Abstract: A ball mill is important grinding equipment in a concentrator, and the accurate detection of the load status ensures that the ball mill runs in the best state, which helps optimize the grinding process, ensure the stable operation of the ball mill equipment, and save energy. The current mainstream detection methods cannot easily detect the movement inside the ball mill. Mill load requires a more efficient and direct detection method. In this study, the SM ϕ500 mm×500 mm ball mill was taken as the research object. Through theoretical analysis and simulation, intelligent grinding media with an embedded triaxial acceleration sensor and physical properties similar to that of ordinary steel ball media were designed to identify the mill load, and grinding experiments with different filling rates and other grinding conditions were conducted. Results revealed that the filling rate and the material to ball ratio are the important factors affecting the −0.074 mm size products. Taking the grinding effect coefficient as an index to distinguish different load states and grinding effects, the best load state can be achieved under the conditions of 40% filling rate, 1∶37 material to ball ratio, and ~6 kg sample weight. The ball mill load was evaluated using the convolutional neural network (CNN) method and optimized support vector machine (SVM) models from the acceleration signal obtained by the intelligent grinding media. For the optimized SVM models, preprocessing of the acquired one-dimensional acceleration signal, including complementary ensemble empirical mode decomposition algorithm denoising, time-domain eigenvalue extraction, and sample entropy, was conducted. The feature vectors of mill load were included in the genetic algorithm and SVM (GA−SVM), grid search and SVM (GS−SVM), and partial swarm optimization and SVM (PSO−SVM) classification models for training. The research results revealed that the recognition accuracy of the PSO−SVM algorithm reaches 98.33%, but the training process tends to be tedious and time-consuming. For the CNN algorithm with excellent applicability in the field of image recognition, the detected acceleration signal data were converted into two-dimensional pictures and directly inputted into the CNN model based on the VGG19 network for classification and recognition. The classification and recognition accuracy of the mill load of the CNN method (i.e., 98.89%) was higher than that of the optimized SVM algorithm. Moreover, the calculation time of the CNN method was shorter than that of the optimized SVM algorithm. The ball mill load status identification method using the intelligent grinding media and CNN method could provide critical solutions and technical support for load detection and online evaluation.