From raw signals to reliable electrochemical sensing: Data preprocessing strategies for machine learning supported energetic compound identification

The widespread use of energetic compounds in armed conflicts, terrorism, and criminal activities highlights the need for rapid, accurate, and field-deployable detection and identification methods. Electrochemical sensing offers a promising solution, as many energetic compounds contain nitro groups that undergo electrochemical reduction, generating compound-specific electrochemical fingerprints. However, overlapping signals and concentration variability complicate robust identification. In this study, we present a Random Forest-based machine learning algorithm to identify ten nitro-containing energetic compounds and one binary mixture using square wave voltammetry. Voltammetric responses were collected over a concentration range of 50 – 200 µg/mL using bare in-house screen-printed electrodes. Six Random Forest models were developed based on different input data: (1) extracted peak parameters, (2) raw voltammetric data, and (3) Discrete Wavelet Transform (DWT)-processed data. Models trained on raw voltammetric and DWT-processed data using default hyperparameters achieved the highest overall classification accuracy in tests with samples representative for real-life scenarios. Confidence scores enabled quantitative evaluation of model predictions, with the raw voltammetric data model delivering the most confident outcomes. This study demonstrates a novel concentration-independent, machine learning-based electrochemical strategy for the accurate energetic compound identification in field applications.