| With the complex and complicated modern chemical process,the production scale is also expanding,which brings higher benefits to the enterprise,and the probability of process failure is increasing gradually.As the main energy consumption part of HVAC,once the failure can not be found in time,it will cause a lot of energy waste and even serious accidents.However,the abnormal condition of HVAC caused by the failure of chiller frequently occurs,so it is a hot research topic to diagnose the faults accurately and timely.With the continuous expansion of the application of information acquisition system,the scope and depth of chemical process state monitoring are also strengthened.The data features collected are increasing in a large amount,and they are complex,high-dimensional and nonlinear.In addition,for traditional pattern recognition fault diagnosis methods,a large number of labeled data is necessary for training effective models,but it is difficult to collect enough labeled data in industrial application scenarios.Therefore,it is urgent to establish a fault intelligent diagnosis algorithm for tag free sample self-learning,which has become one of the research hotspots.In view of the above problems,this paper mainly studies the following contents:(1)Traditional fault diagnosis of chemical process is based on time series data,and the features only learn in time domain,and do not capture the frequency-domain changes.In this paper,the original data of chiller is extracted by wavelet transform to time frequency image,so that it can not only represent the change of data in time domain,but also represent the change in frequency domain and obtain more fault information.(2)In view of the traditional image feature extraction method,feature extractor relies on artificial design,which requires a lot of experience knowledge,low diagnostic accuracy and poor generalization ability,this paper applies the deep convolutional neural network(CNN)with strong feature extraction capability to extract the fault time-frequency diagram of chillers.CNN is mainly data driven feature extraction.According to the study of a large number of samples,it can get the specific feature representation of deep data set.It is more efficient and accurate to express the data set,and the extracted abstract feature is more robust and generalization ability is better.(3)In view of the problem of incomplete samples,this paper combines CNN with twin structure based on similarity measurement,and proposes a wavelet based twin convolution semi supervised neural network(WT SCNN)and applies it to the fault diagnosis of chillers.Firstly,the fault data is transformed into wavelet transform to get the time-frequency map of the fault as the input of the network;Then,the twin structure is constructed by using the convolution layer and pool layer of two identical CNN through weight sharing;Finally,the paired time-frequency graph is input into the network,and CNN extracts the data.The network maps the extracted features to the lowdimensional space.By calculating the Euclidean distance between unknown samples and known samples in the target space,the unknown fault is a known fault or a new fault.The simulation results of chiller are used to verify the proposed algorithm.The experimental results show that the method can diagnose the tag free fault data,realize the self-learning of new faults and the self-growth of the fault library,which has certain intelligence.(4)According to the experiment,the proposed algorithm can train for more than 60000 seconds when the known fault reaches six classes.In view of the problem of long training time,the paper improves the network training time.A twin asymmetric convolutional semi supervised neural network based on wavelet transform is proposed,(WT-SCNN)and applied it to the fault diagnosis of chiller.This algorithm does not change the overall structure of WT-SCNN,but only changes the traditional CNN part to the non symmetric convolution core CNN,changes the convolution core with the original size of 2×2 to the convolution core with the size of 2×1,and keeps the other parts unchanged,so as to reduce the number of weight parameters and the total parameters of convolution layer,speed up the network feature extraction efficiency and shorten the network training time The simulation results of chiller are used to verify the proposed algorithm.The experimental results show that the algorithm can improve the average training speed of network by 20% without changing the accuracy of diagnosis,so that the network can be more real-time and can diagnose the fault in time. |
