Traditional manufacturing systems face significant challenges in detecting operational anomalies due to the absence of advanced sensor networks and intelligent machinery commonly associated with Industry 4.0. Existing solutions often rely on sophisticated, interconnected infrastructures, which are not feasible in conventional settings. This paper introduces a novel methodology for anomaly detection tailored specifically for traditional manufacturing environments, addressing the gap in cost-effective monitoring solutions. The proposed approach models manufacturing systems as complex temporal networks, where each machine or process is represented as a node and job flows between machines form the network edges over time. The novelty of this method lies in the combination of dynamic network theory with unsupervised machine learning. Statistical features extracted from the temporal networks are processed through dimensionality reduction techniques, specifically Principal Component Analysis (PCA) and Deep Neural Autoencoders, to reduce feature complexity while preserving essential information. The reduced feature sets are then analysed using multiple unsupervised anomaly detection algorithms, including Isolation Forest, One-Class Support Vector Machine (OC-SVM), and Local Outlier Factor (LOF). This approach does not require significant infrastructure upgrades, making it suitable for traditional manufacturing plants while still aligning with Industry 4.0 paradigms. By using only normal job flow data, it provides a cost-effective solution where anomalous data is scarce. The results demonstrate that Local Outlier Factor and Isolation Forest, when combined with Autoencoder-based feature reduction, achieved an F1-score exceeding 84%, with precision close to 99% and recall at 74%. This strong performance underscores the methodology's potential for real-world manufacturing environments, bridging the gap between traditional settings and modern Industry 4.0 paradigms.
Anomaly detection in manufacturing systems with temporal networks and unsupervised machine learning
Mattera R.;
2025
Abstract
Traditional manufacturing systems face significant challenges in detecting operational anomalies due to the absence of advanced sensor networks and intelligent machinery commonly associated with Industry 4.0. Existing solutions often rely on sophisticated, interconnected infrastructures, which are not feasible in conventional settings. This paper introduces a novel methodology for anomaly detection tailored specifically for traditional manufacturing environments, addressing the gap in cost-effective monitoring solutions. The proposed approach models manufacturing systems as complex temporal networks, where each machine or process is represented as a node and job flows between machines form the network edges over time. The novelty of this method lies in the combination of dynamic network theory with unsupervised machine learning. Statistical features extracted from the temporal networks are processed through dimensionality reduction techniques, specifically Principal Component Analysis (PCA) and Deep Neural Autoencoders, to reduce feature complexity while preserving essential information. The reduced feature sets are then analysed using multiple unsupervised anomaly detection algorithms, including Isolation Forest, One-Class Support Vector Machine (OC-SVM), and Local Outlier Factor (LOF). This approach does not require significant infrastructure upgrades, making it suitable for traditional manufacturing plants while still aligning with Industry 4.0 paradigms. By using only normal job flow data, it provides a cost-effective solution where anomalous data is scarce. The results demonstrate that Local Outlier Factor and Isolation Forest, when combined with Autoencoder-based feature reduction, achieved an F1-score exceeding 84%, with precision close to 99% and recall at 74%. This strong performance underscores the methodology's potential for real-world manufacturing environments, bridging the gap between traditional settings and modern Industry 4.0 paradigms.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
