<p>Fault detection in rotating machinery is a critical aspect of ensuring the reliable operation of large-scale rotating equipment. Failure to detect faults in a timely manner may result in reduced system performance and effectiveness, potentially leading to safety incidents. Recently, blind deconvolution methods have been widely adopted for fault feature extraction in rotating machinery. The blind deconvolution method aims to extract the original fault signals from the collected signals by designing an appropriate deconvolution filter. Nevertheless, these approaches heavily depend on prior knowledge and are prone to converging to local optima, which restricts their practical applicability in rotating machinery. This study proposes the wavelet threshold ensemble empirical mode decomposition and improved multiobjective optimal deep deconvolution approach to overcome the limitations of blind deconvolution methods in fault detection for rotating machinery. The enhancement of the multiobjective optimal deep deconvolution method is achieved through a refined loss function that improves the processing of both noise and feature components. This modification enables the model to dynamically adapt to the intrinsic characteristics of the input signal. Furthermore, the integration of wavelet threshold ensemble empirical mode decomposition enhances the model’s robustness by effectively attenuating noise while retaining critical signal features. The results demonstrate that the wavelet threshold ensemble empirical mode decomposition and improved multiobjective optimal deep deconvolution method exhibits superior performance in fault feature extraction, noise suppression, and diagnostic accuracy, thereby ensuring the safe and reliable operation of rotating machinery.</p>