Soft Prediction of Coal-mine Gas Concentration Through the Mixture of Gaussian Processes under the Noisy Input Prediction Strategy

Gaussian process regression is a typical nonlinear regression method in machine learning. However, it is rather difficult for a single Gaussian process to fit non-stationary and multi-modal series data. In addition, test input data may be interfered by noise. In order to overcome these problems, this paper proposes a regression method of mixture of Gaussian processes with the noisy input prediction strategy (niMGP), which is further implemented to perform a kind of soft prediction on coal mine gas concentration data. In comparison with traditional regression methods, this method adopts a noise into the test input data so that the prediction results become more robust and accurate. It is firstly demonstrated by simulation experiments that on a synthetic noisy test data with a fixed signal-to-noise ratio, the regression results of mixture of Gaussian processes with the noisy input prediction strategy is better than those with traditional prediction strategy in terms of stability. It is further demonstrated on the actual gas concentration dataset from the fragments recorded by No. 333944 sensor in Songzao Coal Mine with appropriate data enhancements that the noisy input prediction strategy is more stable than traditional prediction strategy. In practice, the prediction sensitivity can also be adjusted by adjusting the distribution variance of noise added to the test input data to achieve the effect of hierarchical warning.