Surface roughness prediction in boring of TC4-DT with multi-sensor data and machining parameters

Surface roughness in boring process is a critical indicator of product quality. With the aim of maximizing quality, there has been a widespread adoption of computer-controlled machine tools that incorporate realtime prediction of surface roughness. However, existing predictive models often rely on truncated multi-sensor data due to computational limits, potentially overlooking essential sensor inputs crucial for accurate prediction. Moreover, while machining parameters like spindle speed, feed rate and cutting depth directly influence surface roughness, their combined use with multi-sensor data in prediction models is rarely explored in existing literature. To bridge these gaps, this paper introduces patch time series CNN-LSTM feature mixed network (PatchCL-MixNet), a novel deep learning approach for predicting surface roughness in boring by integrating multi-sensor data with machining parameters. The method involves segmenting extensive multi-sensor time series data into smaller patches, thereby reducing computational demands while retaining significant temporal information. A specialized encoder-decoder structure, comprising a CNN-LSTM encoder for temporal pattern identification across patches and a channel-mix decoder for feature fusion from both sensor data and machining parameters, is developed. Experimental validation conducted through rigorous boring experiments on TC4-DT titanium alloy demonstrates that our PatchCL-MixNet method outperforms existing models, achieving a 21% improvement in mean absolute percentage error. Furthermore, an ablation study confirms the model’s effectiveness in leveraging both machining parameters and multi-sensor data for enhanced surface roughness prediction.