Abstract: The FastICA analysis method is optimized by adopting fast decent method and successive over relaxation factor to pre-process the initial value of de-mixing matrix. The development aims to improve the convergence property and avoid the uneven phenomenon. Firstly, the better solutions to de-mixing matrix is selected by the fast decent method with the fastest convergence rate. Then, it introduces successive over relaxation factor to limit the decent property of the objective function according to the given norm. The final initial value is obtained by these two steps, and the combination of these two steps can ensure algorithm convergence. The optimized FastICA algorithm is applied to fault diagnosis of bearings, according to the iteration time and curve shape of the convention. It is testified that the optimized method is superior to the convention one in terms of stability and speed, and inherits the separation performance of convention FastICA method.

Key words: FastICA algorithm, initial value sensitivity, fast decent method, successive over relaxation factor, bearing fault diagnosis

摘要： 为提高FastICA算法的收敛平稳性和速度，克服FastICA算法对初始值选取敏感的问题，提出在最速下降法中引入松弛因子优化FastICA算法中解混矩阵初始值的方法。首先，按最速下降法负梯度原理确定初始值目标函数最速收敛方向，以最快速度选取靠近目标函数解的粗优值；然后，通过引入松弛因子[αk]，限制目标函数的下降性质，促使其进入牛顿迭代法收敛区域，最终达到收敛。将优化后的FastICA算法应用于轴承故障诊断中，根据多次仿真次数下迭代时长及时长的波动趋势验证优化FastICA算法在平稳性和速度方面优于传统FastICA算法，且不影响FastICA算法的分离性能，能准确诊断出轴承的故障类型。

关键词: FastICA算法, 初值敏感, 最速下降法, 松弛因子, 轴承故障诊断