This paper presents deep learning-based prognostics and health management (PHM) for predicting fractures of an electric propulsion (eP) drivetrain system using real-time CAN signals. The deep learning algorithm, based on autoencoders, resamples time-series signals and converts them into 2D images using recurrence plots (RP). Subsequently, through unsupervised learning of DeepSVDD, it detects anomalies in the converted 2D images and predicts the failure of the system in real-time. Also, reliability analysis based on fracture mechanics was performed using the detected signals and big data. In particular, the severity of the eP drivetrain system is proportional to the maximum shear stress (τmax) in terms of linear elastic fracture mechanics (LEFM) and can be calculated by summarizing the relationship between cracks (a) and the stress intensity factor (KIII). During this process, the system status can be checked by comparing the stress intensity factor and fracture toughness (KIIIc), and the time from the detection of an abnormal signal in the system to complete failure can be quantitatively determined. Therefore, it is possible to continuously maintain the status of the system by detecting failure signals using deep learning before vehicle parts fail, and with the detected failure prediction signals, a process can be established to enable users to repair defects in the vehicle system before breakdown occurs. By predicting the remaining life of the system and calculating field reliability through these procedures, we introduce innovative technologies aimed at preventing safety accidents, reducing economic costs, and addressing quality issues. In the future, we expect to achieve high business performance by extending and applying this deep learning-based PHM approach to all vehicle components.