关键词: 多尺度自注意力, 红外图像, 波段 RX 加权, 永磁电机, 温度异常点检测

Abstract: Due to current overload, poor heat dissipation, or material aging, temperature anomalies may occur in permanent magnet motors. It was necessary to enhance the contrast between the anomalies and normal areas, remove background interference and effectively identify temperature anomalies. To address this, a method for detecting temperature anomalies in permanent magnet motors was designed using the infrared image band RX weighting algorithm. Use an infrared thermal imager to conduct a comprehensive scan of the permanent magnet motor and obtain its surface infrared image. Construct a variational adversarial auto encoder model based on multi-scale self attention, implement hyperspectral
unmixing and extract background information from infrared images. Design a feature space RX weighted anomaly detection algorithm to detect temperature anomalies in permanent magnet motors. The test results show that the design method can achieve temperature anomaly detection of permanent magnet motors, accurately distinguish between real anomaly points and normal points, reduce false alarm rates, and have a high AUC value. The actual number of pixels occupied by the real abnormal target is 182, and the design method detects 181 pixels of the real abnormal target, with an error of only 1 pixel.

Key words: multi-scale self-attention, infrared image, band RX weighting, permanent magnet motor, temperature anomaly detection