Explainable Event Extraction in Knowledge-Based Maintenance

Abstract

As an integral part of Industry 4.0, industrial maintenance is adopting smart methods, giving rise to knowledge-based maintenance. This transition requires the integration of advanced digital technologies to optimize maintenance strategies, where natural language processing (NLP) plays a crucial role. Transformer models, as a recent advancement in NLP, enable the extraction of relevant information—such as industrial events—from unstructured maintenance data, revealing hidden patterns and informing decision making. This paper conceptualizes the task of event extraction as a classification problem, applying transformer-based models and zero-shot learning. The proposed approach addresses the absence of labeled training data in maintenance sectors and also offers a solution to enhance transparency and explainability for the extracted events. In this paper, we present the performance of four transformer-based models in terms of accuracy and F1-score across four synthetic datasets available in English and German. Furthermore, we apply the Shapley value method to visualize the quantification of each token’s contribution within the text to the model’s prediction (i.e., the extracted event). We also design and develop an intuitive graphical user interface that not only facilitates user interaction but also promotes transparency. The experimental results demonstrate the effectiveness of transformer-based models for industrial event extraction, providing a powerful tool for the maintenance sector to extract valuable insights from unstructured data within historical maintenance logs, sensor data, and technical manuals. Moreover, the integration of explainability through the Shapley value method offers a deeper understanding of the model’s decision-making process, which is essential in real-world applications in manufacturing systems.