Industrial IoT (IIoT) systems consist of a myriad of sensors that generate considerable amounts of data that must be analyzed intelligently and in a timely fashion at the edge. The primary challenges in deploying machine learning models on edge devices are the limited computational power and the lack of sufficient labeled data. This work tackles the problem of selfsupervised learning (SSL) on resource-constrained intelligent edge devices, solving the problems of resource limitation and annotation bottleneck. The architecture incorporates domain-specific pretext tasks for industrial sensor modalities such as vibration, pressure, and temperature to construct embedding features without requiring human-labeled data. We deploy and evaluate the model within a heterogeneous IIoT testbed that consists of real-world edge devices and measure performance based on embedding quality, accuracy of downstream tasks, energy consumption, and latency. The results show that the proposed approach outperforms baseline supervised and semi-supervised federated learning models in sparse label conditions while achieving near real-time inference and low power consumption. This work assists in the deployment of scalable self-supervised intelligence at the edge for predictive maintenance, anomaly detection, and context-aware automation in future industrial systems.