An intelligent surface roughness recognition method based on vibration signal imaging and improved DenseNet

Surface roughness is a critical indicator for evaluating the machining quality of mechanical components, significantly influencing their performance and service life. Traditional online monitoring methods are often limited by the complexity of signal processing and feature engineering. This paper proposes an intelligent surface roughness recognition method based on vibration signal imaging and an improved Densely Connected Convolutional Network (DenseNet). First, vibration signals during the machining process are acquired via accelerometers installed on the tool of a computer numerical control (CNC) lathe, followed by preprocessing and wavelet denoising. Subsequently, a batch normalization approach is innovatively applied to the vibration signals, and the Relative Position Matrix (RPM) encoding technique is utilized to convert one-dimensional time-series signals into two-dimensional images, enhancing the preservation of amplitude features. On this basis, an improved DenseNet-37 deep learning model is constructed, which enhances feature extraction capability and computational efficiency by optimizing the convolutional kernel structure and network depth. Experimental results demonstrate that the combination of single-channel vibration signals, RPM image encoding, and the DenseNet-37 model achieves the best performance in the surface roughness recognition task, with precision, recall, and F1-score reaching 0.8531, 0.8495, and 0.8498, respectively, significantly outperforming comparison models (e.g., ResNet and the Inception series). This research provides an effective solution for the real-time and high-precision online monitoring of surface quality in machined workpieces.