PI-DBN在滚动轴承故障诊断中的研究

Research on PI-DBN in Fault Diagnosis of Rolling Bearings

针对DBN存在的参数冗余程度高、计算量大、训练时间长等问题,提出一种脉冲强度剪枝算法(PI-DBN)。该方法在无监督预训练阶段对深度信念网络(Deep Belief Network,DBN)进行网络剪枝操作,优化网络结构,去除无效网络节点;以脉冲强度变化大小删除不具有重要意义的权重,提高DBN收敛和计算速度,同时保留继承数据特征的信息。结果表明,经过脉冲强度剪枝后权值信息在低维数据集下能够较好地表征数据的特征分布,使其快速收敛;随着数据维度的不断降低,脉冲强度剪枝收敛速度相比原始网络获取的收益更大;当权值矩阵规模较小时,变化较大的权值将具有较强的数据特征表现力。

Aiming at the problems of DBN’s high degree of parameter redundancy,large amount of calculation,and long training time,this paper proposes a pulse intensity pruning algorithm(PI-DBN).This method performs network pruning operations on Deep Belief Network(DBN)in the unsupervised pre-training stage,optimizes the network structure,removes invalid network nodes,and deletes weights that are not of significance with the change of pulse strength to improve DBN convergence and calculation speed,while retaining the information of inherited data characteristics.The results show that the weight information after pulse intensity pruning can better characterize the feature distribution of the data in the low-dimensional data set,and make it converge quickly.With the continuous decrease of the data dimension,the convergence speed of pulse intensity pruning is greater than that of the original network.When the scale of the weight matrix is small,the weights with larger changes will have stronger data characteristics.