Hybrid Butterworth Filter and Autoencoder for Time-Series Anomaly Detection in Agricultural IoT Sensors

Internet of Things (IoT) sensors deployed in agricultural environments frequently suffer from noise interference caused by environmental fluctuations, which degrades the performance of deep-learning-based anomaly detection. This work presents a hybrid pipeline combining a Butterworth low-pass filter and an autoencoder to detect anomalies from sensor time-series data. The Butterworth filter suppresses high-frequency environmental noise before data enters the autoencoder, improving the quality of input features. The autoencoder reconstructs normal sensor signals and identifies anomalous samples by measuring reconstruction error. We validate the proposed method on the NAB real-world sensor dataset and conduct ablation experiments to analyse the contribution of the filtering module. Experimental results show that the pre-filtering step effectively reduces false positives and improves detection accuracy compared with using a raw-data autoencoder baseline. This hybrid approach is lightweight and suitable for edge agricultural IoT devices with limited computational resources.