Document Type
Workshop
Abstract
Rare event prediction is critical in industrial applications, including real-world Industry 4.0 applications. These events, defined by their low occurrence frequency, are often difficult to predict due to the skewed data distribution, which complicates modeling and evaluation. In our research, we provide a comprehensive review of current approaches to rare event prediction across four key dimensions: rare event data, data processing techniques, algorithmic approaches, and evaluation methodologies [1]. By analyzing diverse datasets with multiple modalities, including numerical, image, text, and audio, we categorize the primary challenges and present the gaps in current research. Specifically, we present three novel research contributions aimed at bridging these methodological gaps and advancing the field of rare event and anomaly prediction.
In manufacturing, rare events result in unplanned downtime, reduced equipment lifespan, and increased energy consumption. To address these challenges, we explore the role of data enrichment techniques, such as time series data augmentation, sampling, and imputation, to mitigate the scarcity and improve the representativeness of rare event datasets. Our experiments with 15 machine learning models show that data enrichment significantly enhances predictive performance, with up to a 48% improvement in the F1-score for rare event detection and prediction [2].
Anomalies, a type of rare event, represent a critical area in multiple domains where predictive capabilities are essential for enhancing operational efficiency and minimizing costs. Traditional ML methods are shown to underperform in anomaly prediction, with limitations arising from the rarity of anomalies, the lack of high-fidelity simulation data, and the complex interdependencies between sensor readings. To overcome these obstacles, we introduce the Robust and Interpretable 2D Anomaly Prediction (RI2AP) model, which simultaneously predicts when anomalies will occur and analyzes their dependencies [3]. RI2AP leverages a causal-influence framework for interpretation, making it highly valuable for domain experts. In our studies on rocket assembly simulations, RI2AP demonstrates a 30-point improvement in the F1 score compared to existing models. Real-world deployment of RI2AP for rocket part assembly has shown promising results, demonstrating the model's practicality and robustness.
Single-modality methods often fail to capture the intricate relationships within the data, limiting the precision of anomaly detection. Thus, multiple modalities are needed, with time series and images playing a prominent role. We propose a neurosymbolic AI-based fusion approach, integrating time series and image data [4,5]. Our model achieves higher performance and interpretability in anomaly prediction tasks by employing decision-level fusion, transfer learning, and knowledge-infused learning.
This approach bridges methodological gaps in industrial applications by integrating novel algorithmic strategies, robust evaluation frameworks, multi-modal data, and neurosymbolic AI, offering a comprehensive solution to the challenges of rare event and anomaly prediction.
Publication Info
Preprint version WiML workshop at NeurIPS 2024, 2024.
APA Citation
Shyalika, C., Prasad, R., Wickramarachchi, R., & Sheth, A. (2024, December 10). Towards Rare Event and Anomaly Prediction in Manufacturing: Bridging Methodological Gaps in Industrial Applications. WiML Workshop at NeurIPS 2024. WiML workshop at NeurIPS 2024, Vancouver, BC, Canada.
Rights
© The Authors, 2024