Research On Cathode Condition Diagnosis Of Aluminum Electrolysis Cell Based On Wavelet Analysis And Neural Network

In industrial production,aluminum is prepared by electrolytic method.At the same time of aluminum preparation in the tank,the voltage at the cathode soft bus is measured and analyzed to reflect the state of the cathode in the tank in industrial production.Timely and effective state diagnosis has a very important impact on the normal production and economic benefits of aluminum plants.It is necessary to process the voltage signal collected from the cathode soft bus in the electrolytic cell of the aluminum plant,take the processed data as the input of the wavelet packet function,extract the characteristic signal,determine the wavelet basis function by calculating the minimum energy entropy,determine the decomposition level of the wavelet packet according to the shape of the generated optimal wavelet packet tree,compare wavelet denoising with wavelet packet denoising,and select wavelet packet denoising to process the original data by comparing the signal-to-noise ratio and mean square error.The denoised data is matched with the fault condition in the corresponding time period by wavelet packet feature extraction,and the decomposed feature signal and the corresponding fault condition are taken as the input and output of the subsequent fault diagnosis neural network.The fault diagnosis of aluminum electrolysis cell is carried out by combining a neural network method,a BP neural network is selected as a diagnosis tool and a three-layer neural network model is established,the eigenvalues extracted from wavelet packets are constructed into a data set,and the data set is divided according to a ratio of 3:1 to obtain two parts of data of a training set and a test set,and after diagnosis,the traditional BP method is found to have poor effect on the training set,slow convergence speed and prone to repeated oscillation in the gradient descending process,In order to avoid the defects of the steepest descent method in the learning process,the improved BP algorithm is used.In the training algorithm,the momentum factor is added,the learning rate is changed adaptively in the learning process,and Newton method is combined to optimize.The experimental comparison is made respectively,and the improvement effects of several methods are compared.Finally,the L-M optimization BP algorithm is used to improve the diagnostic accuracy of the test set state.The same data from slot 2 is taken for testing.The network has certain generalization ability.The software system is designed on the MATLAB software platform,including login interface,shutdown interface,data viewing interface,curve display interface and wavelet packet decomposition interface.The basic functions of fault diagnosis are realized.