Iron Ore Information Extraction Based on CNN-LSTM Composite Deep Learning Model

In the mining, processing, and use of minerals, iron ore information identification is crucial. Traditional determination techniques are always accompanied by problems including lengthy experiment cycles, poor accuracy, exorbitant expenses, and significant workloads. In contrast, with the use of data-driven and sophisticated algorithms, modern hyperspectral technology can quickly deliver high-precision iron ore information, increasing efficiency. Magnetite collected from an iron ore mine in the Tangshan area is used as a pilot study, and its spectral data are used as the data source. The raw spectra are preprocessed Savitzky-Golay smoothing, jump point correction, and envelope removal. The bands are subsequently screened by correlation analysis, successive projections algorithm (SPA), and competitive adaptive reweighted sampling (CARS), down to 50 dimensions using principal component analysis (PCA). A convolutional neural network (CNN)-long short-term memory (LSTM) composite model is suggested to concurrently forecast the particle size and water content of magnetite based on its spectral characteristics. According to the model’s results, the particle size classification accuracy is 91.67%, the F1 score is 0.92, the coefficient of determination (R2) for the water content regression is 0.89023, the mean squared error (MSE) is 0.00082, the root mean square error (RMSE) is 0.02872, and the mean absolute error (MAE) is 0.01558. The composite model performs best with superior predictive performance compared to CNN, LSTM, decision tree (DT), random forest (RF), and extreme gradient boosting (XGBoost) models. The findings will push the mining sector toward more intelligence and efficiency, especially in the areas of smart mining and quick mineral appraisal.