Introduction


This investigation presents a rigorous experimental evaluation of a low-cost electronic nose (E-nose) framework, coupled with multivariate chemometric and supervised machine learning algorithms specifically Principal Component Analysis (PCA), Random Forest (RF), and Artificial Neural Networks (ANN) for the identification and discrimination of agricultural biofuel blends. The primary objective is to critically appraise the operational validity and physical limitations of static metal-oxide semiconductor sensing for fuel quality assurance, thereby establishing an empirical baseline to prevent publication bias and guide the robust deployment of bio-based renewable fuels in compression ignition engines.


Materials and Methods


An electronic nose system comprising an array of eight commercial metal-oxide semiconductor (MOS) gas sensors (MQ series: MQ2, MQ3, MQ5, MQ8, MQ9, MQ135, MQ137, and MQ138) was developed and evaluated across 15 distinct fuel classes, including neat fossil diesel (D100), waste cooking oil biodiesel (B100), fusel oil (F100), and their corresponding binary and ternary blends. A strictly balanced dataset of 225 experimental trials was established and partitioned into a training set (n=150) and an independent test set (n=75, with 5 replicates per class). Static steady state voltage equilibrium responses were acquired and standardized. Data structure and variance distribution were explored using unsupervised Principal Component Analysis (PCA), while supervised multi class classification and hyperparameter tuning were conducted using Random Forest (RF) and a Multilayer Perceptron Artificial Neural Network (ANN/MLP, 8-4-15 architecture). Sensor contribution was quantitatively assessed via Mean Decrease in Impurity (MDI), and dimensional sensitivity was examined by contrasting the full eight sensor array against an optimized four sensor subset.


Results and Discussion


Unsupervised exploratory analysis demonstrated that although the first two principal components captured 74.96% of the cumulative variance (PC1: 46.10% and PC2: 28.86%), extensive spatial cluster overlap persisted across all 15 fuel treatments without evident class separation boundaries. Supervised classification on the independent test set yielded poor discrimination, with RF and ANN attaining overall test accuracies of only 6.67% (95% CI: [2.88%, 14.68%]) and 8.00% (95% CI: [3.72%, 16.37%]), and Cohen’s kappa coefficients of κ=0.0000 and κ=0.0143, respectively. Both models converged near the theoretical random chance baseline (1/15≈6.67%), with McNemar’s test confirming no statistically significant performance difference (p=1.0000). Random Forest feature importance revealed a homogenous impurity reduction across the array (MDI spanning narrowly between 0.1008 for MQ9 and 0.1465 for MQ8), evidencing pronounced sensor cross sensitivity and strong collinearity. Furthermore, pruning the array to four sensors marginally altered RF and ANN test accuracies to 10.67% and 5.33%, respectively, demonstrating the adverse impact of collinear dimensions (curse of dimensionality) under small sample constraints and confirming that scalar equilibrium signals lack the kinetic resolution required to decouple highly overlapping volatile organic compound profiles.


Conclusion


The findings demonstrate that low cost MOS electronic nose systems relying solely on steady state equilibrium readings cannot achieve granular 15 class discrimination among chemically proximate diesel, biodiesel, and fusel oil blends, with machine learning complexity failing to compensate for baseline informational deficits. Consequently, future developments must transition from scalar static signals to dynamic transient response profiling, sensor temperature modulation, and continuous chemometric regression modeling to establish dependable, real-time quality control protocols in biofuel refining.