2025

Streltsov, Oleksii; Terziyan, Vagan; Vitko, Oleksandra

In the context of Industry 4.0 and smart manufacturing, effective machine learning models must not only predict outcomes but also understand the underlying causal relationships for tasks such as predictive and prescriptive maintenance. However, traditional Convolutional Neural Networks (CNNs) are often limited by their inability to explicitly capture and utilize these causal relationships within image data, which can lead to suboptimal performance and limited interpretability in industrial applications. This paper addresses these limitations by advancing our prior work on Causality-Aware Convolutional Neural Networks (CA-CNNs), which are designed to identify and utilize hidden causal factors within images. Through a series of empirical validations, we demonstrate that: CA-CNNs outperform conventional CNNs in classification tasks by employing causality maps; pre-trained CNNs enhance the computation of these maps, while weighting features based on causality maps leads to superior performance; the “Lehmer” method improves noise tolerance and adversarial robustness, and the use of “causality shadows” derived from saliency maps increases classification accuracy. We further explore integration strategies, including Knowledge-Informed Machine Learning schemas, to merge the discovered features effectively with their causal relationships. Our findings highlight the significant potential of CA-CNNs and their modifications in industrial applications, offering improved predictive capabilities, and explainability in smart manufacturing. See presentation slides: https://ai.it.jyu.fi/ISM-2024-CA-CNN-Experiments.pptx Find code and details on experiments: https://github.com/Alexiush/weak-causality-and-causal-disposition-in-images