Abstract: To address the issue of unclear and challenging identification of non-contact rotating seal fault signals， this study established an experimental platform and acoustic emission testing system. It involved monitoring acoustic emission signals during various operational conditions， including normal operation and six typical fault scenarios of non-contact rotating seals. A total of 14000 feature samples were effectively collected. By applying the Bayesian optimization algorithm and incorporating continuous wavelet transform， an adaptive convolutional neural network classification model was constructed. Subsequently， the diagnostic performance of the fault recognition model was analyzed using confusion matrices and t-distributed stochastic neighbor embedding. The research results demonstrate that this model successfully classifies and identifies seven different operational conditions of non-contact rotating seals， including normal operation， dry friction， mixed lubrication， spring failure， end-face pitting， local spring failure， and end-face scratching， with an average recognition accuracy of 99.7023%. This achievement underscores the capability of effectively isolating and identifying seal fault sources from acoustic emission signals of non-contact rotating seals in non-stationary， complex， and overlapping environments， thereby establishing a solid theoretical foundation for practical engineering applications.