A high-fidelity noise reduction method for multi-sensor signals and an intelligent prediction method for surface roughness

Abstract The formation of surface morphology is influenced by multiple factors, thereby necessitating the use of data from multiple sensors for predicting surface roughness. Critical information related to surface roughness is often embedded in high-frequency, transient, and low-energy signal features. Therefore, signals used for prediction should exhibit high energy retention and feature preservation rates. However, traditional noise reduction methods often misclassify such energy and features as noise and filter them out. To overcome the loss of effective energy and features during the denoising of multi-sensor signals, this paper proposes a high-fidelity intelligent prediction method for surface roughness. First, a joint noise reduction approach based on ICEEMDAN and wavelet transform is employed. ICEEMDAN decomposes the cutting signal into a series of intrinsic mode functions (IMFs), utilizes energy entropy to intelligently determine the noise ratio, and incorporates an improved semi-soft threshold function for differentiated processing, thereby yielding a denoised cutting signal with high retention rates. Subsequently, the sparrow search algorithm is applied to optimize the convolutional neural network (CNN) for predicting surface roughness, which mitigates errors caused by manual parameter adjustment. The proposed method achieves a root mean square error of 0.01258 μm and a coefficient of determination of 0.962, outperforming traditional models such as BP, SVM, and standard CNN.