Research On Diagnosis Method For Working Conditions Of Aluminum Reduction Cells Based On Signal Fluctuation Of Anode Current

Stability is one of the spotlights in aluminum electrolysis process research. Diagnosing the working conditions of an aluminum reduction cell from its online parameters in order to take proper measures is an important strategy to enhance its stability. Anode current is one of the important parameters that can be collected online, whose signals contain a lot of information concerning the performance of an aluminum reduction cell. Diagnosis method is established by extracting characteristics from anode current signals.As the object of 160kA pre-baked anode cells, anode current signals were collected under either normal cells or typical problem cells. These signals were analyzed by the method of "spectrum-wavelet packet-neural network" with results are as follows:(1) Spectral characteristics of anode current signals in different cell states were investigated by spectral analysis method. Power spectrum estimation based on the Yule-Walker method was put forward to extract the spectral characteristics of anode current signals in different cell states, by which normal, cold and cathode fail cell would be distinguished by recognizing their frequency ranges of main peak as 0.003-0.018Hz,0.023-0.027Hz, and 0.027-0.031Hz respectively. The results of power spectrum estimation analysis have been compared with the results of HHT (Hilbert-Huang Transform) marginal spectrum analysis, which shows it is reliable for power spectrum estimation to analyze the anode current signals.(2) According to the results of power spectrum analysis, the concept of wavelet packets energy characteristics for anode current signal was introduced. Wavelet packets decomposition was performed at level 4 and 5 for anode current signals de-noised with wavelet, and related decomposition energy characteristics vectors were extracted. The results of wavelet packets decomposition was optimized, the appropriate wavelet packets reconstruction was selected, and the energy characteristics vectors of anode current signals were extracted in different cell states. Energy characteristics vectors that were normalized have obvious difference, which make a better distinction between the different cell states, provide the evidence for the diagnosis of cell states.(3) According to the wavelet packet energy characteristics vectors extracted from anode current signals, diagnosis model based on BP neural network was established and verified. Simulation results show that the model of network identification is simple in construction, high accuracy in recognition, and convenient to realize on-line monitoring and real-time identification.