Abstract Accurate diagnosis of tool wear is essential for enhancing both product quality and production efficiency. However, most existing diagnostic methods for tool wear fail to integrate both monitoring and prediction, with the majority of prediction approaches focusing on short-term prediction, often leading to significant cumulative errors. This study proposes a real-time tool wear monitoring and Multi-step Forward Tool Wear Prediction(MFTWP) method based on multi-information fusion. First, an automated data acquisition platform is established to collect milling tool wear images and multi-information. Second, the Segmenting Objects by Locations v2(SOLOv2) model is trained to segment the wear area, and the maximum flank wear width(VBmax) of the milling tool is obtained to provide labeled data for monitoring and prediction models. Simultaneously, multi-domain features from the multi-sensor data are extracted, followed by fusion and post-processing. Kernel Principal Component Analysis(KPCA) is then applied to extract the most representative and interpretable input features for the monitoring model. Subsequently, the Kolmogorov-Arnold Networks(KAN) monitoring model is constructed to establish the relationship between multi-sensor features and real-time tool wear values. The output sequence is used as input to the Transformer prediction model to determine the tool wear sequence for future time steps. Following this, a MFTWP pattern is proposed, along with a Real-time Correction Strategy(RCS) to address the issue of cumulative errors. Finally, A large number of experiments are conducted, demonstrating the outstanding performance and real-time efficiency of the proposed method.
