Fast detection of short circuits in copper electrolytic refining with PCA and a branching perceptron

Abstract

Short circuit faults greatly affect the current efficiency and product quality of copper electrolytic refining. Irregular cloth covers and uneven heating make it difficult to detect short-circuit faults of copper electrolytic refining with infrared images, so complex object detection models believed to do well in feature digs were previously used. Yet, such methods have high computation costs, which limits the detection efficiency and hinders the function expansion to portable devices. We find that with proper feature extraction, the model can be succinct, and the checkbox output of object detection models is not the most applicable form. Therefore, this paper proposes a fast multilabel classification method to detect the short circuit faults of copper electrolytic refining with principal components analysis (PCA) and a branching perceptron. In the method, PCA reduces data dimensions in an unsupervised and reversible way according to the maximum projection variance principle since faults appear as discrepancy signals in images. Then, a branching perceptron is presented for fault identification. Each branch corresponds with a pair of anode and cathode electrodes and the output of the branch predicts the state of the pair. Reversible low-dimensional features obtained with PCA can reduce the pressure on data transfer and storage, and support a more succinct detection model to fasten the training and detection. The binary sequence form of the designed output is more convenient for on-site fault removal and other purposes. The proposed PCA-Perceptron method is verified on real-world data of electrolytic refining of recycled copper.