Condition recognition based on multi-source heterogeneous data and residual temporal network in coal flotation process

Condition recognition in coal flotation is a critical component in achieving flotation process control and optimization. To develop reliable models for condition recognition, two key challenges must be addressed: first, the effective extraction and integration of froth visual feature information from multiple flotation cells; and second, the fusion of froth visual features with process parameters to create a comprehensive representation of flotation conditions. To tackle these challenges, a time series classification model based on multi-source heterogeneous data was proposed for coal flotation condition recognition. First, a bubble size feature representation method combining the tsfresh and maximum relevance minimum redundancy algorithms was proposed. This method accurately describes bubble size distribution features through multi-scale feature extraction and mutual information-based feature selection. Next, a neural network feature evaluation model based on L1 regularization was presented, which quantifies the importance of visual features through the sparse expression of network weights, facilitates feature selection, and applies the resulting feature weights to the subsequent weighting process in classification models. Finally, a residual temporal network that integrates residual feature extraction and temporal modeling was proposed. This model employs one-dimensional convolution and cascaded residual structures for deep feature extraction, while temporal dependencies of condition changes are captured by combining long short-term memory and attention mechanisms. Validation on a flotation froth dataset collected from field sites in Anhui demonstrates that the proposed model achieves a condition classification accuracy of 85.08%, significantly improving the performance of flotation condition recognition compared to existing models.