Under the complex operating conditions of large rotating machinery, bearings can suffer compound failures. To address the challenge of extracting fault features in heavy noisy environments and accurately identifying bearing conditions, adaptive feature-oriented dictionary learning (DL) and sparse classification framework are proposed for bearing compound fault diagnosis. Adaptive fault feature screening is accomplished by intraclass and interclass distance metrics to eliminate redundant feature sets to construct feature training and testing sets. Subsequently, an objective function for multiclassification is constructed utilizing $l_{1}$ -norm constraint to ensure convexity and enhance the solution properties. Finally, after the optimal dictionary is derived by iteration, the classification vector is computed for effective classification. The proposed method is validated to effectively solve the adaptive feature selection problem in the label-consistent $K$ -means singular value decomposition (LC-KSVD) algorithm by simulated signals, rolling bearing acquisition signals from tractor gearbox, and gearbox test bench. Compared with traditional methods, the proposed method exhibits higher classification accuracy and increased robustness to background noise, thus providing valuable insights for practical engineering applications.