A Communication Aware Masked Denoising Autoencoder Framework for Reliability Aware and Fault Tolerant Gas Classification in Electronic Nose Systems


== Abstract ==

<p>Electronic noses (e-noses) have become a promising technology for gas classification in environmental and industrial safety and smart sensing. However, sensor noise, redundancy, drift, and partial sensor failures usually impair performance and reduce reliability in practice. Communication-wise, the multi-sensor array could be viewed as a distributed sensing and information transmission system, noise, sensor failures, and drift are considered channel impairments. To overcome these issues, this paper presents a reliability-conscious, fault-tolerant e-nose model built around a masked denoising autoencoder (MDAE) with probabilistic classification. The approach is based on a masking process that causes the model to learn strong latent representations that can be used to recreate lost or corrupt sensor signals. These acquired representations are then used to perform correct multi-class classification. The framework was evaluated using an MQ gas sensor dataset containing 40,706 samples. After exact duplicate removal, 28,441 unique samples from five valid gas classes were retained for leakage-controlled evaluation. The proposed MDAE+ calibrated Fusion SVM achieved 100.00% accuracy, macro-F1 of 1.000, macro-AUC of 1.000, Brier score of 0.000015, NLL of 0.002172, and ECE of 0.002167. These results confirm excellent classification and probability calibration. While some calibrated baselines performed slightly better on Brier, the proposed framework successfully integrates denoising reconstruction, latent feature learning, sensor fault restoration, and sensor dropout and noise robustness in a novel manner. The efficacy of the masking strategy and denoising architecture is shown through ablation studies, and the stability and generalizability of the proposed framework are verified through 5-fold cross-validation. Furthermore, the robustness analysis shows that the proposed framework performs very well even in the presence of sensor drops and maintains around 72%&ndash;74% accuracy even in severe multi-sensor drops. It exhibits graceful degradation under increasing Gaussian noise, corresponding to a resilient communication system over increasingly degraded channels. The importance of features based on permutation provides a key insight into the importance of particular sensors and justifies the model&rsquo;s learning ability in terms of redundancy awareness. Overall, the proposed framework achieves high classification accuracy and demonstrates robustness, reliability, interpretability, and fault tolerance, and can be applied to intelligent sensing and communication-aware systems in practical applications.</p>



== Document ==
<pdf>Media:Draft_content_349812693-1748-document.pdf</pdf>