基于时频脊线提取与改进稀疏分量分析的RV减速器复合故障盲分离

针对农业机器人的核心传动部件旋转矢量（rotate vector，RV）减速器的故障特征提取问题，该研究提出一种基于时频图像脊线提取与改进稀疏分量分析相结合的RV减速器复合故障盲提取方法。首先利用提出的时频图像脊线提取（ridge extraction from time-frequency images，RETF）方法同步截取机械臂恒速时段的观测振动信号，然后利用提出的sinC函数改进形态滤波（sinC-morphological filtering，SMF）算法、密度峰值聚类（density peak clustering，DPC）算法和正交匹配追踪（orthogonal matching pursuit，OMP）算法相结合的盲源分离方法（SMF-DPC-OMP）实现平稳信号复合故障的分离提取。以sinC函数作为新的结构元素构造平均组合形态滤波器，对恒速时段的振动信号进行形态滤波处理，以提升信号的冲击特性以及稀疏性；利用DPC估计稀疏信号的混合矩阵，进而构建传感矩阵，并结合OMP在频域完成分离源信号重构，最后对重构的时域信号进行快速傅里叶变换完成故障识别。试验台采集的RV减速器的太阳轮和行星轮磨损复合故障信号的分析结果显示，该算法能有效实现RV减速器复合故障的盲分离。RETE算法能够在变转速工况导致时频图较为模糊的情况下，识别出RV减速器的运动状态；SMF-DPC-OMP算法能够在故障源数目未知的情况下，有效完成复合故障的盲分离。与已有方法相比，SMF-DPC-OMP方法能够节省约75%的时间成本，频谱更为简洁，能够抑制精细侧频和干扰分量，适用于关节型农业机器人RV减速器复合故障盲分离，对生产实际中的故障特征提取具有一定的参考意义。

Blind fault separation of RV reducer using time-frequency ridge extraction and improved sparse component analysis

The rotary vector (RV) reducer is one of the most important transmission components inside the joints of robots. The high precision motion control of the robot can be realized in the normal operation and service life of agricultural robots. However, the complex structure of the RV reducer cannot fully meet the variable conditions in actual work and the harsh working environment. Mechanical, cracks, and pitting faults can often occur at the same time or successively in actual operation, due to lubrication, manufacturing errors, and unreasonable forces. The signals collected by the sensor are also the mutual coupling of multiple fault sources. Therefore, it is necessary to explore the composite fault diagnosis of the RV reducer. In this study, a blind fault separation was proposed from the compound signals of the RV reducer using time-frequency image ridge extraction and improved sparse component analysis. The fault feature of the RV reducer was then extracted from the core transmission component of agricultural robots. The composite fault diagnosis of the RV reducer was finally realized with the unknown number of fault sources under the reciprocating motion of the joint arm and the time-varying speed of the agricultural robot. Firstly, the ridge extraction from time-frequency images (RETF) was used to synchronously intercept the vibration signals of the manipulator in the constant speed period. Then, the blind source separation (SMF-DPC-OMP) was combined with the sinC function to improve morphological filtering (SMF), density peak clustering (DPC), and orthogonal matching pursuit (OMP), in order to separate and extract the composite faults of stationary signals. The sinC function was taken as a new structural element to construct the average combined morphological filter. The vibration signal in the constant speed period was subjected to morphological filtering for better impact characteristics and sparsity of the signal. The DPC was used to estimate the mixing matrix of sparse signals, and then the sensing matrix was constructed. The separation source signal was reconstructed to combine with the OMP algorithm in the frequency domain. Finally, the reconstructed time-domain signal was subjected to the fast Fourier transform for fault identification. The test bench was utilized to collect the composite fault signals of the sun wheel and planetary wheel wear of the RV reducer. The results show that the blind separation was effectively realized on the composite faults of the RV reducer. The RETE was used to identify the motion state of the RV reducer, when the time-frequency diagram was blurred, due to the variable speed condition. The SMF-DPC-OMP effectively completed the blind separation of composite faults when the number of fault sources was unknown. The SMF-DPC-OMP saved about 75% of the running time cost than before. A more concise spectrum was achieved to better suppress the fine side frequency and interference components. It is suitable for the composite fault diagnosis of the RV reducer in the articulated industrial robot. The finding can provide a strong reference for the fault feature extraction in the actual production.