Abstract
As the primary equipment in coal mining, shearer anomaly detection is critical for production efficiency and safety. Challenges include the scarcity of labeled data and the machine's complex structure. This paper proposes an unsupervised method using a Multi-scale Correlation Matrix (MSCM) and a CNN-ConvLSTM Autoencoder (C-CLA) network. The MSCM captures interrelationships between features from multiple perspectives, while the C-CLA extracts spatial and temporal features to detect status changes. Validated on real-world datasets, the method outperforms existing techniques by approximately 3%.
Introduction
In underground coal mining, the shearer is the central production tool. Unexpected failures lead to significant downtime and safety risks. Traditional deep learning approaches for anomaly detection often require extensive labeled datasets of failure modes, which are rare and expensive to obtain in industrial settings. Furthermore, local sensor monitoring often fails to capture the systemic health of the machine. This research addresses these gaps by focusing on unsupervised learning and the holistic interdependencies of operational features. By training on "healthy" data, the system learns to recognize the normal operational "signature" of the machine, identifying anomalies as deviations from this baseline.
Methodology
Multi-scale Correlation Matrix (MSCM) Construction
The first stage of the proposed framework involves transforming raw multi-dimensional sensor data into a series of correlation matrices. This approach shifts the focus from individual sensor values to the relationships between them. The researchers utilized 20 dimensions of data covering critical components such as the cutting part, traction part, rocker arm, and main pump. Three statistical methods are employed to construct the MSCM:
- Pearson Correlation: Measures linear relationships between features.
- Spearman Correlation: Captures non-linear, monotonic relationships, providing robustness against outliers and sensor noise.
- Temporal Cross-Correlation: Analyzes how features relate to each other across different time lags, capturing the dynamic nature of the equipment's operation.
CNN-ConvLSTM Autoencoder (C-CLA)
The core of the detection system is the C-CLA network. It utilizes a Convolutional Neural Network (CNN) to extract spatial features from the MSCM heatmaps and a Convolutional Long Short-Term Memory (ConvLSTM) network to capture temporal dynamics. As an autoencoder, the model is trained exclusively on "healthy" data to learn a compressed representation (latent space) of normal operations. During inference, the model attempts to reconstruct the input matrix. Anomalies are identified by calculating the reconstruction residual matrix; a high reconstruction error indicates a breakdown in the expected relationships between machine components.
Experimental Setup and Results
The model was evaluated using 54 days of operational data from the Yujialiang Coal Mine, totaling roughly 2 million data points per dimension. To verify generalization, the researchers also tested the architecture on a separate dataset from a different mining face in Inner Mongolia.
Comparative Performance
The C-CLA method was compared against several state-of-the-art anomaly detection models, including TranAD, Donut, and GAE-M. The proposed method achieved a detection efficiency approximately 3% higher than these alternatives. Ablation studies further confirmed that the multi-scale approach (combining Pearson, Spearman, and Temporal correlations) significantly outperformed models using only a single correlation type.
Detection Thresholds
In testing, the reconstruction error for healthy data segments remained consistently low, around 10%. In contrast, for segments containing known anomalies, the error spiked to over 30%. This clear differentiation allows for the establishment of robust automated alert thresholds, reducing false positives while ensuring critical faults are captured.
Conclusion
This paper demonstrates that unsupervised learning, combined with multi-scale feature correlation analysis, provides a robust solution for shearer anomaly detection in the harsh environments of coal mines. By focusing on the breakdown of relationships between operational parameters rather than simple thresholding of individual sensors, the C-CLA framework offers superior detection capabilities without the need for manual data labeling. Future research will explore transfer learning techniques to adapt the model to different mining environments and equipment models with minimal retraining.
Summary updated