Aiming at the problem that it is difficult to identify the load state due to the inherent characteristics of ball mill vibration signal such as strong randomness, non-smoothness and non-linearity, a load state identification method for ball mill based on phase space reconstruction and PSO-K-means is proposed. Firstly, numerical simulation of two chaotic time series of Lorenz and Rossler used the improved before-and-after autocorrelation coefficient algorithm and derived an accurate and efficient method for calculating the delay time and embedding dimension. Then, feature extraction was performed for phase space attractors under three different load states, and the variation law of the associated dimensional feature quantity was analyzed. Finally, Classification and identification of ball mill load states by inputting the correlation dimensions as feature vectors into the PSO-K-means clustering model. The results show that the PSO-K-means clustering model has high accuracy in load state identification, with 94.2%, 96.3% and 94.8% accuracy under underload, normal load and overload, respectively. The above results confirm that the method can achieve effective identification of ball mill load states.