An Efficient PCA-BP Fault Prediction Scheme for Health Warning of Mechanical Equipment

Abstract

In intelligent manufacturing, redundant monitoring features and complex nonlinear relationships between faults and operating parameters limit the performance of traditional fault prediction models. To address these issues, this paper proposes a PCA-BP-based fault prediction method. Z-score standardization is first applied to normalize feature scales, and PCA is then used to extract principal components with a cumulative explained variance of at least 90%, reducing data redundancy and input dimensionality. A single-hidden-layer BP neural network with 40 hidden neurons is trained using the selected principal components, while logistic regression serves as the baseline model. Experimental results show that the optimal PCA-BP model achieves an $\mathbf{R}^{\mathbf{2}}$ of 0.2518 and an RMSE of 0.4325, outperforming logistic regression in prediction accuracy and nonlinear fitting capability. The proposed method effectively captures fault evolution characteristics and provides an efficient solution for real-time equipment health monitoring and predictive maintenance in intelligent manufacturing.