Industrial-Grade Smart Troubleshooting through Causal Technical Language Processing: a Proof of Concept

The paper introduces a causal diagnosis approach for troubleshooting industrial environments using technical language from Return on Experience (RoX) records. The method employs vectorized linguistic knowledge from a Large Language Model (LLM) and embedded failure modes and mechanisms of industrial assets. A causality-aware retrieval augmented generation system is presented, demonstrated experimentally on a real-world Predictive Maintenance setting. The authors aim to anticipate and mitigate the complex multifaceted problem of asset degradation by introducing quality checks in the manufacturing process and inspection actions in preventive maintenance schedules. The paper emphasizes the importance of a common development standard like ISO 13374 for breaking down complex problems into manageable modules, enhancing project success and interpretability. The method uses the Structural Causal Model (SCM) to represent causal links among variables, capturing directed associations, and the Causal Bayesian Network (CBN) for learning functional associations, treating variables as random and statistically describing their conditional probability distributions. The paper highlights the need for improvement in the utilized causal technology to meet robustness challenges in increasingly complex industrial scenarios. The authors propose a system that uses a "BERTopic" LLM for text embedding, combining MiniLM, UMAP, and HDBSCAN to create a numerical representation of the text. The system employs a Health Assessment block to tackle the challenge of confounding bias by using causal inference techniques. It leverages a discrete causal Bayesian network designed based on RoX data to diagnose input anomalies and estimate potential solutions. The paper discusses an industrial-grade smart troubleshooting system using causal technical language processing, emphasizing the importance of a common development standard like ISO 13374 for breaking down complex problems into manageable modules. The system employs a "BERTopic" Large Language Model for text embedding, combining MiniLM, UMAP, and HDBSCAN to create a numerical representation of the text. The Health Assessment block tackles the challenge of confounding bias by using causal inference techniques. It leverages a discrete causal Bayesian network designed based on RoX data to diagnose input anomalies and estimate potential solutions. The system calculates the conditional probability of root causes, providing an ordered list of potential variables and their probabilities. For solution generation, an intervention is performed to predict the solution's impact on other variables, accounting for causal model assumptions. This approach reduces estimation bias through an adjustment formula. The system is designed to be helpful, positive, and unbiased, with an average accuracy of root cause classification scores over 80%. The data observation includes a specific instance with a root cause of "Part physically damaged" and a detailed solution provided by the system. The paper discusses three key areas for enhancing predictive maintenance in the Smart Troubleshooting setting: vector databases, transportability, and counterfactual analysis. Vector databases aim to improve the granularity of linguistic representation in causal Bayesian networks by embedding unstructured text data into a large vector space. Transportability addresses the need to generalize empirical findings across different environments or populations, considering the specificities of text data in each setting. Counterfactual analysis explores potential outcomes at the individual level, driven by hypothetical speculations over data that may contradict the facts. The paper outlines a comprehensive top-down troubleshooting approach grounded in causal inference principles, designed to be industrially applicable. The focus is on creating a distributed representation of linguistic features through processing technical language, aiming to achieve unbiased causal diagnostics and solutions. This approach is demonstrated in the Predictive Maintenance domain, with results suggesting potential for future research in evaluating generative models across various industrial settings.