Prediction model of flotation molybdenum-bismuth product quality based on RF-BiLSTM

Molybdenum and bismuth are important strategic metal resources, widely used in electronics, metallurgy and other industries. Flotation is a critical step in recovering molybdenum-bismuth concentrate in concentrators, and the quality of the flotation product directly impacts the plant’s economic efficiency. However, at present, most concentrators in China rely on manual shift sampling and offline assaying methods for grade detection. A significant issue with this traditional approach is the delay in obtaining grade information. To address this issue, this paper proposes a dynamic prediction method for flotation product quality using deep learning. Industrial data from concentrators is susceptible to various factors such as environment, equipment failure and other factors, resulting in missing or abnormal data. It is difficult for traditional statistical methods to accurately capture trends and potential links among data. Based on this, this paper proposes an improved random forest interpolation method for missing value repair of mineral processing plant industrial data. This method is effective in filling in missing data and reducing information distortion compared to traditional filling methods such as median. The experimental results show that using this method for data filling enhances the accuracy of downstream prediction task. Additionally, the MAPE value is reduced compared to traditional statistical method, effectively improving data quality. Based on the optimized dataset, BiLSTM prediction model incorporating bidirectional temporal feature extraction is constructed in this paper. Compared with the traditional unidirectional LSTM model, BiLSTM model is able to utilize both forward and backward temporal information to improve the prediction accuracy. Experiments demonstrate that the model shows significant advantages in predicting molybdenum and bismuth grades. Specifically, the molybdenum grade prediction achieves a MAPE value of 0.87% and an R² of 0.89, reducing prediction error by 44.23% compared to the LSTM model. Additionally, the prediction accuracy is sustained even during fluctuating working conditions. It is confirmed that the model has good generalization ability and engineering applicability, and can realize the prediction of molybdenum-bismuth flotation product quality.