Abstract
With advancements in artificial intelligence (AI), there is a growing need to bridge the gap between multimodal learning capabilities and the availability of high-quality datasets for tool wear estimation. Industrial scenarios frequently require domain-specific knowledge, specialized datasets, and efficient deployment on resource-constrained edge devices that demand minimal memory, low latency, and optimized computational performance. While there has been a shift from unimodal sensor-based approaches to multisensory, multimodal strategies, this transition remains in its early stages. Developing feature extraction methods, multimodal fusion techniques, and correlation analysis frameworks is crucial for improving tool wear prediction models. Existing multimodal open-source datasets have several limitations in addressing these challenges:•They are often restricted to a specific set of data modalities, limiting adaptability.•They primarily feature general-purpose objects, which are not well-suited for industrial applications requiring specialized domain knowledge.•They lack support for lightweight models designed for real-time processing on edge devices.•They lack in-depth documentation or dedicated data loaders, limiting reproducibility. To bridge this gap, we introduce the Nonastreda Multimodal Dataset for efficient tool wear state monitoring. The dataset models the multimodal nature of tool wear progression in industrial milling processes, integrating nine data modalities. It comprises 512 samples, each containing RGB images of the shaft milling tool, workpiece, and material chip, along with three scalograms and three spectrograms derived from force signals. Data collection was performed using ten milling tools in an industrial production environment. The dataset is designed to support classification tasks (sharp, used, dulled) and regression tasks predicting three target variables: flank wear (µm), gaps (µm), and overhang (µm). Each sample can be analyzed independently or as part of a temporally correlated sequence. Accompanying scripts for data processing and analysis are available in the repository.