Research On Gas Sensor Fault And Abnormal Signal Recognition Based On Improved Random Forest

The gas sensor is an important component of the coal mine gas monitoring system,aiming at the problem that the coal mine gas sensor is prone to malfunction and abnormal signals in the actual production environment,which leads to the failure to effectively reflect the actual gas concentration and threaten production safety.A method for identifying fault and abnormal signals of gas sensors based on improved random forests is proposed to identify the signal types of gas sensors,thereby completing the judgment of the operating status of gas sensors.The type of signal can be used to determine what fault has occurred or what abnormal signal has occurred in the gas sensor,so as to facilitate timely maintenance or replacement of the sensor.Firstly,the working principle of general mine gas sensors and the types of fault and abnormal signals of gas sensors are analyzed.The causes of six abnormal sensor signals,such as drift,constant deviation,cycle,impact,jamming,and outburst,are analyzed.The gas signal is simulated using a combination of experiment and simulation.Then,CEEMD(Complementary Ensemble Empirical Eode Decomposition)is used to analyze the signal of the gas sensor,and it is verified that CEEMD can solve the problem of EMD mode aliasing while also solving the problem of noise residue in EEMD.Secondly,the selection mechanism of IMF(intrinsic mode function)components is analyzed.Several IMF components generated after decomposition are partially selected through the correlation coefficient principle and variance contribution rate principle to achieve the purpose of dimensionality reduction.Through experiments,it is concluded that the correlation coefficient principle is superior to the variance contribution rate principle.After comparing the experimental results of energy entropy,singular spectral entropy,and power spectral entropy,the IMF’s energy entropy was finally selected as the feature.Then,a genetic algorithm optimized random forest(GA-RF)is used to identify gas sensor faults and abnormal signals.Genetic algorithm is used to optimize the maximum characteristics of nodes and the number of decision trees.After obtaining appropriate parameters,a high-performance random forest is constructed,improving the accuracy of fault and abnormal signal recognition.Using this method to compare with the least squares support vector machine and BP neural network after particle swarm optimization,experiments show that this algorithm has advantages.Finally,a filtering method based on CEEMD decomposition is proposed for abnormal signals such as "large number" interference signals.Experimental results show that this method can effectively filter out "large number" interference signals and prevent false alarms in monitoring systems.