Resource-constrained Edge Research Articles

A Unified Parallel CORDIC-Based Hardware Architecture for LSTM Network Acceleration

Deep Neural Networks (DNNs) have recently become the standard tool for solving various practical problems in a wide range of applications with state-of-the-art performance. Recurrent Neural Networks (RNNs) such as Long Short-Term Memory (LSTM) are a subset of DNNs with fully connected single or multi-layer networks. The complex neurons and internal states of LSTM networks enable them to build a memory of events, making them ideal for time series applications. Despite the great potential of LSTM networks, their heterogeneous operations and computational resource requirements create a vast gap when it comes to the fast processing time required in real-time applications using low-power, low-cost edge devices. This work proposes a novel hardware architecture that combines serial-parallel computation with matrix algebra concepts and efficient low-power computer arithmetics for LSTM network acceleration. The hardware is based on a systolic ring of outer-product-based processing elements (PEs) and a reusable single activation function block (AFB). PEs and AFB are implemented using the coordinate rotation digital computer algorithm (CORDIC) in the linear and hyperbolic modes. Unlike most approaches, the proposed hardware can be configured to perform recurrent and non-recurrent fully connected layers (FC) computations, making it suitable for various low-power edge applications. The architecture is validated on the Xilinx PYNQ-Z1 development board using an open-source time series dataset. The implemented design achieves <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$114 \mu s$</tex-math></inline-formula> average latency and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$1.8GOPS$</tex-math></inline-formula> throughput. The proposed design's low latency and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$0.438W$</tex-math></inline-formula> power consumption makes it suitable for resource-constrained edge platforms.

Federated Transfer Learning for Rice-Leaf Disease Classification across Multiclient Cross-Silo Datasets

Paddy leaf diseases encompass a range of ailments affecting rice plants’ leaves, arising from factors like bacteria, fungi, viruses, and environmental stress. Precision agriculture leverages technologies for enhanced crop production, with disease detection being a vital element. Prompt identification of diseases in paddy leaves is critical for curtailing their propagation and reducing crop damage. However, manually diagnosing paddy diseases in regions with vast agricultural areas and limited experts proves immensely difficult. The utilization of machine learning (ML) and deep learning (DL) for diagnosing diseases in agricultural crops appears to be effective and well-suited for widespread application. These ML/DL methods cannot ensure data privacy, as they involve sharing training data with a central server, overlooking competitive and regulatory considerations. As a solution, federated learning (FL) aims to facilitate decentralized training to tackle the identified limitations of centralized training. This paper utilizes the FL approach for the classification of rice-leaf diseases. The manuscript presents an effective approach for rice-leaf disease classification with a federated architecture, ensuring data privacy. We have compiled an unbalanced dataset of rice-leaf disease images, categorized into four diseases with their respective image counts: bacterial blight (1584), brown spot (1440), blast (1600), and tungro (1308). The proposed method, called federated transfer learning (F-TL), maintains privacy for all connected devices using a decentralized client-server setup. Both IID (independent and identically distributed) and non-IID datasets were utilized for testing the F-TL framework after preprocessing. Initially, we conducted an effectiveness analysis of CNN and eight transfer learning models for rice-leaf disease classification. Among them, MobileNetV2 and EfficientNetB3 outperformed the other transfer-learned models. Subsequently, we trained these models using both IID and non-IID datasets in a federated learning environment. The framework’s performance was assessed through diverse scenarios, comparing it with traditional and federated learning models. The evaluation considered metrics like validation accuracy, loss as well as resource utilization such as CPU and RAM. EfficientNetB3 excelled in training, achieving 99% accuracy with 0.1 loss for both IID and non-IID datasets. MobilenetV2 showed slightly lower training accuracy at 98% (IID) and 90% (non-IID) with losses of 0.4 and 0.6, respectively. In evaluation, EfficientNetB3 maintained 99% accuracy with 0.1 loss for both datasets, while MobilenetV2 achieved 90% (IID) and 97% (non-IID) accuracy with losses of 0.6 and 0.2, respectively. Results indicated the F-TL framework’s superiority over traditional distributed deep-learning classifiers, demonstrating its effectiveness in both single and multiclient instances. Notably, the framework’s strengths lie in its cost-effectiveness and data-privacy assurance for resource-constrained edge devices, positioning it as a valuable alternative for rice-leaf disease classification compared to existing tools.

SDN-Based Reconfigurable Edge Network Architecture for Industrial Internet of Things

Internet of Things (IoT) with edge computing capability enhances efficiency, availability and improves latency of an industrial automation system. However, to provide dynamic services at the resource-constrained edge device, reconfiguration of services is necessary. This paper proposes a programmable edge network to (re)configure different services of industrial IoT, that employs programmable layers at the edge for reconfiguring the sensor/actuator network and application services. The lowermost layer allows reconfiguring the communication-related parameters and the middle layer consists of an SDN controller that can dynamically program different modules and handles actuation decisions from the edge. An interfacing protocol between the layers is proposed to provide reliability by considering the required configuration parameters among layers. At the top-layer, a priority forwarding mechanism is designed for SDN core (control-loop) communication when sensor and actuator are on different edges. The proposed architecture significantly improves the actuation-latency and is highly energy efficient compared to the existing state-of-the-art.

Stateless Q-learning algorithm for service caching in resource constrained edge environment

In resource constrained edge environment, multiple service providers can compete to rent the limited resources to cache their service instances on edge servers close to end users, thereby significantly reducing the service delay and improving quality of service (QoS). However, service providers renting the resources of different edge servers to deploy their service instances can incur different resource usage costs and service delay. To make full use of the limited resources of the edge servers to further reduce resource usage costs, multiple service providers on an edge server can form a coalition and share the limited resource of an edge server. In this paper, we investigate the service caching problem of multiple service providers in resource constrained edge environment, and propose an independent learners-based services caching scheme (ILSCS) which adopts a stateless Q-learning to learn an optimal service caching scheme. To verify the effectiveness of ILSCS scheme, we implement COALITION, RANDOM, MDU, and MCS four baseline algorithms, and compare the total collaboration cost and service latency of ILSCS scheme with these of these four baseline algorithms under different experimental parameter settings. The extensive experimental results show that the ILSCS scheme can achieve lower total collaboration cost and service latency.

Quantized autoencoder (QAE) intrusion detection system for anomaly detection in resource-constrained IoT devices using RT-IoT2022 dataset

In recent years, many researchers focused on unsupervised learning for network anomaly detection in edge devices to identify attacks. The deployment of the unsupervised autoencoder model is computationally expensive in resource-constrained edge devices. This study proposes quantized autoencoder (QAE) model for intrusion detection systems to detect anomalies. QAE is an optimization model derived from autoencoders that incorporate pruning, clustering, and integer quantization techniques. Quantized autoencoder uint8 (QAE-u8) and quantized autoencoder float16 (QAE-f16) are two variants of QAE built to deploy computationally expensive AI models into Edge devices. First, we have generated a Real-Time Internet of Things 2022 dataset for normal and attack traffic. The autoencoder model operates on normal traffic during the training phase. The same model is then used to reconstruct anomaly traffic under the assumption that the reconstruction error (RE) of the anomaly will be high, which helps to identify the attacks. Furthermore, we study the performance of the autoencoders, QAE-u8, and QAE-f16 using accuracy, precision, recall, and F1 score through an extensive experimental study. We showed that QAE-u8 outperforms all other models with a reduction of 70.01% in average memory utilization, 92.23% in memory size compression, and 27.94% in peak CPU utilization. Thus, the proposed QAE-u8 model is more suitable for deployment on resource-constrained IoT edge devices.

Urban Noise Analysis and Emergency Detection System using Lightweight End-to-End Convolutional Neural Network

In recent years, the application of deep learning to environmental sound classification (ESC) has received considerable attention owing to its powerful ability to recognize the context of urban sounds. In general, deep learning models with high accuracy require substantial computing and memory resources. Consequently, to apply complex deep learning models to ESC in the real world, model inference has been performed on cloud servers with powerful computing resources. However, heavy network traffic and security issues occur when inferences are performed on a cloud server. In addition, deploying a deep learning model trained on a single public ESC dataset may not be sufficient for classifying various classes of urban noise and emergency-related sounds. To address these problems, we propose an on-device, real-time urban sound monitoring system that can classify various urban sounds at low system construction costs. The proposed system consisted of an edge artificial intelligence (AI) node and a FIWARE-based server. To enable the real-time inference on a resource-constrained edge AI node, we developed a lightweight convolutional neural network (CNN) by adjusting the input and model configurations to achieve high accuracy with a low number of parameters. The model achieved 94.9% classification accuracy using only 331 K parameters on an integrated dataset that included 17 classes of urban noises and emergencies. Furthermore, a prototype of the proposed system was developed and evaluated to verify its feasibility. The prototype system was built at a cost of less than USD 50 and could perform the entire monitoring process every 3 s. We also visualized the sound monitoring data using Grafana on a FIWARE-based server.

Cloud–Edge Collaborative Inference with Network Pruning

With the increase in model parameters, deep neural networks (DNNs) have achieved remarkable performance in computer vision, but larger DNNs create a bottleneck for deploying DNNs on resource-constrained edge devices. The cloud–edge collaborative inference based on network pruning provides a solution for the deployment of DNNs on edge devices. However, the pruning methods adopted by existing frameworks are locally effective, and the compressed models are over-sparse. In this paper, we design a cloud–edge collaborative inference framework based on network pruning to make full use of the limited computing resources on edge devices. In our framework, we propose a sparsity-aware feature bias minimization pruning method to reduce the feature bias that happens during network pruning and prevent the pruned model from being over-sparse. To further reduce the inference latency, we consider the difference in computing resources between edge devices and the cloud, then design a task-oriented asymmetric feature coding to reduce the communication overhead of transmitting intermediate data. With comprehensive experiments, our framework can reduce end-to-end latency by 82% to 84% with less than 1% accuracy loss, compared to the cloud–edge collaborative inference framework with traditional methods, and our framework has the lowest end-to-end latency and accuracy loss compared to other frameworks.

BeeFlow: Behavior tree-based Serverless workflow modeling and scheduling for resource-constrained edge clusters

Serverless computing has gained popularity in edge computing due to its flexible features, including the pay-per-use pricing model, auto-scaling capabilities, and multi-tenancy support. Complex Serverless-based applications typically rely on Serverless workflows (also known as Serverless function orchestration) to express task execution logic, and numerous application- and system-level optimization techniques have been developed for Serverless workflow scheduling. However, there has been limited exploration of optimizing Serverless workflow scheduling in edge computing systems, particularly in high-density, resource-constrained environments such as system-on-chip clusters and single-board-computer clusters. In this work, we discover that existing Serverless workflow scheduling techniques typically assume models with limited expressiveness and cause significant resource contention. To address these issues, we propose modeling Serverless workflows using behavior trees, a novel and fundamentally different approach from existing directed-acyclic-graph- and state machine-based models. Behavior tree-based modeling allows for easy analysis without compromising workflow expressiveness. We further present observations derived from the inherent tree structure of behavior trees for contention-free function collections and awareness of exact and empirical concurrent function invocations. Based on these observations, we introduce BeeFlow, a behavior tree-based Serverless workflow system tailored for resource-constrained edge clusters. Experimental results demonstrate that BeeFlow achieves up to 3.2× speedup in a high-density, resource-constrained edge testbed and 2.5× speedup in a high-profile cloud testbed, compared with the state-of-the-art. BeeFlow also demonstrates superior robustness in scenarios with heavy system workloads.

Edge device-based real-time implementation of CycleGAN for the colorization of infrared video

Edge computing has gained significant attention as a pivotal technology for practical implementation of machine learning. Nevertheless, resource-constrained edge devices face challenges in meeting the performance requirements of latency-sensitive applications. Moreover, computationally intensive applications such as CycleGAN pose further hindrances to the practical utilization of machine learning. To overcome these challenges, we proposed a hybrid method that combines the implementation of a lightweight model with optimization in deployment. Initially, a lightweight sub-network is derived from a student network obtained through distillation while considering the trade-off between performance and computational complexity. Subsequently, a quantization model is deployed for low-precision inference, resulting in a substantial reduction of inference time. Finally, we apply this approach to a real-life application: a real-time headset for color night vision from infrared videos. For this purpose, three datasets of image pairs comprising long-wave infrared, visible light RGB, and up to near-infrared images are collected. Using these datasets, a lightweight implementation of the CycleGAN model is trained to translate infrared images to RGB images. To ensure efficiency, the model is deployed in a low-precision inference manner using C＋＋ on three different hardware platforms: Jetson Xavier NX, RK3399 Pro, and Raspberry Pi, each with distinct hardware architectures. Image quality and model efficiency are thoroughly analyzed. Experimental results demonstrate that our hybrid method drastically reduces inference time from 112 ms to 17 ms per frame at a resolution of 256 × 256 on the Jetson Xavier NX platform. This improvement is accompanied by only a slight degradation in image quality, enabling a real-time video frame rate close to 60 fps, thereby meeting the requirements for a real-world real-time headset for color night vision applications.

Deep Transfer Learning for Communicable Disease Detection and Recommendation in Edge Networks.

Considering the increasing number of communicable disease cases such as COVID-19 worldwide, the early detection of the disease can prevent and limit the outbreak. Besides that, the PCR test kits are not available in most parts of the world, and there is genuine concern about their performance and reliability. To overcome this, in this paper, we develop a novel edge-centric healthcare framework integrating with wearable sensors and advanced machine learning (ML) model for timely decisions with minimum delay. Through wearable sensors, a set of features have been collected that are further preprocessed for preparing a useful dataset. However, due to limited resource capacity, analyzing the features in resource-constrained edge devices is challenging. Motivated by this, we introduce an advanced ML technique for data analysis at edge networks, namely Deep Transfer Learning (DTL). DTL transfers the knowledge from the well-trained model to a new lightweight ML model that can support the resource-constraint nature of distributed edge devices. We consider a benchmark COVID-19 dataset for validation purposes, consisting of 11 features and 2 Million sensor data. The extensive simulation results demonstrate the efficiency of the proposed DTL technique over the existing ones and achieve 99.8% accuracy while diseases prediction.

Edge-Enhanced QoS Aware Compression Learning for Sustainable Data Stream Analytics

Existing Cloud systems involve large volumes of data streams being sent to a centralised data centre for monitoring, storage and analytics. However, migrating all the data to the cloud is often not feasible due to cost, privacy, and performance concerns. However, Machine Learning (ML) algorithms typically require significant computational resources, hence cannot be directly deployed on resource-constrained edge devices for learning and analytics. Edge-enhanced compressive offloading becomes a sustainable solution that allows data to be compressed at the edge and offloaded to the cloud for further analysis, reducing bandwidth consumption and communication latency. The design and implementation of a learning method for discovering compression techniques that offer the best QoS for an application is described. The approach uses a novel modularisation approach that maps features to models and classifies them for a range of Quality of Service (QoS) features. An automated QoS-aware orchestrator has been designed to select the best autoencoder model in real-time for compressive offloading in edge-enhanced clouds based on changing QoS requirements. The orchestrator has been designed to have diagnostic capabilities to search appropriate parameters that give the best compression. A key novelty of this work is harnessing the capabilities of autoencoders for edge-enhanced compressive offloading based on portable encodings, latent space splitting and fine-tuning network weights. Considering how the combination of features lead to different QoS models, the system is capable of processing a large number of user requests in a given time. The proposed hyperparameter search strategy (over the neural architectural space) reduces the computational cost of search through the entire space by up to 89%. When deployed on an edge-enhanced cloud using an Azure IoT testbed, the approach saves up to 70% data transfer costs and takes 32% less time for job completion. It eliminates the additional computational cost of decompression, thereby reducing the processing cost by up to 30%.

PASS: Patch Automatic Skip Scheme for Efficient Real-Time Video Perception on Edge Devices

Real-time video perception tasks are often challenging over the resource-constrained edge devices due to the concerns of accuracy drop and hardware overhead, where saving computations is the key to performance improvement. Existing methods either rely on domain-specific neural chips or priorly searched models, which require specialized optimization according to different task properties. In this work, we propose a general and task-independent Patch Automatic Skip Scheme (PASS), a novel end-to-end learning pipeline to support diverse video perception settings by decoupling acceleration and tasks. The gist is to capture the temporal similarity across video frames and skip the redundant computations at patch level, where the patch is a non-overlapping square block in visual. PASS equips each convolution layer with a learnable gate to selectively determine which patches could be safely skipped without degrading model accuracy. As to each layer, a desired gate needs to make flexible skip decisions based on intermediate features without any annotations, which cannot be achieved by conventional supervised learning paradigm. To address this challenge, we are the first to construct a tough self-supervisory procedure for optimizing these gates, which learns to extract contrastive representation, i.e., distinguishing similarity and difference, from frame sequence. These high-capacity gates can serve as a plug-and-play module for convolutional neural network (CNN) backbones to implement patch-skippable architectures, and automatically generate proper skip strategy to accelerate different video-based downstream tasks, e.g., outperforming the state-of-the-art MobileHumanPose (MHP) in 3D pose estimation and FairMOT in multiple object tracking, by up to 9.43 times and 12.19 times speedups, respectively. By directly processing the raw data of frames, PASS can generalize to real-time video streams on commodity edge devices, e.g., NVIDIA Jetson Nano, with efficient performance in realistic deployment.

Optimal Device Selection in Federated Learning for Resource-Constrained Edge Networks

Low latency, resource efficiency, and data privacy are some of the crucial requirements in modern communication networks. Federated learning can efficiently address these issues by utilizing the data at the network edge and processing massive amounts of data in parallel at the edge devices, thus ensuring data privacy and low latency. For a large-scale federated learning task spanning many devices, challenges arise due to device heterogeneity, data variability, and limited network resources. Optimal selection of edge devices participating in federated learning is essential to attaining resilient, reliable, and resource-efficient edge networks. In this context, this paper proposes an optimal device selection method to minimize redundant data training and improve network resource utilization without affecting the performance of federated learning over resource-constrained edge networks. The proposed optimal device selection method aims to minimize network resource demands while maximizing data diversity within the aggregated model. The performance of the proposed federated learning framework is evaluated using a publicly available image dataset of handwritten digits, EMNIST (an extended version of the MNIST dataset). Experimental results indicate that the proposed framework can obtain accuracy convergence performance on par with conventional federated learning methods while significantly reducing device usage (up to 50%) and resource utilization (up to 30%) while reaching 99% of achievable accuracy. The proposed method can therefore be effectively applied to resource-constrained edge networks.

Mobility-Aware Registry Migration for Containerized Applications on Edge Computing Infrastructures

Small footprints and fast provisioning times have promoted container adoption for deploying and managing applications on edge computing environments. As keeping all container images locally would quickly saturate the storage of resource-constrained edge servers, application provisioning consists of pulling container images from external repositories, called registries, located in specific locations in the infrastructure. Existing research reduces deployment time on edge infrastructures by defining the location of container registries. Although such an approach yields positive results in specific scenarios, it overlooks that the demand for container images in certain regions can vary over time according to users’ mobility. This work presents a novel strategy that provisions container registries dynamically based on users’ mobility, spinning up new registries when application provisioning times start growing excessively and deprovisioning registries far away from users. Experimental results demonstrate that our approach reduces the application provisioning time issues by 33.19% on average compared to strategies that allocate container registries statically.

Resource-aware multi-task offloading and dependency-aware scheduling for integrated edge-enabled IoV

Internet of Vehicles (IoV) enables a wealth of modern vehicular applications, such as pedestrian detection, real-time video analytics, etc., that can help to improve traffic efficiency and driving safety. However, these applications impose significant resource demands on the in-vehicle resource-constrained Edge Computing (EC) device installation. In this article, we study the problem of resource-aware offloading of these computation-intensive applications to the Closest roadside units (RSUs) or telecommunication base stations (BSs), where on-site EC devices with larger resource capacities are deployed, and mobility of vehicles are considered at the same time. Specifically, we propose an Integrated EC framework, which can keep edge resources running across various in-vehicles, RSUs and BSs in a single pool, such that these resources can be holistically monitored from a single control plane (CP). Through the CP, individual in-vehicle, RSU or BS edge resource availability can be obtained, hence applications can be offloaded concerning their resource demands. This approach can avoid execution delays due to resource unavailability or insufficient resource availability at any EC deployment. This research further extends the state-of-the-art by providing intelligent multi-task scheduling, by considering both task dependencies and heterogeneous resource demands at the same time. To achieve this, we propose FedEdge, a variant Bin-Packing optimization approach through Gang-Scheduling of multi-dependent tasks that co-schedules and co-locates multi-task tightly on nodes to fully utilize available resources. Extensive experiments on real-world data trace from the recent Alibaba cluster trace, with information on task dependencies and resource demands, show the effectiveness, faster executions, and resource efficiency of our approach compared to the existing approaches.

Efficient federated learning on resource-constrained edge devices based on model pruning

Federated learning is an effective solution for edge training, but the limited bandwidth and insufficient computing resources of edge devices restrict its deployment. Different from existing methods that only consider communication efficiency such as quantization and sparsification, this paper proposes an efficient federated training framework based on model pruning to simultaneously address the problem of insufficient computing and communication resources. First, the framework dynamically selects neurons or convolution kernels before the global model release, pruning a current optimal subnet and then issues the compressed model to each client for training. Then, we develop a new parameter aggregation update scheme, which provides training opportunities for global model parameters and maintains the complete model structure through model reconstruction and parameter reuse, reducing the error caused by pruning. Finally, extensive experiments show that our proposed framework achieves superior performance on both IID and non-IID datasets, which reduces upstream and downstream communication while maintaining the accuracy of the global model and reducing client computing costs. For example, with accuracy exceeding the baseline, computation is reduced by 72.27% and memory usage is reduced by 72.17% for MNIST/FC; and computation is reduced by 63.39% and memory usage is reduced by 59.78% for CIFAR10/VGG16.

Cloning-based virtual machine pre-provisioning for resource-constrained edge cloud server

Cloning-based virtual machine pre-provisioning for resource-constrained edge cloud server

BETTER: Bayesian-Based Training and Lightweight Transfer Architecture for Reliable and High-Speed Memristor Neural Network Deployment

Deep learning models implemented using memristors show high scalability and high energy efficiency, promising a compact and efficient computing architecture for resourceconstrained edge computing applications. These technologies integrate both data storage and computation simultaneously in a highly parallel memristor crossbar array architecture. However, the significant variations arising from the inherent physical randomness of memristors lead to a large performance degradation of deep learning models. The challenges of extensive energy costs and transfer time for deployment to maintain performance are faced. In this paper, for the first time, we propose a unified architecture that consists of a Bayesian-based training method and a lightweight transfer scheme. The proposed architecture can tackle the robustness, energy and time consumption issues caused by memristor variations. Our experimental results show that our architecture can double the speed and energy efficiency of deploying deep learning models.

Wireless Channel Adaptive DNN Split Inference for Resource-Constrained Edge Devices

Split inference facilitates deep neural network (DNN) inference tasks at resource-constrained edge devices. However, a pre-determined split configuration of a DNN limits the inference performance in time-varying wireless channels. To accelerate the inference, we propose a two-stage wireless channel adaptive split inference method by considering memory and energy constraints on the edge device. The proposed scheme is able to offer the privacy of the edge device and improves inference performance in time-varying wireless channels by leveraging a U-shaped DNN splitting framework and adaptively determining the splitting points of a DNN in real-time according to time-varying wireless channel gains.

HARBIC: Human activity recognition using bi-stream convolutional neural network with dual joint time–frequency representation

Human activity recognition (HAR) based on wearable devices has progressively advanced in context-aware computing like healthcare, smart homes, and industry 4.0. Since machine learning and Internet-of-Things have been in the spotlight, many deep learning-based approaches have been proposed, achieving very good performance using inertial data. In this study, we explore different joint time–frequency representations of the sensor data, in conjunction with a Convolutional Neural Network (ConvNet) to implement activity classification. We also consider the concurrent use of two different data representations in a cooperative bi-stream ConvNet configuration. Scaling of joint time–frequency representation is explored to reduce the ConvNet model size, and hence the computational complexity of the proposed approach. The classification performance and energy efficiency of the proposed method are evaluated on a real-world public HAR dataset. The performance of pairwise combinations and also, the single-stream of the two prominent representations in the joint time–frequency domain are compared to that of the state-of-the-art. The proposed method is implemented on Google’s Edge TPU platform to investigate the trade-off between the energy footprint and classification performance, which is critical in IoT applications where machine learning inference needs to be performed on resource-constrained edge devices. The highest overall classification accuracy and F1 score achieved by our method was 98.39% and 96.86%, respectively, which represents a significant improvement over the state-of-the-art.