equipped with tunable parameters, the framework is expected to seek a Pareto optimal front considering hardware resource limitation, accuracy and stability, and computation (training) efficiency.

ОЦЕНКА ТЕХНИЧЕСКОГО СОСТОЯНИЯ ГАЗОТУРБИННОЙ УСТАНОВКИ ПРИ ПОМОЩИ МЕТОДОВ МАШИННОГО ОБУЧЕНИЯ ПО ИСКУССТВЕННЫМ ДАННЫМ

Continuous monitoring of the technical condition of gas turbines, defect identification, failure prevention, and optimization of operation, maintenance, and repair processes are relevant tasks for the operators of this equipment. Various machine learning methods that are already being used in the field of gas turbines can help solve these tasks. The limiting factor in this regard is the lack of real operational data. This study examines the possibility of using synthetic data for training and testing machine learning models to determine the level of technical condition of a gas turbine installation. An open dataset created by other researchers using a mathematical model of a marine gas turbine engine was selected for analysis. The research presents the accuracy values obtained by different methods of evaluating machine learning models. The random forest model demonstrated the best results. It was found that when developing machine learning-based solutions for engineering tasks, additional methods for assessing the accuracy of predictions are required. The further development of this work is associated with the development of a proprietary mathematical model of a gas turbine installation capable of considering the influence of specific defects to create datasets for analysis and further research

Machine Learning Enabled Solutions for Design and Optimization Challenges in Networks-on-Chip based Multi/Many-Core Architectures

Due to the advancement of transistor technology, a single chip processor can now have hundreds of cores. Network-on-Chip (NoC) has been the superior interconnect fabric for multi/many-core on-chip systems because of its scalability and parallelism. Due to the rise of dark silicon with the end of Dennard Scaling, it becomes essential to design energy efficient and high performance heterogeneous NoC-based multi/many-core architectures. Because of the large and complex design space, the solution space becomes difficult to explore within a reasonable time for optimal trade-offs of energy-performance-reliability. Furthermore, reactive resource management is not effective in preventing problems from happening in adaptive systems. Therefore, in this work, we explore machine learning techniques to design and configure the NoC resources based on the learning of the system and applications workloads. Machine learning can automatically learn from past experiences and guide the NoC intelligently to achieve its objective on performance, power, and reliability. We present the challenges of NoC design and resource management and propose a generalized machine learning framework to uncover near-optimal solutions quickly. We propose and implement a NoC design and optimization solution enabled by neural networks, using the generalized machine learning framework. Simulation results demonstrated that the proposed neural networks-based design and optimization solution improves performance by 15% and reduces energy consumption by 6% compared to an existing non-machine learning-based solution while the proposed solution improves NoC latency and throughput compared to two existing machine learning-based NoC optimization solutions. The challenges of machine learning technique adaptation in multi/many-core NoC have been presented to guide future research.

DeepSHARQ: hybrid error coding using deep learning

Cyber-physical systems operate under changing environments and on resource-constrained devices. Communication in these environments must use hybrid error coding, as pure pro- or reactive schemes cannot always fulfill application demands or have suboptimal performance. However, finding optimal coding configurations that fulfill application constraints—e.g., tolerate loss and delay—under changing channel conditions is a computationally challenging task. Recently, the systems community has started addressing these sorts of problems using hybrid decomposed solutions, i.e., algorithmic approaches for well-understood formalized parts of the problem and learning-based approaches for parts that must be estimated (either for reasons of uncertainty or computational intractability). For DeepSHARQ, we revisit our own recent work and limit the learning problem to block length prediction, the major contributor to inference time (and its variation) when searching for hybrid error coding configurations. The remaining parameters are found algorithmically, and hence we make individual contributions with respect to finding close-to-optimal coding configurations in both of these areas—combining them into a hybrid solution. DeepSHARQ applies block length regularization in order to reduce the neural networks in comparison to purely learning-based solutions. The hybrid solution is nearly optimal concerning the channel efficiency of coding configurations it generates, as it is trained so deviations from the optimum are upper bound by a configurable percentage. In addition, DeepSHARQ is capable of reacting to channel changes in real time, thereby enabling cyber-physical systems even on resource-constrained platforms. Tightly integrating algorithmic and learning-based approaches allows DeepSHARQ to react to channel changes faster and with a more predictable time than solutions that rely only on either of the two approaches.

Source number of single-channel signals intelligent estimation via signal reconstruction

As wireless communication advances, the spectrum is becoming more crowded and overlapping, making it essential to determine the number of signal sources based on the received signal in signal processing. In various communication scenarios, single-channel reception is preferred over multi-channel reception due to its simplicity, reliability, and cost-effectiveness. However, most existing methods for estimating the number of sources in single-channel systems are adapted from multi-channel methods and require high signal-to-noise ratios (SNRs). To overcome this limitation, we explore a deep learning-based solution for estimating the number of sources in single-channel systems. Our method utilizes a signal reconstruction algorithm based on low-rank representation to recover the essential structure of the signal and suppress noise. The reconstructed signal is then transformed into a radio image, taking into account the correlation and feature completeness of the information. Moreover, we develop a lightweight hybrid classification (LHC) model capable of estimating source number by using small sample images, without the need for extensive data. Our experimental results demonstrate the effectiveness of this method in estimating the number of sources under low SNR and small samples conditions.

Exploiting multi-granularity visual features for retinal layer segmentation in human eyes.

Accurate segmentation of retinal layer boundaries can facilitate the detection of patients with early ophthalmic disease. Typical segmentation algorithms operate at low resolutions without fully exploiting multi-granularity visual features. Moreover, several related studies do not release their datasets that are key for the research on deep learning-based solutions. We propose a novel end-to-end retinal layer segmentation network based on ConvNeXt, which can retain more feature map details by using a new depth-efficient attention module and multi-scale structures. In addition, we provide a semantic segmentation dataset containing 206 retinal images of healthy human eyes (named NR206 dataset), which is easy to use as it does not require any additional transcoding processing. We experimentally show that our segmentation approach outperforms state-of-the-art approaches on this new dataset, achieving, on average, a Dice score of 91.3% and mIoU of 84.4%. Moreover, our approach achieves state-of-the-art performance on a glaucoma dataset and a diabetic macular edema (DME) dataset, showing that our model is also suitable for other applications. We will make our source code and the NR206 dataset publicly available at (https://github.com/Medical-Image-Analysis/Retinal-layer-segmentation).

Nonlinear receding-horizon filter approximation with neural networks for fast state of charge estimation of lithium-ion batteries

An important component of electric vehicles is the Battery Management System (BMS), whose main objective is to monitor the state of charge (SoC) of the battery, which can significantly impact the vehicle’s autonomy. The SoC may be estimated using filtering algorithms; in this context, higher accuracy and computational complexity are of great importance. The present paper aims to propose receding-horizon strategies, namely Moving-Horizon State Estimation (MHSE) and Neural Network Moving-Horizon Estimation (NNMHE), for SoC estimation. MHSE is based on a constrained optimization problem, with information of measured samples along a larger observation window, which ensures high accuracy and robustness but requires a better processing capacity. Simulated results are obtained through this method, and it is demonstrated its capacity to jointly estimate the states and the unknown lumped parameters of the battery model using an augmented states formulation. The accuracy of MHSE in the process is high enough that its results may be used for training the NNMHE, so that a machine learning-based solution, with reduced processing time, is found. In this proposed method, a Neural Network is used to emulate the optimization problem solver, by which faster and approximate results are obtained. This approach is evaluated with an experimental dataset, achieving a coefficient of determination of almost 99% and about 20 times faster, which proves that it is effective and can be readily employed in an embedded systems application requiring less computational resources.

Enabling Proportionally-Fair Mobility Management With Reinforcement Learning in 5G Networks

Mobility management in 5G is challenging, and at higher frequencies, a larger number of cells is needed to provide similar coverage to that in 4G. Consequently, Base Stations (BSs) are placed much more densely and users experience frequent handovers, reducing network capacity. Advanced handover techniques are needed in 5G to perform smooth network operation. In this paper, we formulate an optimization problem, whose goal is to strive for fairness in data rates among users and to reduce handovers. To accomplish that, we consider jointly the decisions when to handover and to which BS a user is to be assigned. This is an integer nonlinear program, and by relaxing it, we obtain an upper bound. Further, due to its NP-hardness, we propose a centralized and a multi-agent Deep Q Network (DQN)-based algorithm, which both find near-optimal user-to-BS assignments. We evaluate our Reinforcement Learning-based solutions for networks of different sizes and users with different velocities. We compare our approaches with baselines and show that they outperform them considerably in terms of fairness and radio link failures while being within 95% of the optimum. Our DQN algorithms also reduce the handover rate by up to 93% and avoid ping-pong handovers almost completely.

Drone Detection and Surveillance

This paper presents a deep learning-based solution for detecting and identifying drones to address security concerns. By leveraging A9G-based GPS tracker modules, the system achieves real-time tracking and differentiation between authorized and unauthorized drones. The proposed model demonstrates high performance with an F-1 score of 93.63%, precision of 96.2%, mAp of 95%, recall of 91.2%, and accuracy of 95.5%. Authorized drone locations can be communicated via SMS for prompt action if an unauthorized drone is detected.

Real-Time Phishing Website Detection using Machine Learning and Updating Phishing Probability with User Feedback

Phishing attacks remain a significant threat to internet users worldwide. Cybercriminals often send out phishing links through various channels such as emails, social media platforms, or text messages, to trick users into disclosing their sensitive infor-mation such as passwords, usernames, or credit card details. This stolen information is then used to perpetrate various types of fraud or sold on the dark web for profit. To combat this problem, various machine learning-based solutions have been developed for detect-ing phishing websites. However, these solutions vary in their effec-tiveness, with some focusing on URL-based algorithms while oth-ers focus on website content. This paper proposes a machine learn-ing-based approach to real-time phishing website detection, with a focus on the website's URL, domain page, and content. The pro-posed framework will be implemented as a browser plug-in, which can identify phishing risks as users visit websites. The framework integrates several techniques, including blacklist interception, whitelist filtering, and machine learning prediction, to improve ac-curacy, reduce false alarm rates, and minimize computation times. The proposed approach also incorporates user feedback to update the phishing probability over time, thereby increasing the accuracy of detecting phishing websites. This feedback loop involves users reporting suspected phishing websites to the system, which then updates the phishing probability calculation with new information to improve its accuracy. The significance of this research lies in its ability to provide real-time phishing detection capabilities, which can help protect internet users from falling victim to phishing at-tacks. Furthermore, the use of machine learning-based algorithms and user feedback ensures that the system is continuously updated to remain effective against new and emerging threats.

AutoCTS+: Joint Neural Architecture and Hyperparameter Search for Correlated Time Series Forecasting

Sensors in cyber-physical systems often capture interconnected processes and thus emit correlated time series (CTS), the forecasting of which enables important applications. The key to successful CTS forecasting is to uncover the temporal dynamics of time series and the spatial correlations among time series. Deep learning-based solutions exhibit impressive performance at discerning these aspects. In particular, automated CTS forecasting, where the design of an optimal deep learning architecture is automated, enables forecasting accuracy that surpasses what has been achieved by manual approaches. However, automated CTS solutions remain in their infancy and are only able to find optimal architectures for predefined hyperparameters and scale poorly to large-scale CTS. To overcome these limitations, we propose AutoCTS+, a joint, scalable framework, to automatically devise effective CTS forecasting models. Specifically, we encode each candidate architecture and accompanying hyperparameters into a joint graph representation. We introduce an efficient Architecture-Hyperparameter Comparator (AHC) to rank all architecture-hyperparameter pairs, and we then further evaluate the top-ranked pairs to select an architecture-hyperparameter pair as the final model. Extensive experiments on six benchmark datasets demonstrate that AutoCTS+ not only eliminates manual efforts but also is capable of better performance than manually designed and existing automatically designed CTS models. In addition, it shows excellent scalability to large CTS.

Privacy-preserving malware detection in Android-based IoT devices through federated Markov chains

The continuous emergence of new and sophisticated malware specifically targeting Android-based Internet of Things devices is causing significant security hazards and is consequently fostering the need for effective detection models and strategies able to work with these hardware-constrained devices. In addition, since such models are often trained on confidential application data, many involved subjects are reluctant to share their data for this purpose. Accordingly, several Federated Learning-based solutions are emerging, which rely on the capabilities of Machine Learning models in malware detection/classification without sharing user data. However, Federated Learning methods are often adversely affected by non-independent and identically distributed data in terms of both the required training time and classification results. Therefore, a promising solution could be to overcome the Federated Learning-related issues by preserving the privacy of end-user data. In this direction, the capabilities of Markov chains and associative rules are extended within a federated environment to face malware classification tasks in the IoT scenario. The presented approach, evaluated on several malware families, has achieved an average accuracy of 99% in the presence of centralized and decentralized unbalanced training/testing data by overcoming the most common state-of-the-art approaches. Also, its runtime performance is comparable with centralized ones by considering several non-independent and identically distributed dataset partitions, splitting criteria, and clients, respectively.

Machine learning-based solution reveals cuproptosis features in inflammatory bowel disease.

Cuproptosis, a new cell death mode, is majorly modulated by mitochondrial metabolism and protein lipoylation. Nonetheless, cuproptosis-related genes (CRGs) have not yet been thoroughly studied for their clinical significance and relationship with the immune microenvironment in inflammatory bowel disease (IBD). We screened CRGs that had a significant correlation with immune status, which was determined utilizing single-sample GSEA (ssGSEA) and Gene Expression Omnibus datasets (GSE75214). Furthermore, utilizing the R package "CensusClusterPlus", these CRGs' expression was used to obtain different patient clusters. Subsequently, gene-set enrichment analysis (GSEA), gene set variation analysis (GSVA), and CIBERSORT assessed the variations in the enrichment of gene function and the abundance of immune cell infiltration and immune functions across these clusters. Additionally, weighted gene co-expression network analysis (WGCNA) and analysis of differentially expressed genes (DEGs) were executed, and for the purpose of identifying hub genes between these clusters, the construction of protein-protein interaction (PPI) network was done. Lastly, we used the GSE36807 and GSE10616 datasets as external validation cohorts to validate the immune profiles linked to the expression of CRG. ScRNA-seq profiling was then carried out using the publicly available dataset to examine the CRGs expression in various cell clusters and under various conditions. Three CRGs, PDHA1, DLD, and FDX1, had a significant association with different immune profiles in IBD. Patients were subsequently classified into two clusters: low expression levels of DLD and PDHA1, and high expression levels of FDX1 were observed in Cluster 1 compared to Cluster 2. According to GSEA, Cluster 2 had a close association with the RNA processes and protein synthesis whereas Cluster 1 was substantially linked to environmental stress response and metabolism regulations. Furthermore, Cluster 2 had more immune cell types, which were characterized by abundant memory B cells, CD4+ T memory activated cells, and follicular helper T cells, and higher levels of immune-related molecules (CD44, CD276,CTLA4 and ICOS) than Cluster 1. During the analysis, the PPI network was divided into three significant MCODEs using the Molecular Complex Detection (MCODE) algorithm. The three MCODEs containing four genes respectively were linked to mitochondrial metabolism, cell development, ion and amino acid transport. Finally, external validation cohorts validated these findings, and scRNA-seq profiling demonstrated diverse intestinal cellular compositions with a wide variation in CRGs expression in the gut of IBD patients. Cuproptosis has been implicated in IBD, with PDHA1, DLD, and FDX1 having the potential as immune biomarkers and therapeutic targets. These results offer a better understanding of the development of precise, dependable, and cutting-edge diagnosis and treatment of IBD.

Deep learning-based blind image super-resolution with iterative kernel reconstruction and noise estimation

Blind single image super-resolution (SISR) is a challenging task in image processing due to the ill-posed nature of the inverse problem. Complex degradations present in real life images make it difficult to solve this problem using naïve deep learning approaches, where models are often trained on synthetically generated image pairs. Most of the effort so far has been focused on solving the inverse problem under some constraints, such as for a limited space of blur kernels and/or assuming noise-free input images. Yet, there is a gap in the literature to provide a well-generalized deep learning-based solution that performs well on images with unknown and highly complex degradations. In this paper, we propose IKR-Net (Iterative Kernel Reconstruction Network) for blind SISR. In the proposed approach, kernel and noise estimation and high-resolution image reconstruction are carried out iteratively using dedicated deep models. The iterative refinement provides significant improvement in both the reconstructed image and the estimated blur kernel even for noisy inputs. IKR-Net provides a generalized solution that can handle any type of blur and level of noise in the input low-resolution image. IKR-Net achieves state-of-the-art results in blind SISR, especially for noisy images with motion blur.

Strength of Deep Learning-based Solutions to Secure Healthcare IoT: A Critical Review

Healthcare applications of IoT systems have gained huge popularity across the globe. From personal monitoring to expert clinical diagnosis, healthcare IoT systems have shown their importance to all possible extents. The ease of use and precise results add to the wide acceptance of such systems. However, this has also led to a magnificent increase in the number of attacks aimed at stealing or manipulating data as well as operations of HIoT-based healthcare assistance. Among the various modes of attacks, network-based attacks are found in the majority. In this work, we perform a critical review of these attacks, the existing countermeasures, and their limitations to understand and proclaim the importance of securing healthcare networks in the best possible manner. We also emphasize the necessity of deep learning-based smart solutions for securing healthcare systems, understanding the potential of deep learning in the security aspects being deployed in other genres of IoT applications. A comparative analysis of deep learning and machine learning-based security solutions is performed to examine their performances.

Online Reinforcement Learning Control of Nonlinear Dynamic Systems: A State-action Value Function Based Solution

In this paper, we present an online reinforcement learning-based solution to the optimal control problem of continuous-time nonlinear input-affine systems. The proposed approach contains a concurrent identifier that estimates time derivatives of states of the system in some arbitrary points. The identifier is utilized to simulate a so-called Bellman error in some unvisited points. The simulated errors together with errors obtained along the trajectory of the system are used to estimate the state-action value function, which is then employed to derive the estimated optimal controller. The designed approach does not explicitly require the input dynamics, which is hard to segregate it from the drift dynamics in optimal regulation problems. In addition, the simulated Bellman errors relax the confining persistence of excitation condition, which is needed for convergence in deterministic systems. A Lyapunov-based analysis was conducted to derive convergence conditions. Simulation studies demonstrated the effectiveness of the developed control scheme.

Reinforcement Learning-Based Joint User Scheduling and Link Configuration in Millimeter-Wave Networks

In this paper, we develop algorithms for joint user scheduling and three types of mmWave link configuration: relay selection, codebook optimization, and beam tracking in millimeter wave (mmWave) networks. Our goal is to design an online controller that dynamically schedules users and configures their links to minimize system delay. To solve this complex scheduling problem, we model it as a dynamic decision-making process and develop two reinforcement learning-based solutions. The first solution is based on deep reinforcement learning (DRL), which leverages the proximal policy optimization to train a neural network-based solution. Due to the potential high sample complexity of DRL, we also propose an empirical multi-armed bandit (MAB)-based solution, which decomposes the decision-making process into a sequential of sub-actions and exploits classic maxweight scheduling and Thompson sampling to decide those sub-actions. Our evaluation of the proposed solutions confirms their effectiveness in providing acceptable system delay. It also shows that the DRL-based solution has better delay performance while the MAB-based solution has a faster training process.

Model-Free Formation Control of Autonomous Underwater Vehicles: A Broad Learning-Based Solution

Model-Free Formation Control of Autonomous Underwater Vehicles: A Broad Learning-Based Solution