Methods for Enhancing Industrial Applications with Explainable Artificial Intelligence

With the growing adoption of Artificial Intelligence (AI) and the successful application of deep learning methods in various domains, AI, particularly deep learning, is increasingly influencing people’s lives. Despite the continuous performance advancements and success of AI models, many manufacturers still hesitate to deploy AI on a larger scale, as failures due to a black-box AI model in production can cause significant financial damage and potentially harm workers. The emerging topic of Explainable Artificial Intelligence (XAI) offers approaches and algorithms that introduce transparency into black-box models by producing explanations for an AI System’s inner workings and decisions. Specifically, in industrial settings, where complex problems and decision-making processes are widespread, enabling transparent automation and decision support is crucial. However, while research in XAI is trending, applying XAI in industrial settings is challenging. For many data types (e.g., images or tabular data), XAI methods are well-studied. Nevertheless, support for time series, ubiquitous in industrial settings, is missing. Furthermore, to use any XAI method in
... mehr
deployment, understanding the explainers’ quality, strengths, and weaknesses is vital to prevent ambiguous and incorrect explanations. Well-performing XAI methods can help users understand the reasons behind a deep learner’s prediction and enable the recognition of spurious correlations learned by a deep learner or missing information in the collected data. However, reverting incorrect predictions and learned patterns is not in the scope of XAI. Nevertheless, reverting such is especially important in industrial settings, where only a limited amount of (often) noisy data is available. Explanations and explanation correction can potentially provide the opportunity to include domain knowledge of users and enable a deep learner to infer the missing context and close this gap. The main contributions of this thesis are to facilitate and enable a more widespread use of XAI in industrial settings by providing methods, frameworks, and in-depth evaluations addressing the application obstacles. Individual contributions of this thesis are summarized as follows:
1. An extensible framework enabling unified access to time series explainers, including a new counterfactual explainer considering various time series transformation mechanisms and thereby enabling plausible explanations on uni- and multivariate time series.
2. A framework and in-depth evaluation of time series explainers, a metric measuring causal coherence of counterfactual approaches and corresponding benchmark datasets.
3. A methodology combining continuous, interactive, and explanatory interactive machine learning to enable human supervised transfer learning in industrial contexts.
The contributions are evaluated in accordance with predefined requirements and to state-of-the-art setups. The outcome (i) allows facilitated access to time series explainers and the generation of more plausible counterfactual explanations, helping to bridge the gap between complex models and contrastive explanations; (ii) indicates the need for additional research specifically for multivariate time series explainers and the generation of causal coherent counterfactual explanation; and (iii) shows that explanation based AI-expert alignment enables continuous learning under human supervision.