SA1D-CNN: A Separable and Attention Based Lightweight Sensor Fault Diagnosis Method for Solar Insecticidal Lamp Internet of Things

Abstract

Sensor faults can produce abnormal and spurious observations in the solar insecticidal lamp Internet of Things (SIL-IoTs) system. Early detection and identification of the sensor node's abnormality are critical to ensure the SIL-IoTs system's reliability. In this study, we propose a lightweight separable 1D convolution neural network that can be implemented in SIL-IoTs nodes to identify sensor faults, reduce detecting delay, and decrease data transmission. However, the reliability of data acquired by sensors is decreased because a SIL-IoTs node releases high voltage pulse discharge (a kind of electromagnetic interference) when pests collide with its metal mesh. This kind of data fluctuation impacts fault diagnosis accuracy. Consequently, fault-related feature maps and temporal signals are characterized via a novel time and channel attention module (TCAM) method, which contributes to separating electromagnetic interference noise from sensor faults of SIL-IoTs nodes. A real-world testbed is applied to validate the effectiveness of the proposed method on sensor fault diagnosis in the SIL-IoTs system. Experimental results demonstrate that the proposed method can detect four typical sensor faults with the best trade-off between accuracy (99.9% average accuracy and 97.6% average F1-score) and efficiency (351KB inference model size and 4.33W average energy consumption on Raspberry Pi).