<![CDATA[Sensor readings in open field monitoring systems are influenced by disruptions, degradation, and operational unreliability. These conditions may result in inaccurate data and unreliable system decisions. However, existing studies focus on detection accuracy and rarely examine the trade-off between detection performance and computational efficiency of Artificial Neural Networks trained using the Levenberg–Marquardt algorithm (ANN–LM) in smart farming environments. This study evaluates the fault-detection capability of ANN–LM for soil moisture sensor readings by analyzing both detection performance (accuracy, precision, recall, and F1-score) and computational efficiency (execution time, CPU usage, and memory consumption), thereby addressing the trade-off between performance and efficiency. Baseline data, hypothetical dataset that represent the soil moisture reading from a smart chilli pepper farming system in normal operating conditions, were used to generate fault-injected datasets representing four common faults: drift, bias, spike, and malfunction. The ANN–LM model was evaluated under five fault-detection scenarios with different network architectures. Model performance was evaluated using accuracy, precision, recall, and F1-score, while computational cost was assessed through execution time, CPU usage, and memory usage. The results show that ANN–LM achieves an accuracy of 0.996–0.999, precision of 1.000, recall of 0.987–1.000, and F1-scores of 0.992–1.000 across all scenarios. Simple ANN architectures give accuracy of 0.997 with reduced execution time (33.74 seconds) and lower CPU usage (50.50%) compared to more complex architectures that require 591.88 seconds and 78.40% CPU usage. Therefore, these results indicate point out that ANN–LM is suitable for smart agricultural systems under resource-constrained conditions.]]>
Copyrights © 2026
