As the complexity and degree of automation of industrial equipment continue to increase, traditional corrective and time-based preventive maintenance paradigms can no longer meet the stringent reliability and safety requirements of modern industrial production. To address the pervasive bottlenecks of existing detection algorithms—namely sensitivity to the quantity of anomalous samples and low computational efficiency that hinders real-time deployment in industrial settings—this paper proposes an innovative predictive maintenance method that fuses multi-source sensor signals, including vibration, temperature, voltage, and current. The method builds a dual-branch predictive network architecture that deeply mines multi-dimensional sensor features under healthy operating conditions, precisely capturing latent patterns of normal operation to enable reliable prediction of expected normal readings at the current time step. Based on the distribution of absolute errors between predictions and measurements, a Gaussian statistical model of normal behavior is established. When monitored data significantly deviate from this model, the system automatically determines an abnormal equipment state and quantifies anomaly severity according to the degree of deviation, thereby generating tailored response strategies. Experimental results show that the proposed dual-branch network achieves a mean absolute error as low as 0.069, representing a 2.53-fold reduction compared with Informer, and an inference speed of 1593 SPS, demonstrating excellent predictive accuracy and real-time performance. Under abnormal operating conditions, the algorithm can automatically trigger corresponding handling mechanisms based on the severity of sensor-data anomalies. This work provides an efficient solution for intelligent operation and maintenance of industrial equipment and offers significant engineering application value.