Traditional aluminum electrolysis fault diagnosis methods have problems such as low accuracy, small forecast advance, and high CPU usage, which make their popularity low in enterprises. Aiming at the above problems, a fault diagnosis method with switchable two-level classifiers is designed. The input data are first judged by the first-level algorithm. If it is determined that there is no fault, the result will be output directly. If it is determined that there is a fault in the electrolytic cell, the data will be transferred to the second-level network for specific fault diagnosis. The first level is based on the Random Forest algorithm with simple structure and good two-class classification effect and is optimized by the improved cuckoo algorithm. The second level is based on an improved DBN-DNN (Deep Belief Neural Network–Deep Neural Network) algorithm, and the training method is given. Experimental results show that this method can switch between different algorithms according to different situations, save computing resources, realize that a computer can monitor multiple electrolytic cells, and reduce investment costs. In addition, the accuracy and forecast advance have been significantly improved, which has promoted the popularization of fault diagnosis systems in aluminum electrolysis enterprises.

传统的铝电解故障诊断方法存在准确率低、预测提前量小、CPU占用率高等问题，在企业中普及度不高。针对上述问题，设计了一种具有可切换两级分类器的故障诊断方法。输入数据首先由一级算法判断。如果确定没有故障，则直接输出结果。如果确定电解槽有故障，则将数据传送到二级网络进行具体故障诊断。第一层基于结构简单、二分类效果好的随机森林算法，并通过改进的布谷鸟算法进行优化。第二层基于改进的DBN-DNN（Deep Belief Neural Network-Deep Neural Network）算法，给出了训练方法。实验结果表明，该方法可以根据不同情况切换不同算法，节省计算资源，实现一台计算机可以监控多个电解槽，降低投资成本。此外，准确率和预测提前性得到显着提高，促进了故障诊断系统在铝电解企业的普及。