基于PSO-BFA和改进Alexnet的滚动轴承故障诊断方法

现实中滚动轴承的工况复杂易变,无法有效地对其进行故障诊断。对此,提出一种基于粒子群优化的细菌觅食(Particle Swarm Optimization and Bacterial Foraging Algorithm,PSO-BFA)和改进Alexnet(第二代卷积神经网络)的滚动轴承故障诊断方法。该方法将Alexnet的结构简化,并分别在其前两层池化层之后添置局部归一化层以降低训练成本;将以小批量样本softmax的交叉熵为损失函数,按Adam迭代优化法小样本、少迭代次数训练改进Alexnet后的变负荷样本诊断精度设计为适应度函数,并结合PSO中粒子移动速度的更新方法更新BFA中细菌的游动方向来寻找改进Alexnet的结构等相关参数;根据PSO-BFA所得的参数,以相同的训练方法大样本、多迭代次数训练改进Alexnet,实现复杂工况下滚动轴承多状态故障诊断。实验结果表明所提出的方法对复杂工况下滚动轴承16种故障状态的诊断是可行的,且有更高的诊断精度、更好的抗干扰和泛化性能。

Rolling bearing fault diagnosis method based on improved Alexnet and PSO-BFA

Actual working conditions of rolling bearing are complex and varying, and bearing faults can’t be diagnosed effectively. Here, a novel rolling bearing fault diagnosis method based on improved Alexnet, i.e., the second generation of convolution neural network, and particle swarm optimization and bacterial foraging algorithm (PSO-BFA) was proposed. Firstly, this method was used to simplify Alexnet structure, and two local normalization layers respectively followed Alexnet’s first two pooling layers and were added to reduce training cost. Then, small batch samples softmax’s cross entropy was taken as a loss function, and Adam iterative optimization method was used to train the improved Alexnet with a small number of samples and a few iterations. The diagnosis accuracy of variable-load samples was designed as swarm intelligence algorithm’s fitness function, it was combined with the updating method of particle movement velocity in PSO to update bacterial swimming direction in BFA, and search improved Alexnet’s structure parameters. Finally, according to parameters obtained with PSO-BFA, the same training method was used to train improved Alexnet with large samples and multiple iterations, and realize rolling bearing multi-state fault diagnosis under complex working conditions. Test results showed that the proposed method is feasible for diagnosing rolling bearing’s 16 fault states under complex working conditions; it has higher diagnostic accuracy, better anti-interference and generalization performances.