Anomaly Detection in Sensor Data for Tomato Farming Using a Federated Learning - Autoencoder

Abstract

Anomalies in tomato farming sensor data may interfere with soil condition monitoring and decision-making processes. This study developed a Federated Learning-Autoencoder model to detect anomalies in tomato farming soil sensor data. The public dataset consisted of three clients or sensor lines with humidity, temperature, and electrical conductivity parameters. The research stages included preprocessing, local model training, Federated Learning model training using Federated Averaging, model evaluation, and dashboard-based monitoring simulation. The evaluation used a confusion matrix, precision, recall, and F1-score. The local L01 P95 model obtained an average precision of 1.0000, an average recall of 0.8291, and an average F1-score of 0.9061. The Federated Learning F01 P95 model obtained an average precision of 1.0000, an average recall of 0.7865, and an average F1-score of 0.8803. The results showed that the Federated Learning model detected spike anomalies competitively without combining raw data among clients.