Federated Learning (FL) has emerged as a decentralised machine learning paradigm for distributed systems, particularly in edge and IoT environments. However, achieving fault tolerance and self-recovery in such scenarios is challenging due to the centralised model aggregation, which poses a single point of failure. This article focuses on the self-recovery of the aggregator, specifically the controlled re-assignment of the aggregator role to the most suitable node. Our proposed solution leverages the RAFT consensus algorithm to facilitate consistent state replication and leader election within the FL system. This is complemented by controlled aggregator re-assignment, which considers various contextual properties to select the optimal node, enhancing the system’s robustness, especially in dynamic and unreliable cyber-physical environments. We implement a proof of concept using the Flower FL framework and conduct experiments to evaluate aggregator recovery time and the traffic overhead associated with state replication. While the traffic overhead scales with the number of FL nodes, our results demonstrate a resilient, self-recovering system capable of handling node failures while maintaining model consistency.

Integrating Artificial Intelligence (AI) with the Industrial Internet of Things (IIoT) has transformed industrial processes, enhancing productivity, quality control, and operational efficiency. However, ensuring the precision and reliability of sensor-generated data remains a critical challenge due to the evolving nature of industrial processes and the limitations of conventional validation methods. Traditional rule-based and supervised learning approaches struggle to adapt to process shifts, drifts, and novel anomalies, making sensor data validation an ongoing issue. This article introduces UDAVA (Unsupervised Learning Approach using Process Mining for Sensor Data Validation in IIoT), a novel AI-driven pipeline designed to automate the identification of reference patterns in sensor data and validate subsequent production cycles by recognizing deviations from expected behaviors. UDAVA employs a multi-stage process that includes preprocessing sensor data, clustering recurring patterns, and assessing deviations. It supports semi-supervised learning by integrating manual labels where available, improving interpretability and accuracy. One of UDAVA’s key strengths lies in its ability to extract features from sensor data rather than relying on raw time series similarity, making it robust against noise and diverse process variations. Additionally, UDAVA integrates process mining techniques—process discovery and conformance checking—to enhance its ability to detect even subtle anomalies and deviations in industrial workflows. We conduct a comprehensive evaluation of UDAVA using three industrial datasets, demonstrating its effectiveness in identifying high-level process behaviors, detecting process shifts and drifts, and ensuring data validation across multiple production cycles. The results highlight UDAVA ’s adaptability across different industrial processes, making it a valuable tool for optimizing operations and ensuring sensor data reliability in IIoT environments.

Automatic Speech Recognition (ASR) systems are increasingly deployed across diverse computing environments, from cloud servers to edge devices. While accuracy has traditionally been the primary evaluation metric, the inference efficiency of these systems, including energy consumption, memory usage, and hardware utilisation, significantly impacts their practical usability. This paper introduces a novel benchmarking framework that assesses ASR models during inference from both performance and sustainability perspectives. We introduce a multi-metric evaluation approach quantifying Word Error Rate (WER), Real-Time Factor (RTF), Energy Per Audio Second (EPAS), inference latency, GPU Memory Efficiency (GME), and Hardware Utilisation Rate (HUR). Our framework includes configurable weighting schemes tailored for various deployment scenarios: balanced general-purpose evaluation, resource-constrained environments, high-throughput batch inference, and real-time processing. To demonstrate the utility of the framework, we benchmark several state-of-the-art ASR architectures (Whisper, Wav2Vec2, HuBERT, WavLM, UniSpeech, and SpeechT5) in both FP16 and FP32 precision on NVIDIA Jetson AGX Orin hardware. The proposed methodology supports researchers and practitioners in making informed model selection decisions based on context-specific inference requirements. By illuminating performance–consumption trade-offs, the metrics framework can help to reduce computational costs and the carbon footprint of ASR systems, while maintaining acceptable accuracy.

Deploying Large Language Models (LLMs) on edge devices presents significant challenges due to computational constraints, memory limitations, inference speed, and energy consumption. Model quantization has emerged as a key technique to enable efficient LLM inference by reducing model size and computational overhead. In this study, we conduct a comprehensive analysis of 28 quantized LLMs from the Ollama library, which applies by default Post-Training Quantization (PTQ) and weight-only quantization techniques, deployed on an edge device (Raspberry Pi 4 with 4 GB RAM). We evaluate energy efficiency, inference performance, and output accuracy across multiple quantization levels and task types. Models are benchmarked on five standardized datasets (CommonsenseQA, BIG-Bench Hard, TruthfulQA, GSM8K, and HumanEval), and we employ a high-resolution, hardware-based energy measurement tool to capture real-world power consumption. Our findings reveal the trade-offs between energy efficiency, inference speed, and accuracy in different quantization settings, highlighting configurations that optimize LLM deployment for resource-constrained environments. By integrating hardware-level energy profiling with LLM benchmarking, this study provides actionable insights for sustainable AI, bridging a critical gap in existing research on energy-aware LLM deployment.

The proliferation of machine learning (ML) has brought unprecedented advancements in technology, but it has also raised concerns about its environmental impact, particularly concerning carbon emissions. To address the imperative of environmentally responsible ML, we present in this paper a novel ML pipeline, named CEMAI, designed to monitor and analyze carbon emissions across the entire lifecycle of ML model development, from data preparation to training and deployment. Our endeavor involves an exhaustive evaluation process underpinned by three industrial case studies. These case studies are structured around the application of ML models to predict tool wear, estimate remaining useful lifetimes, and detect anomalies in the Industrial Internet of Things (IIoT). Leveraging sensor data originating from CNC machining and broaching operations, our research shows empirically the efficacy of carbon emissions as a dependable metric guiding the configuration of an ML development process. The essence of our approach lies in striking a balance between superior performance and minimal carbon emissions. Our findings reveal the potential to optimize pipeline configurations for ML models in a manner that not only enhances performance but also drastically reduces carbon emissions, thereby underlining the significance of adopting ecologically responsible engineering practices.

Abstract Workers, especially those involved in jobs requiring extended standing or repetitive movements, often face significant health challenges due to Musculoskeletal Disorders (MSDs). To mitigate MSD risks, enhancing workplace ergonomics is vital, which includes forecasting long-term employee postures, educating workers about related occupational health risks, and offering relevant recommendations. However, research gaps remain, such as the lack of a sustainable AI/ML pipeline that combines sensor-based, uncertainty-aware posture prediction with large language models for natural language communication of occupational health risks and recommendations. We introduce ERG-AI, a machine learning pipeline designed to predict extended worker postures using data from multiple wearable sensors. Alongside providing posture prediction and uncertainty estimates, ERG-AI also provides personalized health risk assessments and recommendations by generating prompts based on its performance and prompting Large Language Model (LLM) APIs, like GPT-4, to obtain user-friendly output. We used the Digital Worker Goldicare dataset to assess ERG-AI, which includes data from 114 home care workers who wore five tri-axial accelerometers in various bodily positions for a cumulative 2913 hours. The evaluation focused on the quality of posture prediction under uncertainty, energy consumption and carbon footprint of ERG-AI and the effectiveness of personalized recommendations rendered in easy-to-understand language.

The Industrial Internet of Things (IIoT) leverages thousands of interconnected sensors and computing devices to monitor and control large and complex industrial processes. Machine learning (ML) applications in IIoT use data acquired from multiple sensors to perform tasks such as predictive maintenance. While remembering useful learning from the past, these applications need to adapt learning for evolving sensor data stemming from changes in industrial processes and environmental conditions. This paper presents a continual learning pipeline to learn from the evolving data while replaying selected parts of the old data. The pipeline is configured to produce ML experiences (e.g., training a baseline neural network model), improve the baseline model with the new data while replaying part of the old data, and infer/predict using a specific model version given a stream of IIoT sensor data. We have evaluated our approach from an AI Engineering perspective using three industrial case studies, i.e., predicting tool wear, remaining useful lifetime, and anomalies from sensor data acquired from CNC machining and broaching operations. Our results show that configuring experiences for replay-driven continual learning allows dynamic maintenance of ML performance on evolving data while minimizing the excessive accumulation of legacy sensor data.

Unsupervised learning systems using clustering have gained significant attention for numerous applications due to their unique ability to discover patterns and structures in large unlabeled datasets. However, their effectiveness highly depends on their configuration, which requires domain-specific expertise and often involves numerous manual trials. Specifically, selecting appropriate algorithms and hyperparameters adds to the complexity of the configuration process. In this paper, we propose, apply, and assess an automated approach (AutoConf) for configuring unsupervised learning systems using clustering, leveraging metamorphic testing and Bayesian optimization. Metamorphic testing is utilized to verify the configurations of unsupervised learning systems by applying a series of input transformations. We use Bayesian optimization guided by metamorphic-testing output to automatically identify the optimal configuration. The approach aims to streamline the configuration process and enhance the effectiveness of unsupervised learning systems. It has been evaluated through experiments on six datasets from three domains for anomaly detection. The evaluation results show that our approach can find configurations outperforming the baseline approaches as they achieved a recall of 0.89 and a precision of 0.84 (on average).

Manufacturing has enabled the mechanized mass production of the same (or similar) products by replacing craftsmen with assembly lines of machines. The quality of each product in an assembly line greatly hinges on continual observation and error compensation during machining using sensors that measure quantities such as position and torque of a cutting tool and vibrations due to possible imperfections in the cutting tool and raw material. Patterns observed in sensor data from a (near-)optimal production cycle should ideally recur in subsequent production cycles with minimal deviation. Manually labeling and comparing such patterns is an insurmountable task due to the massive amount of streaming data that can be generated from a production process. We present UDAVA, an unsupervised machine learning pipeline that automatically discovers process behavior patterns in sensor data for a reference production cycle. UDAVA performs clustering of reduced dimensionality summary statistics of raw sensor data to enable high-speed clustering of dense time-series data. It deploys the model as a service to verify batch data from subsequent production cycles to detect recurring behavior patterns and quantify deviation from the reference behavior. We have evaluated UDAVA from an AI Engineering perspective using two industrial case studies.

The role of Machine Learning (ML) engineers in mobile development has become increasingly important in recent years, as more and more business-critical mobile applications depend on AI components. Many development teams already include dedicated ML engineers who aim to follow agile development practices in their work, as part of the larger MLOps concept. However, the availability of MLOps tools tailored specifically towards mobile platforms is scarce, often due the limited support for non-native programming languages such as Python, as well as the unsuitability of native programming languages such as Java and Kotlin to support ML-related programming tasks. This paper aims to address this gap and describes a plug-in architecture for developing, deploying and running data ingestion and processing components written in Python on the Android platform. With the possibility to pass a user-defined schema with the data format and structure, the proposed architecture ensures that time-series datasets are correctly interpreted by multiple ML modules dealing with both data ingestion and processing,. The proposed approach benefits from modularity, extensibility, customisation, and separation of concerns, which enable ML engineers to be fully involved in a mobile development lifecycle following agile MLOps practices.

Tracking physical effort from physiological signals has enabled people to manage required activity levels in our increasingly sedentary and automated world. Breathing is a physiological process that is a reactive representation of our physical effort. In this demo, we present DeepVentilation, a deep learning system to predict minute ventilation in litres of air a person moves in one minute uniquely from real-time measurement of rib-cage breathing forces. DeepVentilation has been trained on input signals of expansion and contraction of the rib-cage obtained using a non-invasive respiratory inductance plethysmography sensor to predict minute ventilation as observed from a face/head mounted exercise spirometer. The system is used to track physical effort closely matching our perception of actual exercise intensity. The source code for the demo is available here: https://github.com/simula-vias/DeepVentilation