Multivariate Time Series Classification in Resource-Constrained Embedded Systems

Monitoring systems in oil wells generate large volumes of multivariate time-series data that often need to be used for local decision-making, motivating the use of embedded machine learning solutions for real-time anomaly detection at the edge. Implementing machine learning models on resource-constrained single-board computers requires efficient inference pipelines that balance accuracy, power consumption, and memory usage. This paper presents a complete pipeline for implementing quantized multilayer perceptron models on ESP32 boards for anomaly detection in oil wells. The pipeline covers data acquisition, feature extraction, model training, and post-training quantization with INT8, INT16, and INT32 representations. Experimental results show that while the unquantized model achieves 95% accuracy, the quantized versions degrade accuracy by at least 6% in the INT8 case. On the other hand, the quantized models reduce energy consumption per inference, with gains of up to 55% for INT8 and ~42% for both INT16 and INT32.