Bearings fault alarms by condition monitoring can effectively avoids catastrophic accidents.The fault detection method based on data time series feature reconstruction can avoid the degradation of model accuracy caused by insufficient fault data because only normal data is used for training.However,the fault threshold determination in such methods depends on a large amount of historical data,which has a great impact on the detection accuracy.Therefore,a bearing adaptive threshold fault detection method was proposed based on deep SVDD-CVAE.A CVAE feature compression extraction framework was constructed with ConvLSTM as the basic unit for enhancement extraction of time series signals,so as to extract the weak features of bearing faults.The SVDD was combined to adaptively learn the feature space hypersphere to realize the adaptive determination of the fault detection threshold.Finally,the deep SVDD-CVAE framework was iteratively optimized by global error loss backpropagation.The experimental results show that the proposed method can effectively extract weak bearing fault features and adaptively determine the threshold value with a detection accuracy of 97.7% on IMS bearing dataset.