Cross-condition domain adaptation for coal-gangue recognition with transfer learning and Mel spectrograms

Abstract In order to enhance the performance of coal-gangue recognition (CGR) under longwall top coal caving (LTCC) mining conditions and solve the problem of data acquisition difficulties and limited data information, a cross-condition domain adaptation transfer learning (TL) method for CGR was proposed. Firstly, coal gangue signals collected from both underground and the experimental bench simulating actual operating conditions were preprocessed and converted into Mel spectrograms to enhance the time-frequency feature representation. Subsequently, a deep convolutional neural network (CNN) was designed to enable domain adaptation by aligning feature distributions across varying operating conditions while maintaining shared model parameters. The network integrates a joint loss function that combines Maximum Mean Discrepancy (MMD), Domain Adversarial Loss, and Covariance Loss, ensuring effective feature alignment and robust cross-domain generalization. Finally, the recognition performance for CGR was optimized by evaluating various model architectures and hyperparameter configurations. Additionally, the effectiveness of the Mel spectrogram in recognition performance was assessed by comparing different spectral representations as model inputs. Comparative ablation experiments are further conducted to validate the effectiveness of the proposed joint loss function. The experimental results demonstrate that the proposed method achieves an accurate CGR with a recognition rate of 91.59%. This method effectively solves the issue of insufficient information and difficulty in collecting underground gangue signal data, enhancing the performance of CGR in underground mining environments and demonstrating the feasibility of intelligent LTCC under coal mine intersection conditions.