The transient impacts with sideband modulation caused by some fault of gearbox are the technical basis for fault diagnosis, which will be inevitably interfered by heavy background noise distributed in complex modulation frequency bands. Generally, only the principle components under the selected scale is remained and analyzed as the evidence of fault diagnosis, while some crucial features spread in other scales are ignored. Specially, the selected signal still has lots of in-band noise interference. Motivated by these issues, a new adaptive multi-mode manifold learning (AM²ML) method is proposed to enhance the useful gearbox features distributed complex scales with the in-band noise suppressed. Firstly, a series of mode components are obtained by adaptive variational mode decomposition, where the optimal decomposition level is automatically achieved by the k-value. Time-frequency manifold learning is then respectively employed to mine their corresponding potential structural characteristics. And a reconstructed signal contained multi-scale features are represented via the proportion weights of the correlation coefficients. Therefore, the denoised signal of each manifold mode will be rebuilt by phase preserving and a series of inverse transform while the in-band noise is suppressed. With the weight coefficients of each mode, the final multi-scale features are synthesized. As the result shows, the sub-band energy ratio is introduced to evaluate the effectiveness of the method. The sub-band energy ratio of AM²ML method is about 3 times of VMD method, and about 2 times of fast-kurtogram method. It can be seen that the multiple modes can be adaptively decomposed with the complex scale structural characteristics enhanced by manifold learning, and the expected characteristics distributed in all scales can be well rebuilt via a dynamic weight reconstruction in a self-learning way. The effectiveness of the proposed AM²ML method is verified by the self-made gearbox.

由齿轮箱故障引起的边带调制的瞬态影响是故障诊断的技术基础，不可避免地会受到复杂调制频段中分布的大量背景噪声的干扰。通常，仅保留选定尺度下的主成分并进行分析，以作为故障诊断的依据，而忽略其他尺度下散布的某些关键特征。特别地，所选信号仍然具有很多带内噪声干扰。受这些问题的影响，新的自适应多模式流形学习（AM ²提出了一种ML）方法来增强有用的变速箱特征，即在复杂的比例尺上分配带内噪声的情况。首先，通过自适应变分模式分解获得一系列模式分量，其中最佳分解级别由k值自动获得。然后分别采用时频流形学习来挖掘其相应的潜在结构特征。并通过相关系数的比例权重来表示包含多尺度特征的重构信号。因此，在抑制带内噪声的同时，将通过相位保持和一系列逆变换来重建每个流形模式的去噪信号。利用每种模式的权重系数，可以合成最终的多尺度特征。结果显示 引入子带能量比来评估该方法的有效性。AM的子带能量比² ML方法约为VMD方法的3倍，约为快速峰图方法的2倍。可以看出，通过流形学习可以增强复杂模式的结构特征，从而适应性地分解多种模式，并且可以通过动态权重重构以自学习的方式很好地重建各个尺度下分布的预期特征。自制变速箱验证了所提出的AM ² ML方法的有效性。