Fault Classification Based On Semi-Supervised Ladder Network And Application In Air-conditioning System

Thanks to the fast development of informatization and intelligence,more sensors are deployed into industrial processes,providing with more abundant process data.Therefore,data-driven fault diagnosis has drawn widespread attentions from academic and industrial fields.However,there exist some problems in actual industrial data,such as label missing,high dimension and low value density problems.Due to the above negative characteristics,it becomes harder to obtain high-accuracy fault classification models through traditional pattern recognition methods.To address this problem,the fault classification method based on semi-supervised ladder network is studied in this thesis.From the perspective of classification process optimization,two fault classification methods are proposed based on semi-supervised ladder network feature extraction and multiple classical supervised classification model fusion.From the perspective of skip-layer structure optimization,two enhanced semi-supervised ladder network methods are proposed.From the perspective of application,the main algorithms proposed are applied to conduct fault diagnosis in air-conditioning system.Main research and results of this thesis are listed as follows:Firstly,aiming at solving the problems like lacking large-scale labeled samples and less effectiveness of raw data features in real-world industrial process,from the perspective of classification process optimization,two fault classification models are proposed based on semi-supervised ladder network feature extraction and multi-model fusion.Semi-supervised ladder networks act as trainable semi-supervised feature extractors in these models,and the new features generated will be used to train basic classifiers,since the classical supervised classification models need more effective features and the classification mechanism in semi-supervised ladder network is relatively simple.Furthermore,in order to balance the influence of different classifiers and tolerate different data distributions,voting and stacking ensemble methods are applied to the proposed fault classification models respectively.The experiment results illustrate the effectiveness and stability of the proposed fault classification models.And the model based on stacking ensemble method is better and more stable,has better performance than original semi-supervised ladder network.Secondly,aiming at the problem that the structures of the models obtained by semi-supervised ladder network classification process optimization are complex and cause the separation of training and classification,two fault classification models named semi-supervised residual ladder network and semi-supervised dense ladder network are proposed.The main improvements lie in the network architecture and the loss function.Here the idea of skip-layer information transmission in semi-supervised ladder networks is extended by introducing the residual connections of ResNet and the dense connections of DenseNet.These skip-layer connection structures could enhance the information transfer and feature reuse.Meanwhile,the predicted output loss of the corrupted encoder is also added into the original loss function,in order to ensure the consistency between the training target and the testing target.In experiment results,the enhanced models proposed are proved to show more convincing performance than the original semi-supervised ladder network and other traditional classification methods.Among all methods,the semi-supervised dense ladder network maximizing information flow between layers in the ladder network exhibits the best performance.Finally,this thesis tests the practicability of the proposed algorithms upon on air-conditioning system.The air-conditioning system dataset is collected from the simulation of a standard multi-split variable water air-cooled air-conditioning system provided by UTC Carrier Corporation.Nine types of faults occurred in main control loops are simulated.The actual diagnosis results also prove the effectiveness of the classification algorithms proposed in this thesis.And the semi-supervised dense ladder network based on skip-layer structure optimization exhibits the best performance and has high application value.