In rotating machinery, the contamination of lubricating oil by metal particles can lead to oil film failure, which in turn accelerates equipment wear. Due to its sensitivity to weak signals, acoustic emission technology is particularly useful for early fault detection, where traditional monitoring methods fall short. Therefore, this study proposes a filtering approach that combines a median index with discrete wavelet transform (MI-DWT) for acoustic emission signals, greatly enhancing the signal-to-noise ratio. Furthermore, an adaptive second-order synchrosqueezing wavelet transform (AS2-WSST) is introduced to convert the filtered signal into a time–frequency representation that is then fed into a CNN for identification. Experimental results indicate that, with learning rates of 0.001 and 0.0001, the recognition accuracy reaches 92.35% and 100%, respectively.