Research on Infrared Image Denoising and Fault Diagnosis Methods for Rotating Machinery

Rotating machinery sits at the heart of petrochemical, power generation, and other heavy industries. Its reliability directly shapes production output and cost. Infrared thermography enables non-contact, real-time monitoring of equipment surface temperatures. This makes it a useful tool for early fault detection. Yet images captured in factories typically carry mixed noise from sensors, transmission, and the environment. Such noise often obscures the thermal signatures that indicate faults. This paper puts forward a complete edge-deployable pipeline that denoises infrared images and then diagnoses faults. The first stage is a denoising method built on the stationary wavelet transform (SWT) and a parallel dual-branch U-Net. Under mild, moderate, and severe mixed noise, it reaches 40.99, 35.08, and 31.76 dB PSNR, respectively. The second stage is a lightweight CNN-Transformer-CBAM classifier. It uses depthwise separable convolution and a CBAM attention module. On an 11-class induction motor dataset, it scores 98.67% accuracy (five-run average). The model also achieves 98.75% on a 9-class transformer dataset and 99.50% on a 5-class pump dataset, demonstrating cross-device generalization. The model occupies only 2.758 MB. After INT8 quantization, it runs on the RK3588S edge board. Average CPU inference takes 103.026 ms. The 99th-percentile latency stays under 124.778 ms. These numbers support the feasibility of factory-floor monitoring.