A concise peephole model based transfer learning method for small sample temporal feature-based data-driven quality analysis

摘要

Insufficient samples and low analysis efficiency are two main problems for data-driven quality analysis. To avoid negative transfer from source data to target data and improve transfer efficiency for temporal feature extraction, in this paper, a novel transfer learning model and algorithm with feature mapping, feature learning and domain adaptation was proposed based on concise peephole model (TLCPM). A feature mapping model based on deep convolution network was firstly presented by establishing deep convolution network to automatically learn and minimize the feature distance of source data and target data. The feature learning method was then constructed by concise peephole model to extract feature of temporal sample, which has less training parameters and more concise network structure than traditional peephole model. And domain adaptation combining one-layer fully connected network with TLCPM is used to realize the transfer learning ability of CPM model and minimize the probability output distribution distance. Finally, a bolt tightening test bench was set up and a tiny bolt-tightening data set was acquired to validate the proposed method. The results showed the TLCPM can be properly applied to analyze small sample temporal feature-based data and achieved good comprehensive performance. It took 0.5 min to obtain high test accuracy (95.14%), which was better than the results of other state of the art algorithms. Moreover, additional validation experiments were carried on. The results revealed that TLCPM also had better prediction accuracy on learning different categories of small samples.