Modern edge-cloud systems face challenges in ef-ficiently scaling resources to handle dynamic and unpredictable workloads. Traditional scaling approaches typically rely on static thresholds and predefined rules, which are often inadequate for optimizing resource utilization and maintaining performance in distributed and dynamic environments. This inefficiency hin-ders the adaptability and performance required in edge-cloud infrastructures, which can only be achieved through the newly proposed in-place scaling. To address this problem, we propose the Multi-Agent Reinforcement Learning-based In-place Scaling Engine (MARLISE) that enables seamless, dynamic, reactive control with in-place resource scaling. We develop our solution using two Deep Reinforcement Learning algorithms: Deep Q-Network (DQN), and Proximal Policy Optimization (PPO). We analyze each version of the proposed MARLISE solution using dynamic workloads, demonstrating their ability to ensure low response times of microservices and scalability. Our results show that MARLISE- based approaches outperform heuristic method in managing resource elasticity while maintaining microservice response times and achieving higher resource efficiency.

In this paper, we demonstrate a smart resource auto-scaling solution based on Multi-Agent Deep Reinforcement Learning (MADRL). Our decentralized approach dynamically adapts compute resources to services to ensure a smooth response to fluctuating user demands. The demonstration, consisting of a visualization and control dashboard, a cloud edge hardware configuration and an AI-based wireless fingerprinting service, shows in-place real-time scaling of resources in the virtual environment and its impact on service response time and performance.

In many cases, a machine learning model must learn to correctly predict a few data points with particular values of interest in a broader range of data where many target values are zero. Zero-inflated data can be found in diverse scenarios, such as lumpy and intermittent demands, power consumption for home appliances being turned on and off, impurities measurement in distillation processes, and even airport shuttle demand prediction. The presence of zeroes affects the models’ learning and may result in poor performance. Furthermore, zeroes also distort the metrics used to compute the model’s prediction quality. This paper showcases two real-world use cases (home appliances classification and airport shuttle demand prediction) where a hierarchical model applied in the context of zero-inflated data leads to considerable performance improvements. In particular, for home appliances classification, the weighted average of Precision, Recall, F1, and Area Under the Receiver Operating Characteristic Curve (AUC ROC) was increased by 39%, 49%, 88%, and 48%, respectively. Furthermore, it is estimated that the proposed approach is also four times more energy efficient than the state-of-the-art (SOTA) approach against which it was compared to. Two-fold modeling approaches significantly outperform regular regression, especially when predicting the occurrence of demand events. SOTA results were achieved using Gradient Boosting trees to determine whether an event will occur and Visual Geometry Group (VGG) or Support Vector Regressor (SVR) models for the subsequent classification/regression. The code has been released at two separate repositories. • We reframe predictions on zero-inflated data, posing it as a two-fold problem. • A two-fold model enables optimizing each stage for a different but complementary goal. • Two-fold models achieve SOTA results on two real-world use cases. • For home appliances classification, the weighted average AUC ROC was increased by 48%. • For shuttle transfer demand, two-fold models performed significantly better (P = 0.05).

Time‐series classification is a relevant step supporting decision‐making processes in various domains, and deep neural models have shown promising performance in this respect. Despite significant advancements in deep learning, the theoretical understanding of how and why complex architectures function remains limited, prompting the need for more interpretable models. Recently, the Kolmogorov–Arnold Networks (KANs) have been proposed as a more interpretable alternative to deep learning. While KAN‐related research is significantly rising, to date, the study of KAN architectures for time‐series classification has been limited. In this paper, we aim to conduct a comprehensive and robust exploration of the KAN architecture for time‐series classification utilizing 117 datasets from UCR benchmark archive, from multiple different domains. More specifically, we investigate (a) the transferability of reference architectures designed for regression to classification tasks, (b) the hyperparameter and implementation configurations for an architecture that best generalizes across 117 datasets, (c) the associated complexity trade‐offs, and (d) KANs interpretability. Our results demonstrate that (1) the Efficient KAN outperforms MLPs in both performance and training times, showcasing its suitability for classification tasks. (2) Efficient KAN exhibits greater stability than the original KAN across grid sizes, depths, and layer configurations, especially when lower learning rates are employed. (3) KAN achieves competitive accuracy compared to state‐of‐the‐art models such as HIVE‐COTE2 and InceptionTime, while maintaining smaller architectures and faster training times, highlighting its favorable balance of performance and transparency. (4) The interpretability of the KAN model, as confirmed by SHAP analysis, reinforces its capacity for transparent decision‐making.

With the process of democratization of the network edge, hardware and software for networks are becoming available to the public, overcoming the confines of traditional cloud providers and network operators. This trend, coupled with the increasing importance of AI in 6G and beyond cellular networks, presents opportunities for innovative AI applications and systems at the network edge. While AI models and services are well-managed in cloud systems, achieving similar maturity for serving network needs remains an open challenge. Existing open solutions are emerging and are yet to consider democratization requirements. In this work, we identify key requirements for democratization and propose NAOMI, a solution for democratizing AI/ML workflows at the network edge designed based on those requirements. Guided by the functionality and overlap analysis of the O-RAN AI/ML workflow architecture and MLOps systems, coupled with the survey of open-source AI/ML tools, we develop a modular, scalable, and distributed hardware architecture-independent solution. NAOMI leverages state-of-the-art open-source tools and can be deployed on distributed clusters of heterogeneous devices. The results show that NAOMI performs up to 40% better in deployment time and up to 73% faster in AI/ML workflow execution for larger datasets compared to AI/ML Framework, a representative open network access solution, while performing inference and utilizing resources on par with its counterpart.

The shift towards software-centric network infrastructures is driven by the increasing need for networks to be responsive, flexible, and scalable in the face of an expanding set of connected devices. The digital twin (DT) approach, mirroring physical entities in a digital format, has emerged as a key enabler of network reliability and availability. Incorporating artificial intelligence (AI) into DTs enhances the resilience of networks by providing in-depth analysis and increasingly automated mitigation strategies against operational disruptions. In this work, we propose a new AI-based information extraction module that is part of the DT Monitoring component able to process RSS data, extract and characterize abrupt anomalies. The output of this component is used to maintain an anomaly history in the Link Abstraction within the DT and subsequently inform possible automatic mitigation actions. We design the AI-based information extraction module to identify and characterize three types of RSS based anomalies. Our extensive performance analysis on 10 versions of the “You Only Look Once” architecture reveals that YOLOv8n produces a good tradeoff between performance and computational complexity. We show that our approach performs on par with the state of the art for anomaly detection, while enabling anomaly characterization by location, duration, and severity. By employing two SotA explainability algorithms, we also provide insights into the important regions of the input that trigger the selected model’s classification and characterization decisions.

Energy management systems (EMS), as enablers of more efficient energy consumption, monitor and manage appliances to help residents be more energy efficient and thus more frugal. Recent appliance detection and identification techniques for such systems rely on machine learning. However, machine learning solutions for appliance classification on existing low-frequency household metering have not yet been thoroughly investigated. In this paper, we propose CARMEL, a new approach for identifying home appliances from load monitoring in building EMS based on a new data representation technique and a new model that leverages spatio-temporal correlations in the new representation. The proposed data representation technique performs dimensionality expansion of time series that scales linearly, rather than quadratically and, together with the proposed model, outperform the state of the art image transformation models by 5 percentage points. Evaluation on 5 different low-frequency household metering datasets, considering 29 appliances in total, shows that the proposed representation and the corresponding resource-aware deep learning architecture (1) achieve an average weighted F1 score of 0.92 and (2) require only 230 labeled samples and 3x fewer epochs to transfer to new households.

Nowadays, modern man-made infrastructures are being upgraded with information and communication technologies that form large wireless networks. Such large wireless networks must be monitored to ensure reliable operation by the proactive detection and correction of link failures or abnormal network behaviour in view of uninterrupted business operations. In this paper, we present a new method for detecting wireless link anomalies based on graph neural networks. The proposed method transforms time series data into graphs using a Markov Transition Field transformation. The data resulting from the transformation then trains a new graph neural network architecture to learn to successfully discriminate between 4 different link layer anomalies with an average F1 score of 0.956. The resulting model achieves competitive results with superior detection ability for the more subtle slow degradation anomaly while having up to ≈230 times fewer trainable parameters compared to the state-of-the-art, which subsequently makes it computationally much more efficient.

Driven by various academic, standardization and regulatory initiatives, recent research on spectrum resource utilisation has focused also on technology and transmission classification using various deep learning (DL) architectures. However, especially in unlicensed bands it is often hard to obtain labelled data of sufficient quality for training DL for all transmissions that may occur. Therefore in this paper we investigate clustering techniques that require no labelled data or prior knowledge on continuous spectrum sensing sweeps over a 200 kHz wide band in the unlicensed European 868 MHz frequency band. Using several clusterability tests we show that the sweeps can be clustered, however the number of clusters is not clear. By analyzing the sweeps with three state-of-the-art techniques, K-Means, Agglomerative Hierarchical Clustering, and Hierarchical Density-Based Spatial Clustering (HDBSCAN) 9-10 clusters are discovered. The quantitative evaluation shows that HDBSCAN outperforms the other two and the qualitative analysis shows that HDBSCAN seems to be able to better discriminate between the same technology transmitting at/from different power levels/distance, however it has a relatively poorer class balance compared to K-Means and tends to group more transmissions than needed in the majority no transmission cluster.

Time series classification is an important task in many fields. In intrusive and non-intrusive load monitoring (N)ILM, time series data are obtained from power measurements of electrical appliances that are not known in advance, therefore extracting the type of appliance from the data is a relevant problem in smart grids. We propose a transformation that encodes a time series into an image that can be effectively recognised by well known image classification algorithms. The transformation is based on plotting the time series into a matrix and filling the area under it, producing a more pronounced representation of its shape. We perform an extensive evaluation on 1 synthetic and 4 measured datasets. Our experiments on synthetic and mixed measured data yielded F1 scores of 99.2% and 85.9%, respectively, and outperformed the state-of-the-art on three out of four tested datasets. Additionally, we conclude that our method tends to work better on longer time series segments, as the resulting images contain more distinguishing features.

In recent years, much work has been done on processing of wireless spectrum data involving machine learning techniques in domain-related problems for cognitive radio networks, such as anomaly detection, modulation classification, technology classification and device fingerprinting. Most of the solutions are based on labeled data, created in a controlled manner and processed with supervised learning approaches. However, spectrum data measured in real-world environment is highly nondeterministic, making its labeling a laborious and expensive process, requiring domain expertise, thus being one of the main drawbacks of using supervised learning approaches in this domain. In this paper, we investigate the utilization of self-supervised learning (SSL) for exploring spectrum activities in a real-world unlabeled data. In particular, we assess the performance of SSL models, based on the reference DeepCluster architecture. We carefully consider the current state-of-the-art feature extractors, taking into account the performance and complexity trade-offs. Our findings demonstrate that SSL models achieve superior performance regarding the feature quality and clustering performance compared to baseline feature learning approaches. With SSL models we achieve significant reduction of the feature vectors size by two orders of magnitude, while improving the performance by a factor ranging from 2 to 2.5 across the evaluation metrics, supported by visual assessment. Furthermore, we showcase how adapting the reference SSL architecture to domain-specific data is followed by a substantial reduction in model complexity up to one order of magnitude, without compromising, and in some cases, even improving the clustering performance.

The number of end devices that use the last-mile wireless connectivity is dramatically increasing with the rise of smart infrastructures and requires reliable functioning to support smooth and efficient business processes. To efficiently manage such massive wireless networks, more advanced and accurate network monitoring and malfunction detection solutions are required. In this article, we perform a first-time analysis of image-based representation techniques for wireless anomaly detection using recurrence plots (RPs) and Gramian angular fields and propose a new deep learning architecture enabling accurate anomaly detection. We elaborate on the design considerations for developing a resource-aware architecture and propose a new model using time series to image transformation using RPs. We show that the proposed model: 1) outperforms the one based on Gramian angular fields by up to 14% points; 2) outperforms classical ML models using dynamic time warping by up to 24% points; 3) outperforms or performs on par with mainstream architectures, such as AlexNet and VGG11 while having their weights and up to ≈8% of their computational complexity; and d) outperforms the state of the art in the respective application area by up to 55% points. Finally, we also explain on randomly chosen examples how the classifier takes decisions.

Location based services, already popular with end users, are now inevitably becoming part of new wireless infrastructures and emerging business processes. The increasingly popular Deep Learning (DL) artificial intelligence methods perform very well in wireless fingerprinting localization based on extensive indoor radio measurement data. However, with the increasing complexity these methods become computationally very intensive and energy hungry, both for their training and subsequent operation. Considering only mobile users, estimated to exceed 7.4 billion by the end of 2025, and assuming that the networks serving these users will need to perform only one localization per user per hour on average, the machine learning models used for the calculation would need to perform 65×1012predictions per year. Add to this equation tens of billions of other connected devices and applications that rely heavily on more frequent location updates, and it becomes apparent that localization will contribute significantly to carbon emissions unless more energy-efficient models are developed and used. This motivated our work on a new DL-based architecture for indoor localization that is more energy efficient compared to related state-of-the-art approaches while showing only marginal performance degradation. A detailed performance evaluation shows that the proposed model produces only 58% of the carbon footprint while maintaining 98.7% of the overall performance compared to state of the art model external to our group. Additionally, we elaborate on a methodology to calculate the complexity of the DL model and thus the CO2footprint during its training and operation.

Machine learning (ML) techniques play a significant role in detecting anomalous wireless links. However, to date, to the extent of our knowledge, there is no robust classifier that would work in a realistic scenario where various anomalies could appear concurrently in the time-series gleaned from the network monitoring tools. In this paper, we propose a new deep learning based classifier and show that is able to outperform the state of the art for existing link layer anomalies. Our evaluation results demonstrate that the state-of-the ML models perform with an average accuracy of about 63%, whereas the average accuracy of the proposed DL model is around 90%, indicating a significant improvement of 27 percentage points anomaly detection performance.