基于TUCKER-DBN的机械故障识别方法研究

针对传统深度信念网络（DBN）存在的分类精度不高、训练速度较慢、仅适用于一维信号等不足，将DBN结合TUCKER分解，提出一种新的故障识别方法。该方法首先利用TUCKER分解大幅度压缩数据，提取其核心张量作为故障特征，然后将核心张量输入到DBN分类器中进行训练和识别。将该方法与传统的DBN故障识别方法进行对比研究，在采集的120个样本中，选择30个样本进行故障识别测试实验。结果表明:使用TUCKER-DBN识别方法的识别率为93%，较传统的DBN故障识别方法的识别率更高；并且使用TUCKER-DBN识别方法的训练时间比传统DBN故障识别方法所用的时间更短。

Research on Mechanical Fault Recognition Method Based on TUCKER-DBN

Aiming at the shortcomings of traditional deep belief network (DBN), such as low classification accuracy, slow training speed, and only being suitable for one-dimensional signals, a new fault identification method was proposed by combining DBN with TUCKER decomposition. Firstly, the method uses TUCKER to decompose the compressed data, extracts its core tensor as fault features, and then inputs the core tensor into the DBN classifier for training and recognition. This method was compared with the traditional DBN fault identification method. Among the 120 samples collected, 30 ones were selected for fault identification test. The results show that the recognition rate of TUCKER-DBN is 93%, which is higher than that of the traditional DBN method. Moreover, the training time with TUCKER-DBN recognition method is shorter than that of traditional DBN fault recognition method.