Predictive maintenance plays a critical role in reducing unplanned downtime and extending the operational life of industrial equipment. Thermal imaging, with its ability to detect early-stage faults via abnormal heat signatures, offers a powerful non-contact diagnostic tool. However, existing solutions often rely on cloud-based processing or manual inspection, which limits their responsiveness and scalability. This paper presents an embedded AI system for real-time thermal image analysis aimed at fault diagnosis and monitoring, deployed on a low-power ESP32-S3 microcontroller. Two lightweight convolutional neural network (CNN) architectures, Sequential Shallow Model (SSM) and Deep Structured Model (DSM) were developed and trained on a labeled dataset of motor fault thermal images. The models were converted to TensorFlow Lite and deployed in both float32 and int8 formats for on-device inference. A built-in web interface supports direct image upload, preprocessing, and fault classification visualization, all without reliance on external infrastructure. Quantized models achieved up to an 8 \(\times \) speedup in inference (as low as 23.8 ms), with memory usage reduced to under 20 KB, while maintaining classification accuracy (F1-score up to (98.6%)) on par with their full-precision counterparts. The proposed system demonstrates the feasibility of fast, compact, and autonomous AI-driven thermal monitoring at the edge, offering a scalable solution for industrial predictive maintenance aligned with Industry 4.0 goals.