关键词: 电解, 泄漏事故, 算法, 神经网络, 故障预测

Abstract:

The slot leakage accident will cause enormous economic losses and casualties during the production process of aluminum reduction cell. Therefore, the accurate early warning of aluminum reduction cell leakage accident is significant for enterprises. As for the issue that the operation data of aluminum reduction cell is multidimensional, nonlinear, and the sample size of failure is smaller than that of normal operation, a data-driven model for early warning model of slot leakage accident is proposed. The support vector machine (SVM) algorithm with good robustness and high efficiency was adopted to avoid over-fitting due to small fault sample size. Then, cross-validation and sparrow search optimization algorithms (SSA) were used to optimize the parameters of SVM. The F₁ score and AUC value of the optimized model were increased by 0.166 and 0.163 respectively. To better mine the characteristic information in the running data of aluminum electrolysis, the kernel principal component analysis (KPCA) method was used to reduce the data dimension. Eight principal components were selected when the cumulative variance contribution was 80%. The running time of the model was thereby decreased by about one minute. The F₁ score and AUC value of optimized model were also increased by 0.046 and 0.038, respectively. Thirteen characteristic parameters were selected from many production parameters of aluminum electrolysis as input of model. Then, the missing data in the collected data were interpolated by K nearest neighbor (KNN) algorithm. When setting the classification labels, three characteristic parameters of temperature, Fe content, and Si content of the electrolysis process with obvious changes were selected as auxiliary classification conditions and combined with the actual situation. Thus, the fault feature range and the size of fault samples were expanded, indirectly solving the problem of small fault sample size to a certain extent. The experimental results showed that the F₁ score and AUC value of the early warning of aluminum reduction cell leakage accident model reached 0.995 and 0.998, respectively. The feasibility and effectiveness of machine learning application in the field of early warning of aluminum reduction cell leakage accident are verified. It demonstrates that the KPCA-SSA-SVM model performed better than other classification models, which is more favorable for early warning of aluminum reduction cell leakage accident with small fault sample data. The results have important practical significance in prolonging the remaining life of aluminum reduction cell, improving the safety in the production process, and promoting the intelligent production of aluminum electrolytic enterprises.

Key words: electrolysis, leakage accident, algorithm, neural networks, fault prediction