The algorithm preselects valid candidate support vectors, reduces the size of the training sample set, and improves the training speed. Secondly, a new membership function is established to enhance the model's ability to perform optimal classification when constructing FSVM. The support vector set is obtained by training FSVM with pre- selected training samples. Finally, particle swarm optimization is used to select the optimal support vector subset. The average classification error was used as a fitness function. When the final particle is output, the membership degree of the sample is compared with the set threshold, and the sample with greater membership degree is selected as the new support vector. The experiment proves the learning performance of support vector machine. According to the historical data statistics and analysis results, SVM method is used to predict the new alarm data, which can be quickly and accurately judged. It solves the problem that traditional data mining methods cannot deal with dynamic changes, and uses support vector machine for classification prediction, which has obvious improvement in training time and prediction accuracy, and has good generalization ability, and can mine valuable information from massive data.