It is well known that direct ionization mass spectrometry(MS) has the advantages of few pretreatments,rapidness,and convenience,which is widely used in the fields of rapid detection of various substances,such as explosives,drugs,food additives,etc.However,for the direct ionization mass spectrometry,there exist large fluctuations in the signal intensities which usually obey the lognormal distributions.As a result,how to identify these signals and classify the samples becomes a real challenge.The common analysis methods for mass spectrum data include the extracted ion current(EIC) and Gaussian mixture model(GMM).Both of these analysis methods make only use of the peak intensity information,but ignoring the peak shape features such as peak position,full width at half maximum(FWHM) and correlation information among multi peaks.In addition,the detection accuracies of these methods also depend on the preset thresholds,which is usually a tedious work to optimize.However,machine learning unlike EIC and GMM could take advantage of all the characteristics of mass spectrum signals to improve the detection performance without a preset threshold.In this work,acetylsalicylic acid(115 samples) was adopted as an analog for explosives in the mass spectrometry detection.In order to improve the detection accuracy especially for dielectric barrier discharge ionization(DBDI) source and mass spectrometer in detection of the samples with low concentrations,the data preprocessing and classification algorithms based on EIC,GMM and machine learning of the direct ionization mass spectrometry were studied.Moreover,2,4,6-trinitrotoluene and ammonium nitrate(110,90 samples,respectively) with a concentration of 1 ng/mL,and blank samples(366 samples) under the same detecting conditions were investigated and analyzed.Results showed that the signal intensities of either the blank samples or the explosive samples with the same concentration,all obey the lognormal distributions,spanning several orders of magnitude.As set forth,the detection results based on the EIC and GMM were found all sensitive to the preset thresholds.However,the random forest(RF) algorithm could be used to improve the classification accuracy,compared with EIC and GMM methods.And the F_score numbers for EIC,GMM and RF are 0.74,0.89 and 0.96,respectively.In addition,with the same analysis methods,MS/MS could significantly reduce the background interference and improve the detection accuracy,compared with the full scan mass spectrometry.In practice,it usually takes 3-6 seconds to acquire the mass spectrum data for each sample.In this work,the analysis time for each single sample data is much less than 0.1 second,which means the algorithm could be adopted to a real time detection.In conclusion,direct ionization mass spectrometry combined with machine learning could meet the needs for rapid,real time and accurate detection of explosives,which exhibits good application prospects.