Research On Fault Diagnosis Method For Aluminum Reduction Cell Based On LMD And ELM

Reduction cell, the typical equipment in the aluminum electrolytic production, its running state affects the economic and technical index, daily production and the life of itself. But, reduction cell is disturbed by magnetic field, electric field and great heat at the same time; this makes many effects including the difficulty of identify fault features, the shortage of production parameters with the property of continuous measurement and the difficulty of reduction cell’s condition monitoring and fault diagnosis. Therefore, as the object of 306 k A pre-baked anode cell, cell voltage signals were collected under normal cells, metal pad wave cells, cathode failure cells, lower polar distance cells and scabbing under cell bottom cells. Furthermore, the de-noising for the signal, the extraction of fault feature and the classification of faults were carried out to study.Firstly, some common failure modes of reduction cell and the relationship between failure mode and cell voltage signal were introduced. The direct method was used to estimate the cell voltage signals’ power spectrum under five conditions. The experimental results showed that the main frequency domain was 0-0.1Hz and the frequency features of power spectrum were similar, the signals need to be analyzed further. Then, all the cell voltage signals were de-noised by wavelet threshold algorithm.Local mean decomposition algorithm is often used in fault feature extraction for mechanical vibration signal because it can keep more frequency and amplitude characteristics than other methods. Process signal is different from mechanical vibration signal because of the stationary signal amplitude and the low sample frequency. Therefore, LMD algorithm was improved in this paper by adjusting the signal value range, doing the nonlinear transformation, magnifying signal local characteristics and resolving signal. Then, relative energy and total energy of PF weight were selected to be the fault features. The experimental results showed that the features could indicate the states of reduction cell accurately by the improved algorithm.Finally, extreme learning machine algorithm was adopted to identify the fault types. For the disadvantage of hidden layer node parameters that were selected random in ELM, particle swarm optimization algorithm was employed to search for the optimal hidden layer node parameters in order to improve the classification precision. Then, to overcome the disadvantage of easily getting into the local extreme of PSO algorithm, an improved of the inertia weight was proposed. Nonlinearly decreasing weight reduction was adopted and classification accuracy was selected to be the prerequisite of optimization procedure in order to achieve less number of iterations and higher network training speed. The experimental results showed that both high training speed and classification accuracy.