Hybrid Fault Diagnosis Method Of Chemical Process Based On LSTM And Dynamic Model

In order to guarantee the safe operation of the chemical plant,early detection and diagnosis needs to be carried out.The symptoms of the equipment and the trend of deterioration should be discovered as soon as possible,and necessary measures should be taken to eliminate the hidden dangers of the accident.The single fault diagnosis method is difficult to accurately locate the cause of faults in the complicated chemical process,and the diagnosis accuracy and speed are not easy to meet the industrial requirements.Therefore,it is very necessary to combine multiple methods to realize fault diagnosis of the chemical device.Based on the above issues,this paper uses the Tennessee Eastman(TE)process and an industrial catalytic cracking simulation system as examples to combine the Long Short Term Memory Network(LSTM)with the dynamic model to realize fault detection,fault identification and prediction of fault parameters.Chemical plants are complicated nonlinear dynamic systems with many variables,so it is difficult to choose feature variables.In this paper,the dynamic simulation system of chemical process is established based on the equations of material balance and energy balance.When the current state of the device deviates from the normal state,the fault diagnosis step is activated.This work uses the nonlinear dimensionality reduction algorithm t-SNE to reduce the dimensionality of high-dimensional data sets composed of multiple variables.Then,the dimensionality reduced data is divided into training set,validation set and test set,and input into LSTM to realize fault identification after training and testing.Finally,the LSQ is used to determine the critical fault parameters in the abnormal state.The PLS is used to fit the relationship between the system output and the fault parameters to predict the fault parameters.Because PLS does not require iterative calculations,the employment of PLS instead of LSQ can speed up calculations and save computing time.The application results show that the hybrid method can accurately distinguish the fault and predict the fault parameters,realizing the location of the fault causes and the prediction of the development trend of the parameters.In the five repeated classification tests of 9 kinds of faults in TE example,the accuracy rate of fault recognition is up to 97.6% and the average accuracy rate is 94%,indicating that the LSTM network can accurately identify the fault status of chemical process from the original data.Fault 8 is chosen as example.When estimating the fault parameters with an optimized inversion method,the number of function estimations is reduced by 92.31% and the running speed is increased by about 13 times compared with the basic case,proving the effectiveness of this method.