Surface roughness identification in turning using motor current signals and optimized envelope spectrum features

Abstract In machine tool turning, accurate identification of surface roughness is crucial for product quality stability. Conventional sensor-based methods that capture images, vibrations, and acoustic emission signals are costly and vulnerable to environmental interference. By contrast, motor current signature analysis is cost-effective, operationally stable, and reliable. However, high-frequency noise, power line interference, and harmonic components in current signals hinder surface roughness identification and limit the recognition accuracy. To address these challenges, this study introduces a novel methodology for identifying surface roughness in machine tool turning operations based on spindle motor current signals. The approach begins with preprocessing the raw current signals through wavelet packet denoising, which effectively suppresses the noise interference. Subsequently, feature reconstruction (FR) techniques were employed to transform the denoised signals into the frequency domain, enabling a deeper exploration of the inherent relationships between the current signal characteristics and surface roughness patterns. Building on this foundation, the power spectral density (PSD) function was utilized to further refine the envelope feature spectrum, thereby selecting the optimal frequency bands that best represented the surface roughness attributes. Concurrently, Pearson correlation coefficients were incorporated to analyze the correlation between these frequency bands, validating the precision and reliability of the extracted features. Finally, the optimal feature spectrum was integrated with a convolutional neural network (CNN) architecture to establish an intelligent recognition model capable of predicting surface roughness in turning processes. The experimental findings demonstrated that the envelope feature spectrum optimization technique, which was based on PSD analysis, extended the capabilities of conventional FR methods and significantly enhanced the CNN-based identification performance. The proposed model achieved 98.61% classification accuracy, an F1-score of 0.964 63, and a mean deviation distance of 0.16 during validation testing, thereby confirming the method’s technical efficacy and measurement accuracy.