To address the issues of low efficiency and insufficient accuracy in traditional mining equipment anomaly detection, a mining equipment anomaly detection model integrating bidirectional long short memory network and an improved subtraction average optimizer was proposed. Attention mechanism was introduced to achieve dynamic weighting of key features. Convolutional neural network was combined to extract local temporal features, and guided regularized random forest was employed to perform feature selection, thereby reducing data dimensionality and suppressing noise interference. The improved subtraction average optimizer introduced chaotic mapping initialization and golden-sine step-size adaptive adjustment strategy, which significantly enhanced the convergence speed and training stability of the model. The results show that on the mining equipment multisensor dataset and the mining machinery fault simulation dataset, the accuracy of the proposed model reaches 94.17% and 95.48%, respectively, and its comprehensive performance surpasses that of existing mainstream detection models. In ablation experiments, when all key modules are integrated, the accuracy, precision, recall, and F₁ score of the model reach 98.25%, 97.62%, 98.05% and 97.84%, respectively. In field tests on key equipment such as belt conveyors and crushers at a large open-pit mine, the detection accuracy of the model increases to 97.32%, with an average detection time of only 0.39 s. This indicates that the model can detect abnormal states of mining equipment efficiently and accurately, providing reliable technical support for ensuring mining production safety.