K-means algorithm is a numerical, unsupervised, and uncertain iterative method. This method is simple and fast, and has been proved to be a very effective clustering method in practical applications. This article focuses mainly on the two shortcomings of the K-means algorithm, that is, the algorithm is sensitive to the outliers and noise points of the data sample, and the objective function has a local optimal solution. The improvement of this algorithm is proposed by using the stochastic gradient descent method to replace the traditional K-means algorithm to obtain the average value of the distance within the cluster to update the cluster center. After continuously reducing the learning rate, the algorithm is implemented in the direction of gradient descent, pulling the cluster center nearer to the data sample. Through this improvement, the influence of isolated points on the cluster center is effectively avoided. At the same time, for the situation where the objective function has a