Wavelet transform (WT) is a widely applied method in signal processing that can transform the original time-domain signal into a time–frequency representation. Traditional bearing fault diagnosis methods based on WT often rely on predefined kernel functions, limiting their adaptability to extract intrinsic signal characteristics. To address the challenge, this paper proposes a multi-level feature extraction network (MLFEN) that synergizes adaptive wavelet decomposition with convolutional neural network (CNN). The proposed MLFEN introduces a trainable multi-level feature extraction layer, where learnable wavelet kernel functions can adaptively decompose raw vibration signals into multiple sub-bands. This layer extracts fault-related features at different frequency scales, which are input into the subsequent CNN backbone to be further extracted and classified. The experimental validation on the XJTU-SY bearing dataset demonstrates the superior diagnostic performance of the proposed MLFEN. Furthermore, the loss curve and the features visualization demonstrate the ability of MLFEN to converge rapidly during the training process and its excellent clustering ability for different fault categories.