Study On Fault Monitoring And Visualization Of Fermentation Process Based On Convolution Neural Network

The fermentation process is closely related to our life.From the 12 th Five Year Plan period to the 13 th five year plan,it shows that the state attaches great importance to the fermentation and pharmaceutical industry.The fermentation industry has been transformed from the traditional high energy consumption mode to the high capacity mode,which has created commercial value and achieved the important position of the industry.However,its vigorous development has gradually highlighted the safety problems in the process.If the safety problem is not solved in time,it may lead to economic loss or harm,casualties,etc.Therefore,when there is a fault in the process,it is necessary to detect the abnormality in time,monitor the fermentation process and eliminate the fault immediately.With the continuous development of complex industrial system and the gradual change of process mode,the traditional data-driven method can not meet the needs of fault monitoring of industrial process in the new era.Inspired by the deep structure network,aiming at the nonlinear and dynamic characteristics of fermentation process,this paper studies the fault monitoring method with convolution neural network as the core to improve the accuracy of the monitoring model and complete the fault online monitoring.The main research work is as follows:(1)Implementation of a multivariable time-frequency analysis based on Hilbert Huang methodWhen the fermentation process breaks down,the amplitude,phase and frequency of the corresponding variable signal will change with time.The time-frequency analysis of the variable can characterize its signal characteristics at different times and frequencies.According to the nonlinear and time-frequency characteristics of fermentation process,using EMD to express the nonlinear variable signal linearly,we can accurately capture the local characteristics of the signal,and then merge the decomposed variables for Hilbert transform.Thus,the time-frequency map of time-frequency features is obtained,which makes the variable features better map to time-frequency domain,and realizes multi variable time-frequency analysis.(2)A feature extraction method based on CNN is studied to monitor process faults onlineIn this paper,a deep convolution neural network method is studied to extract the adaptive feature of the time-frequency image after the time-frequency analysis,input it to the softmax layer for classification,and complete the online fault monitoring.In order to improve the training speed and accuracy of convolution neural network,a batch normalization(BN)method is introduced into convolution neural network.In view of the fact that the whole connection layer is prone to over fitting due to its needto optimize and update many parameters,the global mean pooling is used to take the mean value of the characteristics of the last layer after convolution pooling,so that the anti over fitting effect is better and the network generalization ability is improved.(3)Based on CNN Kiviat graph,a method of fault visual monitoring in fermentation process is proposedIt is difficult to express the process of fault classification monitoring by deep neural network feature extraction.Therefore,the visual fault monitoring method which combines visualization and fault monitoring is studied to reveal the process of fault monitoring vividly.Time display Kiviat map can display the depth features of data extracted by CNN in three-dimensional map,and realize the visual monitoring of fermentation process fault.(4)Experimental study on the actual data of E.coli fermentationThe convolution neural network feature extraction method and time display Kiviat visualization method proposed in this paper are applied to the actual production data of E.The results show that the method proposed in this paper can effectively extract the deep features of variable data,significantly improve the accuracy of fault monitoring,ensure the monitoring speed and superior monitoring ability,and guide the operators to find abnormalities in time and effectively prevent accidents.