An Investigation of Topological Machine Learning in Industrial Domain: Application to Tyre Noise Prediction

Small dataset sizes, high variability, and complex structured measurements often limit the direct applicability of conventional Machine Learning (ML) in real-world industrial scenarios. In many industrial contexts, acquiring large, high-quality datasets is impractical due to production constraints, measurement costs, or proprietary restrictions. This creates a need for approaches that can extract meaningful information from limited data and provide a deeper understanding of the underlying structures.In this work, we propose to use Topological Machine Learning (TML) as a novel analytical perspective, integrating traditional ML with Topological Data Analysis (TDA). TDA captures persistent topological features via persistence diagrams (PDs), revealing geometric and morphological patterns in a robust and compact form. By transforming PDs into vectorized representations compatible with ML models, TML provides a new way to examine structural properties of complex industrial data. We illustrate this approach using a dataset of 311 profilometric tyre tread images to predict tyre-generated noise, which is a critical factor for acoustic comfort and vehicle safety. Experiments demonstrate that topological descriptors encode highly informative structural patterns linked to acoustic emission. Of the representations tested, Persistence Images (PI) achieved the greatest predictive accuracy, with the Random Forest model outperforming other regressors in terms of mean absolute error. Results highlight the ability of TML to extract information from complex textures, enabling noise prediction in small-sample industrial scenarios.