Study On Localization Algorithm For Coal Mine Based On Gradient Boosting Regression Tree

Disasters occur frequently in the process of underground coal mining.The accuracy and stability of the underground positioning algorithm are related to the efficiency of post-disaster rescue.A well-established positioning system can greatly reduce property and life losses.This paper studied the existing positioning technology and its advantages and disadvantages,and focused on the location fingerprint based positioning method.The paper proposes a underground localization algorithm based on gradient boosting regression tree to improve the accuracy of the positioning system.An adaptive weighted Gaussian filter is proposed to process the position fingerprint set to improve the stability and reliability of the data.The establishment of the positioning system is divided into three stages:location fingerprint acquisition,model training and online positioning.In the off-line acquisition phase,this paper fully considers the complex multipath effect of underground tunnels.It is difficult to accurately express the propagation loss of signals by ordinary empirical propagation models.Therefore,the ray tracing algorithm is used to simulate the results of multiple transmission superposition of underground signals,which is closer to the real environment.In addition,due to the severe time-varying of the underground signal,an adaptive Gaussian weighting filter is proposed to filter out the singular signals in the data set.The signal with a close distance from the signal transmitting point gives a larger weight,and the signal with a far distance is given a smaller signal weight.Thereby a more stable and reliable training data set is obtained.In the model training phase,this paper proposes a underground localization algorithm based on Gradient Boosting Regression Tree(GBRT),which can be regarded as a combination of forward distribution algorithm and addition model.A plurality of continuous regression trees are established for the processed fingerprint data set,and the optimal segmentation point of each regression tree is determined based on the heuristic search and the mean square error minimization criterion,and then the fitted coordinates and real coordinates of the regression tree are calculated.Each subsequent regression tree fits the residual value of the previous regression tree,and finally adds the fitted coordinates of all the regression trees to obtain the gradient-enhanced regression tree model.Firstly,the paper introduces the establishment process of the algorithm,then analyzes the influence of regression tree depth,sub-sampling ratio,maximum iteration number and learning rate on the positioning accuracy,and finally determines the model parameters suitable for the underground fingerprint dataset.In the online positioning stage,the signal strength vector whose position coordinate is unknown is input into the GBRT model,and the coordinate value corresponding to the signal strength vector can be output,thereby realizing the positioning.The final simulation of the paper verifies the validity and reliability of the GBRT algorithm.In the random positioning of 50 points,the average error is 0.54 meters,the average error on the X-axis is 0.28 meters,and the average error on the Y-axis is 0.29 meters.Then the paper compares the error cumulative distribution function and pedestrian trajectory positioning results of five algorithms,including K-Nearest Neighbor(KNN),Random Forest(RF),Support Vector Machine(SVM),Multi-layer Perceptron Regressor(MLPR)and GBRT.The positioning results of the GBRT algorithm are significantly better than the other four algorithms.Considering the characteristics of pedestrian walking,the paper finally performs 5-point average filtering on the positioning result of GBRT,and the filtered positioning trajectory is closer to the real trajectory.The simulation results show that the underground localization algorithm based on gradient lifting regression tree has higher precision and can meet the requirements of underground positioning.