With the rapid advancement of Industrial Internet of Things (IIoT) technology, predictive maintenance for equipment in the oil and gas recovery industry has become a key technology for improving equipment reliability and reducing maintenance costs. However, traditional predictive maintenance methods based on centralized cloud computing rely heavily on cloud processing of large amounts of device-generated data. This approach often leads to network congestion and excessive computational delays, particularly when the data volume at the device end increases dramatically, thus failing to meet the demands of latency-sensitive and computation-intensive tasks. To address this issue, this paper proposes a cloud-edge-end collaborative predictive maintenance (PDM) framework based on adaptive Deep Neural Network (DNN) inference acceleration and introduces a dynamic partitioning algorithm to optimize delay and resource consumption during DNN inference. By coordinating the cloud, edge, and terminal devices, the framework efficiently allocates computational tasks and schedules resources, significantly reducing inference latency and improving resource utilization. Experimental results show that, under various network conditions and data volumes, the proposed method significantly reduces inference delay and enhances prediction accuracy compared to traditional predictive maintenance methods. Overall, the proposed framework demonstrates significant advantages in terms of inference latency and prediction accuracy, effectively improving the efficiency of PDM tasks in the oil and gas recovery industry.