Convolutional Neural Network-Based Feature Extraction for Automated Pollutant Classification in Electrochemical Environmental Sensor Data

The article presents a multi-scale convolutional neural network, EC-PollutantNet, which automatically identifies environmental pollution using the electrochemical sensor data. The suggested model uses parallel convolutional subnets with the kernel size of 3, 5, and 7 to extract the features at various scales of time and combine them with the squeezing-and-excitation attention units to allow channel-wise weighting. The key novelty of EC-PollutantNet lies in its electrochemical-domain-specific multi-scale design: parallel convolutional subnets with kernel sizes of 3, 5, and 7 are specifically chosen to align with the characteristic frequency ranges of voltammetric signals, capturing sharp redox peaks (k = 3), medium-width overlapping features (k = 5), and broad diffusion-limited waveforms (k = 7) in a design not previously applied to electrochemical pollutant classification. A two-headed classification approach justifies the quality of features with the help of an end-to-end SoftMax classifier and a Convolutional Neural Network–Linear Discriminant Analysis (CNN-LDA) hybrid classifier, where the LDA head operates on pooled latent vectors extracted from the global average pooling layer. On a collection of 4,600 voltammograms of six pollutant classes (Pb2+, Cd2+, Hg2+, Cu2+, phenol, and mixtures), EC-PollutantNet achieves 94.8% classification accuracy and a macro-F1 score of 0.95, surpassing traditional machine learning baselines by 8.3% and single-scale CNN by 3.3%. No data augmentation with synthetic voltammograms was employed; all augmentation strategies (amplitude scaling, baseline wandering, peak shifting, additive noise, and peak broadening) are applied to real measured signals. Class imbalance is addressed through stratified sampling and macro-F1 evaluation. The model has a compact format (86K parameters) that allows inference times of less than 2 ms, which is suitable for real-time environmental monitoring scenarios. Ablation results confirm that preprocessing contributes 5.1%, multi-scale architecture 3.3%, and attention mechanisms 1.0% to the overall performance gain. The model demonstrates strong cross-electrode generalization with an accuracy decrease of less than 4%, validating its applicability across different electrochemical sensing systems.