Convolutional neural networks (CNNs) have emerged as powerful tools for pattern recognition and fault diagnosis, owing to their exceptional feature extraction capabilities. In the realm of rotating machinery fault diagnosis, it is imperative for intelligent models to exhibit robust interpretability and resilience to noise. To this end, numerous studies have integrated traditional signal processing methods, such as wavelet and Fourier transforms, to bolster these attributes. In light of this, a novel adaptive wavelet feature fusion network (AWFFN) has been developed to distill salient features from noisy signals. This architecture is designed to address the challenge of noise vibration signals in rotating machinery, offering a fewer parameter set compared to conventional CNNs while maintaining a high tolerance for noise. The AWFFN employs an adaptive lifting scheme to decompose noisy signals, enhancing feature extraction without the need for wavelet selection, a departure from traditional wavelet-based CNNs. Furthermore, the framework incorporates a mechanism akin to multi-resolution analysis, facilitating the fusion of multi-scale features and enabling the model to efficiently identify fault-related features with fewer parameters. Experimental outcomes from both gearbox and motor bearing datasets attest to the framework’s superior performance, underscoring its proficiency in fault diagnosis, noise resistance, and interpretability.