This research introduces an new fault diagnosis framework for detecting faults in milling machines (MM). To accurately capture transient fault characteristics, the proposed method begins by preprocessing acoustic emission (AE) signals through wavelet packet transform (WPT). Unlike conventional methods that extract features from a single WPT node, this framework utilizes all available WPT bases to extract a rich set of features. Additionally, since time-domain (TD) features are effective in identifying weak faults, key statistical attributes are also extracted from AE signals in the time domain. These extracted features are then consolidated into a comprehensive feature matrix (CFM) for further analysis. However, the resulting high-dimensional feature space may include redundant or less-informative attributes, which could impact classification accuracy. To refine the feature set, an adaptive feature refinement (AFR) strategy is employed. This strategy evaluates the relevance of each feature based on inter-class compactness and intra-class separability, ensuring that only the most discriminative features are retained. The identified features are subsequently analyzed using the K-nearest neighbor (KNN) algorithm for classification. Experimental findings demonstrate that the proposed approach outperforms current advanced techniques in detecting milling machine faults, achieving an accuracy of 96.2%, offering improved accuracy and reliability.