K-means is a widely used and efficient unsupervised clustering algorithm. However, studies have shown that when dealing with high-dimensional or non-spherically distributed datasets, the K-means algorithm has significant limitations in determining the number of clusters and selecting initial centroids. To thoroughly explore and optimize the initial centroid selection mechanism and the problem of determining the number of clusters in the K-means algorithm, a K-means algorithm based on knowledge induction and useless center driven is proposed. This algorithm first introduces a detection mechanism for high-density knowledge points, constructs a candidate centroid set by identifying high-density knowledge points in the dataset; then infers the optimal number of clusters based on Gaussian mixture model theory; subsequently adopts a useless center screening strategy to optimally select the candidate centroids, and finally determines the optimal initial centroid set. Experiments on real datasets show that the proposed optimized algorithm generally outperforms other comparison algorithms in clustering performance. This algorithm effectively solves the clustering problem of non-spherical data distribution and exhibits relatively superior clustering performance in scenarios with complex data structures.