Embedded Load State Classification Using Goertzel-Based Spectral Features and LoRaWAN Communication

This work presents the design and implementation of an embedded sensor system for the classification of electrical load operating states based on current signal analysis. The system performs local signal acquisition, feature extraction, and classification directly on a resource-constrained microcontroller, while transmitting only compact classification results via LoRaWAN. The load current is measured using a galvanically isolated current sensor and conditioned by an analog offset and amplification stage to match the input range of the microcontroller’s ADC. Spectral features are extracted using the Goertzel algorithm, targeting the mains frequency and selected harmonics, which enables efficient frequency-domain analysis with minimal computational overhead. These features serve as inputs to a compact fully connected neural network that classifies the operating state of the load. Model training and validation are performed offline using Julia, employing an identical signal processing pipeline to ensure consistency between training and embedded inference. The trained model parameters are subsequently integrated into the firmware, allowing real-time inference on the embedded device. Only the classification result and selected system parameters are transmitted via LoRaWAN to The Things Network backend, where the data is decoded and visualized. Experimental results demonstrate reliable discrimination between different operating states of a household appliance, while maintaining low communication bandwidth and energy consumption. The presented approach illustrates the feasibility of combining embedded signal processing, lightweight machine learning, and low-power wireless communication for intelligent sensor systems.