Abstract
As the "Mining of Things" (MoT) integrates thousands of sensors into underground environments, reducing the data load on cable channels has become a critical challenge. Gas monitoring data, characterized by long periods of stability punctuated by sudden, safety-critical spikes, is particularly demanding. This paper proposes a Multi-Scale 1-bit Compressive Sensing (MS 1-bitCS) algorithm that leverages the statistical "regular patterns" of gas data. By employing non-uniform quantization based on safety priority, the algorithm achieves high compression ratios on redundant data while preserving sensitive information, overcoming the overload quantization distortion typical of traditional 1-bit systems.
The Challenge of the Mining of Things
Modern coal mines are increasingly reliant on dense sensor networks to monitor safety-critical parameters such as gas concentration, wind speed, and equipment health. This infrastructure, often referred to as the Mining of Things, generates massive volumes of time-continuous data that can overwhelm underground communication channels.
While Compressed Sensing (CS) offers a way to sample signals below the Nyquist rate, standard 1-bit CS—which reduces measurements to just their sign—often fails in industrial safety contexts. It suffers from "overload quantization distortion," where the fixed resolution treats all data points equally, potentially blurring the sharp, sudden changes (mutations) in gas levels that signify immediate danger.
The Multi-Scale 1-bit CS Approach
The proposed MS 1-bitCS algorithm introduces "prior attention" into the compression process. Instead of treating every bit of data with equal importance, it divides gas concentrations into five scales (Safe, Warning, and Alarm) based on concentration percentages.
Key Methodological Innovations:
- Attention-Based Division: Each scale is assigned a "Rule of Attention." Safe levels (<10% of the threshold) receive low attention, while alarm levels receive 100% attention.
- Non-Uniform Quantization: The algorithm uses a variable quantization step ($\delta$) that is inversely proportional to the square of the attention level. This effectively "zooms in" on critical data, providing finer resolution for dangerous spikes and coarser resolution for safe, redundant data.
- Adaptive Reconstruction: The researchers utilized a modified fixed-point iteration algorithm alongside Orthogonal Matching Pursuit (OMP) and Compressive Sampling OMP (CoSaMP) to reconstruct the signals from the compressed measurements.
Field Application and Datasets
The algorithm was validated using real-world data from the Yangzhuang coalmine (Huaibei Mining Group). The dataset, spanning April 2 to April 12, 2013, captures the typical profile of gas monitoring: long stretches of low-level activity interrupted by sudden mutations or equipment calibrations.
Key Results and Insights
- High Fidelity for Critical Data: The MS 1-bitCS algorithm achieved a Peak Signal-to-Noise Ratio (PSNR) of 31.8dB for mutation signals, successfully tracing spikes that traditional methods distorted.
- Efficiency and Accuracy: For general curve reconstruction, the system reached a PSNR of up to 44.6dB, proving it can accurately recover the original signal even under significant compression.
- Improved Reliability: By optimizing the measurement formula, the probability of successful signal reconstruction increased from 2% in standard 1-bit CS to 83.3% in experimental trials.
Conclusion
The Multi-Scale 1-bit CS algorithm provides a robust framework for managing the data explosion in MoT environments. By prioritizing "interesting" data—the 5% of signals that indicate danger—over the 95% of redundant safe data, the system ensures safety-critical information is preserved without overloading the network. Future work will focus on optimizing observation matrices to handle even more complex, non-sparse signal fluctuations.
Summary updated