在现代化工业生产中，数控机床是重要的生产工具，所以数控机床能否正常地运行对于生产过程的稳定性起着决定性的作用。然而，数控机床整机在长期的高强度运行过程中，容易因为累积性能的损耗而出现渐变故障，渐变故障发展到一定规模就会严重干扰着机床的正常运行。所以准确预测渐变故障的发生过程是保证数控机床生产能力的关键所在，本研究以数控机床中的元动作单元为研究对象，以预测数控机床元动作单元的渐变故障为研究目标，展开研究。
  首先，对数控机床的蜗杆转动元动作单元进行了故障模式分析，建立了蜗杆转动元动作单元渐变故障振动信号采集实验平台，通过实验平台采集元动作单元的振动信号。为了解决从传感器采集到的原始信号受到噪声影响的问题，提出了基于包络熵的双层分解振动信号降噪预处理方法，该方法以包络熵值为依据通过双层分解将原始信号分解为若干分量，然后依据皮尔逊相关系数进行重构得到重构信号，减少了大多数噪音对于渐变故障预测的影响，在此基础上，提取出重构信号的时域和频域特征。 
  其次，对提取出的故障信号特征进行特征选择解决特征指标数据维度过高的问题，运用ReliefF算法对每个特征指标与目标变量之间的相关性进行分析，并对这些特征指标进行赋权。提出了COA-SVM故障分类算法，以不同权重为阈值将不同特征指标以COA-SVM故障分类算法的分类准确率为验证，进行多次重复试验，构建最优渐变故障预测特征指标集。
  再次，在构建出最优渐变故障预测特征指标集的基础之上，为了实现对数控机床元动作单元渐变故障的精准预测，构建基于多折叠交叉验证的Stacking 架构集成学习预测模型。该模型以GBDT、LightGBM、XGBoost、随机森林和LSTM五种算法为基学习器，将特征数据集以5折交叉验证的方法进行学习预测，利用集成学习架构综合基学习器的预测结果实现对数控机床元动作单元渐变故障的准确预测。
  最后，设计元动作单元渐变故障预测方法的有效性实验，通过蜗杆转动元动作单元试验台提取元动作单元的全渐变故障周期振动信号，按照所提出方法对元动作单元渐变故障进行预测，并通过消融实验对模型中基学习器的作用进行验证，然后通过与其他模型的对比实验验证该方法的预测性能。经试验验证，本研究提出的方法在元动作单元渐变故障预测方法展现出良好的性能。

In modern industrial production, CNC machine tools are essential production equipment, and their proper functioning plays a decisive role in ensuring the stability of the production process. However, during prolonged high-intensity operation, CNC machine tools are prone to gradual faults due to cumulative performance degradation. When these gradual faults reach a certain scale, they severely disrupt the normal operation of the machine tools. Therefore, accurately predicting the occurrence of gradual faults is key to maintaining the production capacity of CNC machine tools. This study focuses on the meta-action units within CNC machine tools, with the research objective of predicting gradual faults in these units.

First, a fault mode analysis was conducted on the worm gear rotation meta-action unit of the CNC machine tool, and an experimental platform for collecting vibration signals of gradual faults in the worm gear rotation meta-action unit was established. Vibration signals from the meta-action unit were collected through this platform. To address the issue of noise interference in the raw signals acquired from sensors, a double-layer decomposition vibration signal denoising preprocessing method based on envelope entropy was proposed. This method decomposes the original signal into several components through double-layer decomposition based on envelope entropy values and then reconstructs the signal using the Pearson correlation coefficient, thereby reducing the impact of most noise on gradual fault prediction. On this basis, time-domain and frequency-domain features of the reconstructed signal were extracted.

Second, feature selection was performed on the extracted fault signal features to address the issue of high dimensionality in the feature indicator data. The ReliefF algorithm was used to analyze the correlation between each feature indicator and the target variable, and weights were assigned to these feature indicators. A COA-SVM fault classification algorithm was proposed, using different weights as thresholds to select feature indicators. Repeated experiments were conducted with the classification accuracy of the COA-SVM algorithm as the validation criterion to construct an optimal feature indicator set for gradual fault prediction.

Third, based on the constructed optimal feature indicator set for gradual fault prediction, a Stacking architecture ensemble learning prediction model based on multi-fold cross-validation was built to achieve precise prediction of gradual faults in CNC machine tool meta-action units. This model employs five algorithms—GBDT, LightGBM, XGBoost, random forest, and LSTM—as base learners. The feature dataset was learned and predicted using a 5-fold cross-validation method, and the ensemble learning architecture was utilized to synthesize the prediction results of the base learners, enabling accurate prediction of gradual faults in CNC machine tool meta-action units.

Finally, an experiment was designed to validate the effectiveness of the proposed gradual fault prediction method for meta-action units. Vibration signals covering the entire gradual fault cycle of the meta-action unit were extracted from the worm gear rotation meta-action unit test bench. The proposed method was applied to predict gradual faults in the meta-action unit, and ablation experiments were conducted to verify the role of the base learners in the model. Comparative experiments with other models were then performed to validate the prediction performance of the proposed method. Experimental results demonstrated that the method proposed in this study exhibits excellent performance in predicting gradual faults in meta-action units.