Accurate recognition of fine-grained operating modes in traveling grate pelletizingTraveling grate pelletizing process processes (TGPPs) is critical for maintaining production stability and product quality. However, feed fluctuations cause variations in pellet residence times and the thermal environment experienced, resulting in different operating modes that are difficult to distinguish due to their subtle differences. To address these challenges, a multiscale temporal attention recognition framework is proposed. Firstly, a dynamic time-window alignmentDynamic time-window alignment strategy is introduced to match the process data sequence with the pellet quality data, ensuring that the selected process data accurately represents the actual operating mode. Secondly, a multiscale one-dimensional convolutional neural network is developed to jointly extract both short- and long-term temporal features from the process data sequence. Third, attention mechanismsAttention mechanism are integrated to selectively emphasize critical variables and time segments. Validation on industrial data from a large-scale TGPP demonstrates the recognition performance underscoring its significant potential for practical deployment.