Predictive Maintenance for Bird Feeder Motors via Time-Series Current Analysis

Abstract

Automated outdoor electronic device systems have been deployed widely for their high efficiency. However, environmental vulnerabilities could be a severe issue. To be more specific, high humidity outdoors causes feed agglomeration, therefore leading to fatal problems like mechanical jamming and motor burnout. According to current techniques, although deep learning model achieves the high accuracy of diagnosis, the heavy load is not suitable for a low-cost edge artificial intelligence (AI) controller. At the same time, traditional signal processing methods based on threshold values can easily cause false alarms in a noisy environment. This paper adopts lightweight sensorless predictive maintenance method, combining it with the motor current characteristic analysis (MCSA) and random forest classifier, and extracting the variance and peak amplitude features via Fast Fourier Transform (FFT) in the frequency domain. The result shows that by this method, the accuracy on the test data is up to 96.80%, which is superior to the standard deep neural network (89.30%) and maintains an extremely low inference latency, which is suitable for edge AI deployment.