Evaluating Edge AI for Real-Time Industrial Anomaly Detection

Thesis Abstract

Industrial manufacturing increasingly relies on edge AI to meet the demand for real-time anomaly detection while preserving data locality, reducing latency, and mitigating bandwidth constraints. However, deploying and evaluating edge-enabled anomaly detection systems in diverse industrial environments poses challenges related to model generalization under heterogeneous sensor setups, resource-constrained devices, and evolving process conditions. This study addresses these gaps by empirically evaluating edge AI pipelines for real-time industrial anomaly detection, focusing on performance, reliability, and operational impact in live manufacturing settings. The aim is to determine how edge AI architectures compare with cloud-centric and hybrid approaches in terms of detection accuracy, latency, energy consumption, and maintainability, and to identify factors that influence their effectiveness in production. Specific objectives are (1) to quantify detection performance across multiple fault types using time-series sensor data collected from three distinct production lines; (2) to measure end-to-end latency, inference energy, and throughput under varying workload scenarios; (3) to assess model robustness to concept drift and sensor failures through controlled stress-testing; (4) to evaluate deployment feasibility, including resource utilization on edge devices (ARM-based and x86-based gateways) and data governance implications; and (5) to develop a prescriptive framework for selecting edge AI configurations tailored to process characteristics and operational constraints. The methodology adopts a mixed-methods, empirical field study design. The population comprises industrial lines in a mid-sized manufacturing facility equipped with vibration, temperature, and acoustic sensors. A sample of 40 sensors across three lines serves as the primary data source, with 20 bespoke edge devices deployed for on-device inference and 5 gateway systems collecting centralized cloud-assisted analyses for comparison. Data collection instruments include synchronized time-series datasets, edge device telemetry (CPU/GPU utilization, memory, energy draw), and operator feedback captured through structured interviews. The study employs supervised anomaly detection models based on lightweight architectures (e.g., TinyML CNNs and LSTM variants) trained on historical fault-free and fault-label data, with parallel experiments implementing cloud-based and hybrid pipelines. Validity and reliability are ensured through cross-validation, repeat runs under identical conditions, and calibration against ground-truth maintenance logs. Analytical techniques comprise descriptive statistics to profile operational baselines, followed by inferential analyses. Performance is evaluated using precision, recall, F1-score, and area under the ROC curve for each fault type. Latency is analyzed via end-to-end response times from sensor event to anomaly alert, using repeated-measures ANOVA to compare edge, cloud, and hybrid configurations under varying workloads. Energy efficiency is assessed through Wilcoxon signed-rank tests on device-level power consumption. Robustness to concept drift is examined with time-based drift simulations and adaptive retraining experiments, analyzed through regression models and change-point detection methods. A qualitative component analyzes operator perspectives via thematic analysis of interview transcripts, exploring perceived reliability, maintainability, and trust in edge-enabled systems. Theoretical grounding draws on the Resource-Based View to interpret capability configurations and the Theory of Distributed Cognition to explain human-system interaction in fast-paced manufacturing contexts. Expected findings indicate that edge AI substantially reduces inference latency (mean reduction of 45–60 ms per event) and lowers bandwidth needs by up to 70% without compromising detection accuracy beyond a 1–2 percentage-point difference compared with cloud-based baselines, given well-curated feature sets. Edge deployments are anticipated to exhibit lower energy per inference in idle-to-moderate loads, while scaling challenges arise under peak machine states necessitating strategic offloading. The study is expected to reveal situational guidelines for model complexity, data preprocessing, and hardware selection that optimize performance under drift and sensor faults. The contribution to knowledge includes an empirically validated framework for selecting edge AI configurations in real-time industrial anomaly detection, a comparative performance map across edge, cloud, and hybrid pipelines, and methodological guidance for robust field deployments. Concluding recommendations emphasize governance of data locality, continuous monitoring for model degradation, scheduled recalibration cycles, and a decision-support toolkit enabling practitioners to tailor edge AI strategies to process characteristics, reliability requirements, and organizational constraints.