Towards Sustainable Manufacturing: A Reliability Prediction Method for Remanufacturing Processes Based on Causal Knowledge Enhancement Under Small Sample Characteristics

Remanufacturing is a core pathway for value-added utilization of waste resources and promoting sustainable manufacturing. Its process reliability directly determines product service safety and recycling benefits. Given that remanufacturing targets high-value personalized end-of-life components, stringent reliability requirements are imposed to avoid economic losses from “secondary scrapping”. However, the small-batch production mode results in limited data, and complex causal interactions among failure morphologies, process parameters, and reliability render traditional models prone to capturing spurious correlations, leading to prediction deviations. Therefore, a method is proposed that extracts causal knowledge from historically similar process samples to guide reliability prediction. First, a component process similarity assessment and data fusion method based on process symbol entropy is proposed. The assessment of dissimilarity between process routes is conducted by quantifying global structural consistency through the Longest Common Subsequence (LCS) and local distributional variance through positional entropy. Furthermore, the reliability data of similar processes are fused to expand the sample set. Second, a knowledge modeling method for mapping the influencing factors of process reliability based on causal gradient derivation is proposed. The PC algorithm, combined with a bootstrap strategy, is applied to extract the causal structural relationships among influencing factors from the training portion of the augmented dataset. Subsequently, causal effects are estimated and converted into causal gradients, which quantitatively reveal the network transmission pathways and influence intensities of various factors on process reliability. Finally, a process reliability prediction model based on causality knowledge supervision is established. The prior causal gradient is embedded as a regularization term in the training of the Multilayer Perceptron network. By adjusting the direction of network parameter updates, the model’s predictive accuracy is enhanced. Taking the machine tool spindle as the research object, the proposed method is validated under small-sample conditions. The results demonstrate significant improvements in both R2 and RMSE compared with standard models. The similarity-based data fusion raises the R2 of the conventional MLP from 0.7145 to 0.9012, and the causal gradient regularization further increases it to 0.9320 with the RMSE reduced from 0.3483 to 0.2548.