面向变载荷工况的滚动轴承多模态半监督CNN-SVM智能诊断架构研究

Research on Multi Modal Semi Supervised CNN-SVM Intelligent Diagnosis Architecture for Rolling Bearings under Variable Load Conditions

摘要: 针对变载荷工况下滚动轴承故障诊断中存在的跨工况特征分布漂移、目标域标签稀缺及深度模型缺乏物理可解释性等问题，提出一类融合多维度流形重构、CNN-SVM特征解耦与半监督伪标签学习的多模态智能诊断架构。首先，在剥离全局稳态趋势的基础上，联合时域高阶统计量、频域能量解调与连续小波变换构建高维特征映射，并利用主成分分析(PCA)构建低秩正交投影流形，在保留主体解释方差的同时滤除19.5%的尾部冗余成分。其次，构建CNN-SVM串联诊断网络，以前置CNN提取高阶拓扑表征，依凭后端SVM结构风险最小化原则确立全局决策边界。再次，嵌入基于动态置信度阈值的伪标签闭环迭代机制，自适应挖掘目标域无标签样本的隐式聚类结构以修正决策面，有效抑制小样本约束下的局部过拟合。最后，引入基于集成树基尼不纯度的特征溯源算子量化特征权重，并通过0~3 HP梯度载荷实验进行综合验证。结果表明：该架构显著缓解了分布偏移引发的收敛阻滞，极限重载下的诊断准确率达91.94%，特征重要性演化分析结果揭示了载荷递增诱发的频带能量迁移规律，诊断核心判据由基础幅频指标(如频谱质心)向表征信号复杂度的深层拓扑指标(如谱滚降、谱熵)转移。该方法不仅实现了复杂工况高精度预测，还确立了数据驱动决策与机械耗散动力学间的物理映射关系。

Abstract: Aiming at the problems of cross condition feature distribution drift, scarce target domain labels, and lack of physical interpretability of deep models in rolling bearing fault diagnosis under variable load conditions, a multimodal intelligent diagnosis architecture was proposed, which integrated multi-dimensional manifold reconstruction, CNN-SVM feature decoupling, and semi supervised pseudo label learning. Firstly, based on the detachment of the global steady-state trend, a high-dimensional feature map was constructed by combining high-order time-domain statistics, frequency-domain energy demodulation, and continuous wavelet transform. Principal component analysis (PCA) was used to construct a low rank orthogonal projection manifold, which could filter out 19.5% of tail redundant components while retaining the main explanatory variance. Secondly, a CNN-SVM serial diagnostic network was constructed to extract high-order topological representations from the pre CNN, and establish global decision boundaries based on the principle of minimizing the structural risk of the back-end SVM. Thirdly, a pseudo label closed-loop iterative mechanism based on a dynamic confidence threshold was embedded to adaptively mine the implicit clustering structure of unlabeled samples in the target domain to correct the decision surface, effectively suppressing local overfitting under small sample constraints. Finally, a feature traceability operator based on integrated tree Gini impurity was introduced to quantify feature weights, and comprehensive verification was carried out through 0~3 HP gradient loading experiments. The results showed that the architecture significantly alleviated the convergence blockage caused by distribution offset, and the diagnostic accuracy under extreme overload reached 91.94%, and the evolution analysis of feature importance revealed the frequency band energy transfer law induced by increasing load, and the diagnostic core criterion shifted from basic amplitude frequency indicators (such as spectral centroid) to deep topological indicators that could characterize signal complexity (such as spectral roll off and spectral entropy). This method not only achieved high-precision prediction of complex working conditions, but also established a physical mapping relationship between data-driven decision-making and mechanical dissipative dynamics.