Single-cell mass spectrometry analysis enables metabolic profiling of individual cells，helps to reveal the heterogeneity among cells，which is of great significance in oncology research．Bladder cancer is the most common malignant tumor in the urinary system at present．Accurate identification on the types of bladder cancer cells has an important value in life science and clinical application in the selection of treatment plan，prognosis judgment and drug resistance evaluation of patients．In this paper，single-cell mass spectrometry combined with machine learning was used to identify bladder cancer cells．The metabolic profiles for different bladder cancer cell subtypes were investigated by single-cell mass spectrometry analysis system，and classification algorithms were studied. Based on the collected single cell metabolic data，,t,-distributed stochastic neighbor embedding（,t,-SNE） clustering algorithm was used for dimensionality reduction analysis on the data，and the difference between the single cell metabolic profile was visualized in the two-dimensional space．In order to accurately identify different types of bladder cancer cells，linear discriminant analysis，random forest，support vector machine and logistic regression were respectively used to establish machine learning classification models，and grid search method and 5-fold cross-validation were used to optimize the model parameters．Then，five repeats of 10-fold cross-validation were performed on all data sets，and the averaged statistical result was taken as the final result．Accuracy，sensitivity，specificity，receiver operating characteristic（ROC） analysis and other indicators were used to comprehensively evaluate the performance of the model．The results showed that the metabolites of a single bladder cancer cell，such as ADP，ATP，glutamic acid，pyroglutamic acid，glutathione，etc，were successfully detected by the single-cell mass spectrometry system．There were significant differences among different types of bladder cancer cells，as well as large differences among single cells of the same type，indicating the high heterogeneity of single cell in the tumor．In addition，the four machine learning models all had good typing ability for bladder cancer cells，with a comprehensive accuracy not less than 94.9%，a sensitivity not less than 88.6% and a specificity not less than 93.3%．Compared with other methods，the random forest algorithm has the highest classification accuracy，sensitivity and specificity，which are all up to 100%，and the area under the ROC curve（AUC） of the model is up to 1，indicating that this method has obvious advantages in classification performance. The method presented in this paper realized the detection of metabolites and differentiation of cell subtypes at single cell level of bladder cancer，paving the way for more single cell metabolomics research in future.