Gasoline Adulteration Analysis Based on Raman Spectroscopy and a Multi-Scale Convolutional Neural Network

Abstract

The quality and safety of gasoline are critical for energy distribution and market regulations. Consequently, rapid and nondestructive detection methods based on spectroscopy combined with intelligent algorithms have garnered significant attention. However, existing approaches often separate adulterant identification and concentration prediction, limiting their application in complex adulteration scenarios. To address these limitations, this study develops an attention-based multi-scale convolutional neural network, termed Multi-Scale Attention Fusion Convolutional Neural Network, based on Raman spectroscopy, to jointly perform adulterant classification and concentration regression within a unified framework. Four adulteration systemsgasoline blended with methanol, ethanol, kerosene, and dieselwere constructed, and 1200 Raman spectra were collected across adulteration ratios ranging from 5% to 30%. The performance of the model was systematically evaluated via ablation studies and comparisons with various backbone networks. Additionally, the proposed model was compared with multi-task deep learning models based on ResNet18 and LSTM architectures. The results indicate that the proposed model achieves more stable performance in both classification and regression tasks. In the ablation study, the removal of the Feature Pyramid Network module led to a 1.25% decrease in classification accuracy and a 29.21% increase in root mean square error, underscoring its critical role in multi-scale feature representation. Traditional single-task models were also evaluated for comparison. A Support Vector Machine, K-Nearest Neighbors, and Random Forest were applied to classification tasks, while Support Vector Regression (SVR) and Partial Least Squares Regression were used for regression. Even with optimal variable importance in projection wavelength selection, the best-performing SVR model achieved an R ² of only 0.9785. In contrast, the proposed model achieved a classification accuracy of 99.92% and an R ² exceeding 0.9886 in regression tasks, demonstrating superior qualitative and quantitative detection capability compared with conventional chemometric and single-task learning methods.

1. Introduction

Gasoline adulteration is a persistent global issue that primarily involves the addition of low-cost solvents such as kerosene and ethanol to high-quality gasoline for economic gain. , Such adulteration not only degrades the antiknock performance and combustion efficiency but also accelerates lubricant degradation and mechanical wear during prolonged use, thereby shortening the lifespan of internal combustion engines. Moreover, the combustion of adulterated fuels generates high levels of toxic pollutants, posing significant risks to the atmosphere and public health. Therefore, it is essential to develop analytical techniques capable of rapidly identifying adulteration and performing quantitative assessments in the context of increasingly stringent regulations regarding the fuel supply chain and environmental emissions.

In the context of fuel adulteration, gas chromatography and its multidimensional variants provide high-resolution chemical fingerprinting information. , Lee et al. employed fast gas chromatography combined with partial least-squares regression to simultaneously evaluate the gasoline octane number and concentrations of aromatic and oxygenated compounds within minutes, thereby enabling the identification of adulterated samples.

Fourier transform infrared (FTIR) spectroscopy, particularly attenuated total reflectance FTIR (ATR-FTIR), has been widely applied for the screening of gasoline and diesel adulteration because of its rapid analysis and minimal sample preparation. Moura et al. quantitatively identified kerosene adulterants in commercial diesel using infrared spectroscopy combined with partial least squares (PLS) regression. , Additionally, Wang et al. quantified the methanol content in methanol–gasoline blends using fluorescence excitation–emission matrix (EEM) spectroscopy combined with second-order calibration algorithms. Romanel et al. demonstrated that ethanol blending in gasoline can be detected by measuring the transverse relaxation time (T₂), which reflects molecular mobility and interactions within the fuel matrix. However, despite their demonstrated analytical capabilities, these techniques generally exhibit several limitations, including high instrumentation costs, complex operational or calibration procedures, and limited suitability for rapid on-site application. More critically, most existing approaches are designed either for single-adulterant quantification or for separate qualitative and quantitative analyses, thereby limiting their effectiveness in real-world scenarios involving complex mixtures of multiple adulterants at varying concentrations.

Raman spectroscopy, a molecular vibration-based fingerprinting technique, offers several advantages, including non-destructive analysis, minimal sample preparation, and high sensitivity to complex organic mixtures, making it a promising tool for fuel quality evaluation and adulteration detection. − Dantas, W. F. C., et al. employed Raman spectroscopy combined with correlation-constrained multivariate curve resolution–alternating least squares (MCR-ALS) to simultaneously identify and quantify multiple fuel adulterants. Li et al. investigated the combined use of mid-infrared and Raman spectroscopy for the qualitative identification and quantitative analysis of kerosene in gasoline. They also achieved accurate modeling of the methanol content in methanol–gasoline blends using portable Raman spectroscopy combined with feature selection and random forest regression. , Deep learning has demonstrated significant advantages in Raman spectral denoising, baseline correction, feature extraction, and classification and regression of complex spectral data. It has also enabled the integrated identification of adulterant types and quantification of adulteration levels in petroleum products through multi-task deep learning networks. − In recent years, the integration of spectroscopic techniques with advanced machine-learning and deep-learning methods has emerged as a prominent research direction in analytical chemistry. Recent studies have explored convolutional neural networks, attention mechanisms, and spectral data fusion strategies to improve the feature extraction and predictive performance. Nevertheless, many of these approaches rely on single-scale feature representations, use separate models for classification and regression tasks, or lack effective mechanisms to emphasize chemically informative spectral regions, which restricts their capability for integrated qualitative and quantitative analysis in complex gasoline-adulteration systems.

Accordingly, there is a pressing need for a unified analytical framework capable of simultaneously performing qualitative and quantitative identification, effectively capturing multi-scale spectral features, and enhancing model interpretability and robustness. Existing Raman spectroscopy-based studies have three primary limitations. From the perspective of problem formulation, most methods treat adulterant classification and concentration prediction as separate tasks rather than as unified problems. From a spectral modeling perspective, single-scale or single-task models are insufficient for capturing both local peak variations and global compositional trends. From a chemical interpretation perspective, the model outputs are often weakly correlated with chemically meaningful spectral features. To address these limitations, this study develops an attention-based multi-scale convolutional neural network (MSAF-CNN) that integrates gasoline adulterant classification and concentration regression into a unified framework, rather than treating them as independent analytical tasks. The contributions of this study can be summarized as follows: First, regarding analytical problem formulation, gasoline adulteration detection is framed as a unified qualitative–quantitative task for multi-adulterant systems rather than treating classification and regression separately. This formulation provides a more comprehensive analytical strategy compared with conventional approaches that mainly focus on individual detection objectives. Second, in terms of the spectral modeling strategy, the proposed MSAF-CNN integrates attention mechanisms with a multi-scale feature pyramid within a multi-task framework, enabling the joint extraction of local spectral features and global compositional information. Compared with conventional single-scale feature extraction methods, this strategy improves the capability of representing spectral variations associated with different adulterants and concentration levels. Third, in terms of chemical interpretability, the model links sensitive spectral regions to Raman-active vibrational features associated with different adulterants, thereby providing a partially interpretable connection between data-driven modeling and molecular composition. Compared with traditional chemometric approaches and conventional single-task deep learning models, the proposed method achieves unified qualitative and quantitative analysis within a single end-to-end architecture, reducing model redundancy while improving the predictive accuracy and computational efficiency.

From an analytical chemistry perspective, this study integrates vibrational spectroscopy with data-driven intelligent modeling by introducing a rapid, nondestructive, and highly integrated analytical strategy. The proposed method further enhances the capability of Raman spectroscopy to analyze the complex mixtures.

In summary, this study addresses the key limitations of existing methods, including single-task constraints, insufficient multi-scale feature representation, and limited adaptability to multiadulterant systems, and provides a practical and scalable framework for integrated qualitative and quantitative analysis of gasoline adulteration. By combining unified task formulation with multi-scale spectral feature learning, this work provides a distinct analytical framework compared with previous Raman-spectroscopy-based gasoline adulteration studies. The proposed approach is expected to be further extended to more specific applications and to support the large-scale detection and analysis of the gasoline quality.

2. Materials and Methods

2.1. Instrument and Sample Acquisition

In gasoline adulteration practices, methanol, ethanol, diesel, and kerosene are among the most commonly used adulterants; therefore, they were selected as the target analytes for gasoline adulteration detection in this study. Methanol and ethanol are low-cost oxygenated organic solvents with good miscibility in gasoline within certain proportion ranges and are often illicitly added to increase the fuel volume or enhance apparent combustion characteristics. However, their addition significantly alters the chemical composition and combustion performance of the gasoline. Diesel and kerosene are petroleum distillates that share similarities with gasoline in origin and partial composition, making it difficult to distinguish them after blending on the basis of appearance or simple physicochemical indicators. Their presence can adversely affect engine performance and emission characteristics. These adulterants are representative in terms of their motivation, chemical composition, and impact on the gasoline performance. Their selection enables a systematic investigation of how different adulteration types influence the Raman spectral characteristics of gasoline, thereby enhancing the practical applicability and reliability of adulteration detection methods.

Gasoline samples of grades 92#, 95#, and 98#, as well as 0# diesel, were collected from two gas stations in the Banan District of Chongqing, China. Methanol, ethanol, and kerosene were purchased online as analytical reagent (AR)-grade chemicals. Considering that fuel adulteration in practice is typically conducted at low proportions to avoid regulatory detection, and that high adulteration levels are readily noticeable, the adulteration ratios in this study were limited to a maximum of 30%. Gasoline was blended with different adulterants at volume ratios of 95:5, 90:10, 85:15, 80:20, 75:25, and 70:30 to prepare four adulterated samples: ethanol–gasoline (GE), methanol–gasoline (GM), kerosene–gasoline (GK), and diesel–gasoline (GD).

The specific preparation procedures are as follows. First, 4.5 mL of ethanol was mixed with gasoline (0.5 mL) in clean test tubes. The mixture was shaken for 2 min and then transferred to a liquid chromatography vial. The remaining adulterated samples were prepared by using the same procedure. In total, 240 samples were collected.

The Raman spectra of all samples were acquired using an Optosky ATR3000-785 nm portable Raman spectrometer. The laser power was set to 300 mW and the integration time per scan was 1 s, with three scans performed for each measurement. Spectra were collected over the range of 200–2600 cm^–1. For each sample, five spectra were recorded at different locations, and 50 spectra were obtained for each adulterant proportion, resulting in a total of 1200 Raman spectra. To ensure an objective evaluation of the model’s true generalization performance and to avoid potential data leakage, we adopted a strict sample partitioning method to ensure that repeated measurement data from the same sample are not shared between the training and validation sets.

2.2. Adulteration Identification Model Design

Raman spectra, as one-dimensional signals, exhibit distinct structural patterns across multiple spectral scales. Fine-scale regions capture subtle peak features and their morphological variations, whereas broader scales describe the overall spectral distributions and global compositional trends. Relying on a single-scale representation is insufficient for concurrently encoding localized discriminative features and holistic spectral information. In contrast, multi-scale modeling provides a more complete representation of spectral complexity and facilitates improved learning of intricate spectral patterns. In the context of gasoline adulteration detection, variations in fuel composition and adulteration levels manifest differently in both the local peak structures and the global spectral profiles. These heterogeneous characteristics can be characterized more effectively using multi-scale feature representations. To address this, an attention-integrated, multiscale, multitask deep learning framework, MSAF-CNN, was developed. The proposed model leverages cross-scale spectral information to simultaneously perform classification and concentration prediction, integrating fine-grained feature localization with global contextual understanding, thereby improving both predictive accuracy and model robustness.

The overall architecture of the MSAF-CNN model is shown in Figure a. The model uses a preprocessed gasoline Raman spectral sequence x ∈ R ^{B ×1× L} as input, where B represents the batch size and L denotes the spectral length. Initially, multi-scale feature extraction was conducted using a one-dimensional FPN enhanced with a Convolutional Block Attention Module (CBAM). In the bottom-up pathway, spectral features spanning local to global scales are progressively extracted with channel-wise and spatial attention mechanisms applied at each level to highlight informative spectral peaks while reducing noise interference. In the subsequent top-down pathway, higher-level semantic features are iteratively upsampled and fused with the corresponding lower-level features via lateral 1 × 1 convolutions, facilitating cross-scale feature integration and refinement. Finally, feature maps from all scales are upsampled to a uniform length and concatenated along the channel dimension, yielding a multi-scale fused representation F _FPN ∈ R ^{B ×1× L} .

1.

(a) MSAF-CNN overall structure of the model; (b) CBAM attention structure; (c) CBAM_FPN feature pyramid architecture.

To reduce feature dimensionality and enable efficient processing in subsequent fully connected layers, the fused feature maps were compressed from 512 to 256 channels using one-dimensional convolution. The compressed features were then transformed into a fixed-length spectral embedding F _C ∈ R ^{B ×256×64} via adaptive average pooling. Subsequently, the embedding was flattened and passed through a two-layer shared fully connected network to produce a task-shared representation h ∈ R ^{B ×256} .

Building on shared representation, the network branches into two task-specific heads: one for adulterant classification and the other for the concentration estimation. Joint optimization was achieved using a weighted combination of cross-entropy and mean-squared error loss functions. By incorporating attention mechanisms within a multiscale feature pyramid, the model captures both local spectral peak characteristics and global spectral trends, thereby enhancing the performance in both classification and regression tasks for complex spectral data.

2.3. Construction of Adulteration Identification Model

2.3.1. CBAM Module

One-dimensional Raman spectroscopy poses several challenges, including densely clustered spectral lines, significant noise interference, and an uneven feature distribution. To overcome these issues, recent research has incorporated attention mechanisms into deep learning models to extract salient spectral information more effectively. Among these approaches, the convolutional block attention module has been widely utilized for spectral feature extraction owing to its lightweight structure, low computational cost, and flexibility in integration with various network architectures.

The architecture of the CBAM module is shown in Figure b. In this design, channel and spatial attention are applied sequentially to iteratively refine the input features, allowing the network to adaptively weigh different feature channels and spectral positions during learning. For one-dimensional sequential data such as Raman spectra, this mechanism enhances the characteristic peaks associated with material composition while attenuating the background noise and redundant spectral information that may impair discrimination. This approach is chemically meaningful in gasoline adulteration analysis, because informative spectral variations are typically localized within specific wavenumber regions corresponding to molecular vibrational modes rather than being uniformly distributed across the entire spectrum.

Let the input one-dimensional Raman spectral feature map be denoted as F ₁ ∈ R ^{C × L} , where C is the number of channels and L is the spectral length. In the channel attention branch, global average pooling and global max pooling are applied along the spectral dimension to produce two channel descriptors: F _avg ∈ R ^{C ×1} and F _max . These descriptors capture the mean activation and strongest response of each channel across the spectrum. The two descriptors are then passed through a shared multilayer perceptron (MLP) to perform nonlinear transformation. The outputs are combined via element-wise summation and subsequently activated by a sigmoid function to generate channel attention weights.

$${\mathit{M}}_{\mathit{C}}({\mathit{F}}_{\mathbf{1}})=\mathit{\sigma }(\mathit{Net}({\mathit{F}}_{\mathit{avg}}^{\mathit{C}})+\mathit{Net}({\mathit{F}}_{\mathit{max}}^{\mathit{C}}))$$ 1

where σ ( x ) denotes the sigmoid activation function. The resulting channel attention weights were then applied to the original features via element-wise multiplication to produce a channel-refined feature representation.

$${\mathit{F}}_{\mathbf{1}}^{\prime }={\mathit{M}}_{\mathit{C}}({\mathit{F}}_{\mathbf{1}})\otimes {\mathit{F}}_{\mathbf{1}}$$ 2

In Raman spectral analysis, the channel attention mechanism adaptively learns the relative importance of feature channels for fuel classification and adulteration concentration prediction, thereby strengthening the representations associated with variations in chemical composition. Building on this, the CBAM incorporates a spatial attention mechanism to capture the significance of different wavelength positions within the spectral sequence. Specifically, average pooling and max pooling are performed along the channel dimension on the feature map F ₁ , producing two one-dimensional spectral descriptors, F _avg ∈ R ^{1× L} and F _max ∈ R ^{1× L} . These descriptors were concatenated along the channel dimension and subsequently processed by a one-dimensional convolution layer, followed by a sigmoid activation function to generate spatial attention weights.

$${\mathit{M}}_{\mathit{S}}({\mathit{F}}_{\mathbf{1}}^{\prime })=\mathit{\sigma }(\mathit{Conv}([{\mathit{F}}_{\mathit{avg}}^{\mathit{S}};{\mathit{F}}_{\mathit{max}}^{\mathit{S}}]))$$ 3

where [ a ; b ] denotes the concatenation operation. The resulting spatial attention weights are then applied to the input features via element-wise multiplication, yielding feature representations modulated by both the channel and spatial attention.

From an application standpoint, the spatial attention mechanism emphasizes discriminative peak locations within the Raman spectral sequences, enabling the model to focus on regions closely associated with compositional variations. In contrast, the channel attention mechanism selectively enhances informative feature channels at the representation level. Through their complementary interaction, the CBAM module improves the representational capacity for complex Raman spectra without introducing substantial computational overhead, thereby providing a robust feature basis for subsequent multi-scale fusion and integrated qualitative–quantitative modeling. Furthermore, the attention-enhanced representations corresponded to chemically meaningful Raman bands, such as oxygen-related vibrational regions in alcohol-adulterated samples and hydrocarbon deformation regions in kerosene- and diesel-adulterated fuels.

2.3.2. CBAM_FPN Multi-Scale Feature Extraction Module

The Raman spectral signals encode chemical information across multiple hierarchical and spectral scales. At finer scales, the peak shapes correspond to specific functional groups or molecular vibrations, whereas broader spectral patterns reflect overall composition and structural characteristics. Single-scale feature representations are inadequate for simultaneously capturing detailed local features and global spectral context, thereby limiting the modeling of complex spectral patterns. Consequently, the incorporation of multiscale feature fusion strategies for hierarchical spectral modeling has emerged as an effective approach for improving both qualitative and quantitative prediction performance.

Motivated by the considerations above, FPN augmented with the CBAM attention mechanism was incorporated into the backbone architecture to enable hierarchical feature extraction and effective cross-scale fusion, as illustrated in Figure c. For Raman spectra in gasoline adulteration analysis, this multi-scale design is consistent with the underlying chemical characteristics of the signal: narrow spectral bands correspond to specific molecular vibrational features, whereas broader patterns reflect the overall hydrocarbon composition and distribution of adulterants within the fuel system.

The multi-scale feature fusion framework is illustrated in Figure c, comprising both bottom-up and top-down pathways. Lower-level features preserve the detailed spectral peak information, whereas higher-level features encode more abstract global representations. Following the initial convolutional extraction of the input Raman spectral signal, the feature map at the i -th layer is defined as

$${\mathit{C}}_{\mathit{i}}\in {\mathit{R}}^{{\mathit{C}}_{\mathit{i}}\times {\mathit{L}}_{\mathit{i}}}\mathit{i}=\mathbf{1},\mathbf{2},\mathbf{3},\mathbf{4}$$ 4

Here, C _i denotes the number of channels in the i -th layer, and L _i represents the corresponding feature length. As the network depth increases, the number of channels typically expands, whereas the feature length decreases because of the downsampling operations. Consequently, feature representations evolve from localized spectral peak information to more abstract global features with a richer semantic content. During the bottom-up feature extraction process, the representations at each layer are updated through one-dimensional convolution and further refined via adaptive modulation by the CBAM module, as formulated in the following equation:

$${\mathit{C}}_{\mathit{i}+\mathbf{1}}={\mathit{f}}_{\mathit{CBAM}}({\mathit{f}}_{\mathit{conv}}({\mathit{C}}_{\mathit{i}}))$$ 5

To enable effective cross-scale fusion, the Feature Pyramid Network employs one-dimensional convolutions to standardize channel dimensions across feature levels. Empirical evaluation with different channel settings indicates that a unified dimension of C _f = 128 achieves optimal performance; therefore, the feature transformation is defined as follows:

$${\mathit{L}}_{\mathit{i}}=\mathit{Con}{\mathit{v}}_{\mathbf{1}\times \mathbf{1}}({\mathit{C}}_{\mathit{i}})\mathit{i}=\mathbf{1},\mathbf{2},\mathbf{3},\mathbf{4}$$ 6

After channel standardization, the feature maps can be expressed as L ₁, L ₂, L ₃, L ₄ ∈ R ^{B×128×(L ₀/2 i–1)}, where L ₀ denotes the length of the input Raman spectrum.

Building on this, the Feature Pyramid Network performs hierarchical multi-scale fusion through a top-down pathway. Specifically, higher-level features were upsampled and combined with the corresponding lateral features at the same resolution through element-wise addition, followed by convolutional refinement.

$${\mathit{P}}_{\mathbf{4}}=\mathit{Smoot}{\mathit{h}}_{\mathbf{4}}({\mathit{L}}_{\mathbf{4}})$$ 7

$${\mathit{P}}_{\mathbf{3}}=\mathit{Smoot}{\mathit{h}}_{\mathbf{3}}({\mathit{L}}_{\mathbf{3}}+\mathit{Upsample}({\mathit{P}}_{\mathbf{4}}))$$ 8

$${\mathit{P}}_{\mathbf{2}}=\mathit{Smoot}{\mathit{h}}_{\mathbf{2}}({\mathit{L}}_{\mathbf{2}}+\mathit{Upsample}({\mathit{P}}_{\mathbf{3}}))$$ 9

$${\mathit{P}}_{\mathbf{1}}=\mathit{Smoot}{\mathit{h}}_{\mathbf{1}}({\mathit{L}}_{\mathbf{1}}+\mathit{Upsample}({\mathit{P}}_{\mathbf{2}}))$$ 10

The resulting feature maps satisfied P _i ∈ R ^{B ×128×( L ₀ /2 i –1 )}. During feature fusion, small-kernel convolutions are applied to the aggregated features to perform localized integration, suppress aliasing effects, and reduce the interpolation artifacts introduced by upsampling. This operation also facilitates the integration of local channel information, yielding smoother and more locally consistent representations at each scale, which, in turn, supports stable processing in subsequent modules.

At the pyramid output stage, feature maps from all scales are upsampled to a common spectral length L ₀ and concatenated along the channel dimension to construct the final multi-scale fused representation, as defined below:

$${\mathit{F}}_{\mathit{FPN}}=\mathit{Concat}({\mathit{P}}_{\mathbf{1}},\mathit{Up}({\mathit{P}}_{\mathbf{2}}),\mathit{Up}({\mathit{P}}_{\mathbf{3}}),\mathit{Up}({\mathit{P}}_{\mathbf{4}}))$$ 11

The output satisfies F _FPN ∈ R ^{B ×4 C _f × L ₀} .

This architecture allows the model to retain both chemically relevant local spectral features and global profile variations, which are critical for differentiating oxygenated adulterants from hydrocarbon-based counterparts and for accurately estimating their concentration levels.

2.4. Model Training and Evaluation Methods

All experiments were conducted on the Google Colab cloud platform by using Python 3.12 and the PyTorch 2.9.0 deep learning framework. Model training and inference were executed on an NVIDIA Tesla T4 GPU with CUDA 12.6 acceleration, which was automatically configured in the Colab environment.

Training the MSAF-CNN model for gasoline adulteration identification requires preprocessing of the acquired Raman spectral data. A preprocessing pipeline was sequentially applied. First, the raw spectra were smoothed using a Savitzky–Golay filter (window length = 15, polynomial order = 3). The purpose of setting these parameters is to maintain the characteristic shape of the Raman peak while effectively suppressing high-frequency noise. Next, baseline correction was performed on the smoothed spectra using the adaptive iteratively reweighted penalized least squares (airPLS) algorithm (penalty parameter λ = 100, difference order p = 1, maximum iterations = 15). This configuration is widely used in Raman chemometrics and is the best method to eliminate the broad fluorescence background in our dataset. Corrected spectra were obtained by subtracting the estimated baseline from the smoothed signals. Finally, all the spectra were normalized to the range [0, 1] using min–max scaling. To ensure experimental transparency and reproducibility, all traditional chemometric baselines (VIP-PLS, CARS, Random Frog, SVM, KNN, Random Forest, PLSR, and SVR) adhered to this identical preprocessing workflow. Hyperparameters for comparison models were fixed to traceable settings: VIP-PLS utilized 10 latent variables; CARS was configured with 50 iterations (0.9 to 0.2 retention ratio); and Random Frog was executed for 5000 iterations. Models such as SVM (RBF kernel, C = 1.0, γ = scale) and PLSR ( n _components = 17) were evaluated using 5-fold stratified cross-validation ( random _state = 42). Model optimization was performed using the Adam optimizer with a ReduceLROnPlateau scheduler (factor = 0.5, patience = 5) to facilitate convergence. The batch size was set to 32, and each fold was trained for 100 epochs. To ensure reproducibility, random seeds were fixed at 42.

During model development, the MSAF-CNN architecture was implemented using the PyTorch framework. The model parameters were optimized over 100 training epochs by using the Adam optimizer. The initial learning rate was set at 0.001, and an adaptive scheduling strategy based on the validation loss was employed. Specifically, the learning rate was automatically reduced by a predefined factor when the validation loss did not decrease over several consecutive epochs. This strategy stabilizes training in later stages by reducing update magnitudes and facilitates more precise convergence within the optimal region of the loss landscape.

The proposed model adopts a multi-task learning strategy in which the classification and regression objectives are jointly optimized via a weighted loss formulation. During training, the classification branch employs cross-entropy loss to quantify the discrepancy between the predicted fuel categories and ground-truth labels, whereas the regression branch utilizes the mean squared error (MSE) loss to model the adulterant concentration. This joint optimization constrains the prediction errors in both tasks. The overall loss function is defined as follows:

$$L_{cls}=-\frac{1}{B}\sum _{i=1}^B\sum _{c=1}^C{y}_{i,c}\log ({ŷ}_{i,c})$$ 12

$$L_{reg}=\frac{1}{B}\sum _{i=1}^B{({\mathrm{r̂}}_i-r_i)}^2$$ 13

where B denotes the batch size, and C represents the number of classes. y _{i , c} indicates the ground-truth label of sample i for class c , and y ^ _{i , c} denotes the predicted probability for that class. Furthermore, r _i represents the true adulteration concentration of sample i , and r ^ _i is the corresponding predicted value.

During training, the regression task was more challenging than the binary classification task and converged more readily. To balance their contributions in joint optimization, the classification and regression loss weights were set at 0.5 and 2.0, respectively. Prior to loss calculation, regression targets were Z-score-normalized within each fold to align the numerical scales of both tasks. These specific weights were determined through systematic sensitivity analysis, evaluating multiple candidates on the validation set to achieve an optimal balance between the classification accuracy and regression RMSE. By increasing the contribution of the regression term to the overall loss function, the model emphasizes accurate estimation of adulteration concentration while maintaining stable classification performance, thereby ensuring that back-propagation gradients effectively drive learning for quantitative analysis without being dominated by the rapidly converging classification branch. This configuration was empirically verified to facilitate the stable and synchronized convergence for both branches. During optimization, the network parameters were iteratively updated via back-propagation, and the validation loss was computed after each epoch using a held-out validation set. Continuous monitoring of the validation loss enables an assessment of model convergence and generalization performance. The model parameters corresponding to the minimum validation loss were retained as the optimal solutions. After completing the training for each fold, the selected model was evaluated and its performance on both the classification and regression tasks was comprehensively analyzed.

After completing the cross-validation procedure, the model results were aggregated to compute the overall performance metrics and visualization techniques were employed to analyze the model. This enables an intuitive assessment of the predictive performance and highlights the key features emphasized by the model.

To comprehensively assess the model performance on both classification and regression tasks, accuracy (ACC), misclassification rate (MR), and macro-averaged F1 score (MF1) were adopted as evaluation metrics for classification. The formulations are given in eqs (14), (15), and (16). Higher values (closer to 1) indicate a stronger discriminative capability and improved classification performance. For regression, the coefficient of determination ( R ² ) and root mean square error (RMSE) were employed as the evaluation metrics. Their definitions are given as follows:

$$R^2=1-\frac{\sum _{i=1}^N{(y_i-{ŷ}_i)}^2}{\sum _{i=1}^N{(y_i-\mathrm{y̅})}^2}$$ 14

$$RMSE=\sqrt{\frac{1}{N}\sum _{i=1}^N{({ŷ}_i-y_i)}^2}$$ 15

Here, N denotes the number of samples, y _i represents the true adulteration concentration of the i -th sample, y ^ _i denotes the corresponding predicted value, and ^y is the mean of all true concentrations. A higher coefficient of determination ( R ² ), approaching 1, indicates stronger explanatory power and a better regression fit. In contrast, a lower root mean square error (RMSE) reflects smaller prediction errors and greater stability when estimating the adulteration concentration.

3. Results and Discussion

3.1. Raman Spectral Data Analysis

Figure presents a comparison of the average Raman spectra for different adulterants across an adulteration range of 5–30%. Prior to analysis, all spectra were subjected to smoothing, baseline correction, and normalization; consequently, the observed differences primarily reflected variations in the relative spectral responses associated with changes in the adulteration level. Across the four adulteration scenarios, the overall spectral profiles at different ratios remained largely consistent with those of the base fuel and the positions of the principal characteristic peaks showed minimal variation. This suggests that, within the studied concentration range, the Raman response is predominantly governed by the base fuel composition. Rather than introducing new vibrational modes, adulteration primarily alters the relative contributions of the existing spectral components. This observation aligns with the physicochemical nature of fuel adulteration, which involves compositional changes, rather than fundamental structural transformations.

2.

Average spectra of GE (a), GK (b), GD (c), and GM (d) at different ratios and spectral comparison of different adulteration types (e).

Although the overall spectral profiles of the adulterated samples remain highly similar, subtle yet systematic variations can be observed within the specific wavenumber regions. Taking ethanol adulteration (Figure a) as an example, the characteristic bands at approximately 400, 500, 893, 1052, 1090, and 1447 cm^–1 exhibit a clear monotonic increase in intensity with an increasing adulteration ratio, accompanied by distinguishable changes in peak shape. In contrast, the spectral responses associated with adulteration with methanol, kerosene, and diesel were more complex. For methanol (Figure d), noticeable fluctuations occurred around 1030, 1380, and 1447 cm^–1; however, these variations did not follow a simple linear relationship with the concentration. Similarly, kerosene and diesel (Figure b and Figure c) primarily show changes near 1380 and 1447 cm^–1 but also lack a clear concentration-dependent linear trend.

The sensitive Raman regions can be attributed to specific vibrational modes that reflect the micro-level chemical changes induced by adulterants. The 600–800 cm^–1 region corresponds to ring skeletal deformation and low-frequency fingerprint vibrations, which are indicative of distinct chemical signatures for different adulterant species. The region around 1000 cm^–1 arises from ring breathing and C–C vibrations, showing high sensitivity to compositional differences in the aromatic (alkane) ratio. The bands at 1052–1090 cm^–1 correspond to C–O stretching and are significantly enhanced in oxygenated-adulterated samples, such as those containing ethanol and methanol. Furthermore, the 1150–1300 cm^–1 region, encompassing aromatic C–H bending alongside additional C–O stretching modes, provides a critical window for identifying oxygen-containing impurities. The peak near 1447 cm^–1 is associated with CH₂/CH₃ deformation and serves as a key indicator of aliphatic hydrocarbon composition changes. Finally, the spectral feature at 1550–1700 cm^–1 is dominated by aromatic CC skeletal stretching (centered near 1600 cm^–1), which serves as a direct chemical probe for monitoring variations in aromatic content caused by potential adulteration. Kerosene and diesel, which are composed mainly of hydrocarbons, exhibit broader and more complex spectral features. These variations are related to the molecular structure.

Figure e compares the average Raman spectra of the different adulteration systems within a concentration range of 5–30%. After smoothing, baseline correction, and normalization, the observed variations mainly reflected the relative spectral changes associated with adulteration levels. Across all systems, the overall spectral profiles and peak positions remain largely unchanged, indicating that the Raman response is dominated by the base gasoline matrix, whereas adulteration primarily modifies the relative contributions of the existing vibrational components.

Within this concentration range, spectral changes are primarily reflected in the gradual variations in the peak intensity and shape across multiple regions rather than the appearance of new characteristic peaks. These variations were distributed over different spectral bands, indicating that adulteration cannot be reliably characterized by a single wavenumber. Therefore, both the local spectral features and the global spectral patterns should be considered in subsequent qualitative and quantitative analyses.

3.2. Experimental Results and Comparative Analysis

3.2.1. Five-Fold Cross-Validation Results

The final results of the proposed model after training and validation are summarized in Table . The table reports the classification and regression metrics for each fold in the five-fold cross-validation, along with their mean and standard deviation. In this study, cross-validation was performed at the spectral level, with all of the spectra randomly divided into five subsets. As shown in Table , the model exhibited consistently strong classification performance, achieving an average accuracy (ACC) of 99.92%, with perfect accuracy (1.0) observed several times. This indicates a high capability of distinguishing among different fuel categories. In addition, the macroaverage recall (MR) and macroaverage F1 score (MF1) remained stable across folds with mean values of 0.9992 and low standard deviations, demonstrating minimal performance variation and strong generalization ability.

1. Five-Fold Cross-Validation Results of MSAF-CNN for All Evaluation Metrics.

	MR	MF1	ACC	R2	RMSE
fold1	0.9959	0.9959	0.9958	0.990574	0.008669
fold2	1	1	1	0.986343	0.009685
fold3	1	1	1	0.991956	0.007738
fold4	1	1	1	0.988415	0.008971
fold5	1	1	1	0.989859	0.008539
Mean and standard deviation	0.9992±0.0018	0.9992±0.0018	0.9992±0.0019	0.9894±0.0021	0.0087±0.0007

In regression tasks, the model also demonstrated a stable predictive performance. Under five-fold cross-validation, the average coefficient of determination (R ²) reached 0.9894, with consistently high values across folds and a maximum of 0.9919, indicating effective modeling of the relationship between the adulteration concentration and Raman spectral features. The average root-mean-square error (RMSE) was 0.0087 with a low standard deviation, suggesting that prediction errors were consistently small across different folds within the data set.

Under the current cross-validation setting, the model exhibited a stable performance. As shown by the baseline comparisons in subsequent sections, differences between models primarily lie in their average performance and variability under the identical experimental conditions. Therefore, these results should be interpreted as reflecting the performance of the present dataset rather than demonstrating absolute superiority.

3.2.2. Performance Comparison of Different Backbone Network Models

A series of comparative experiments were conducted to comprehensively evaluate the proposed model. Under the multitask learning framework, model performance was first compared across different backbone architectures using the same training strategy. Subsequently, several single-task models were introduced to examine the differences between multitask and single-task learning in both the classification and regression tasks. Finally, the key architectural parameters were analyzed by evaluating the model performance under different configurations to validate the effectiveness of the design and the parameter settings. As the classification metrics MR, MF1, and ACC showed minimal differences, ACC was selected as the primary metric for subsequent evaluation. In the comparison of deep learning models, the same data set, data partitioning strategy, multitask framework, and evaluation criteria were used, with only the backbone network varied. This setup minimizes the influence of confounding factors and ensures that the comparison focuses on the feature extraction capabilities of the different backbone architectures.

Within the multitask learning framework, backbone architecture plays a critical role in spectral feature extraction. Therefore, a comparative analysis of different backbones was conducted under the unified experimental conditions. To ensure a fair comparison, the data set, training protocol, and multitask framework were kept constant, while only the backbone was varied to construct MDTL models based on ResNet18 and LSTM. ResNet18 and LSTM were selected as representative baselines, reflecting two widely used paradigms in spectral analysis: convolution-based feature extraction and sequence-based modeling. This design enables a direct comparison of the proposed MSAF-CNN model to architectures that differ substantially in their feature-learning mechanisms.

The results in Table reveal clear performance differences between the three backbone architectures. The proposed MSAF-CNN model has higher classification accuracy, a larger coefficient of determination (R ²), and a lower prediction error along with improved stability across folds. This indicates that integrating attention mechanisms with multi-scale feature fusion enhances the representation of key spectral information, thereby improving both classification and regression performance. In comparison, ResNet18_MDTL performs well in classification but shows slightly inferior regression accuracy, whereas LSTM_MDTL exhibits larger regression errors, suggesting that sequence-based modeling alone is insufficient to capture complex spectral characteristics. In addition, the ablation study included a standard multitask CNN without the proposed feature enhancement modules, providing a direct baseline for assessing the contributions of CBAM and FPN. Consequently, the comparisons cover both representative deep backbone architectures and conventional multitask CNN baselines under consistent experimental settings. Overall, the CBAM-FPN is better suited as a backbone for multitask learning in spectral analysis.

2. Performance Comparison of Different Backbone Network Models.

Backbone models	ACC	R2	RMSE
MSAF-CNN	0.9992±0.0019	0.9894±0.0021	0.0087±0.0007
ResNet18_MDTL	0.9967±0.0035	0.9858±0.0089	0.0098±0.0025
LSTM_MDTL	0.9967±0.0055	0.9805±0.0028	0.0119±0.0009

3.2.3. Comparative Results of the Ablation Study

The ablation results in Table show that the model incorporating both CBAM and FPN achieved the best performance across all metrics. The full model outperformed the variants in terms of accuracy, coefficient of determination, and root-mean-square error (RMSE), indicating a strong complementary effect between attention mechanisms and multi-scale feature fusion. When the CBAM module is removed and only the FPN is retained, the performance degrades, particularly in R ² and RMSE. This suggests that CBAM plays a critical role in highlighting informative features, suppressing redundant information, and improving regression fitting.

3. Comparative Results of the Ablation Study.

Methods	CBAM	FPN	ACC	R2	RMSE
MSAF-CNN	√	√	0.9992±0.0019	0.9889±0.0025	0.0089±0.0008
FPN_MDTL	×	√	0.9983±0.0023	0.9847±0.0028	0.0105±0.0007
CNN_CBAM_MDTL	√	×	0.9867±0.0136	0.9810±0.0088	0.0115±0.0023
CNN_MDTL	×	×	0.8674±0.0478	0.9544±0.0100	0.0181±0.0023

A comparison of the two CNN-based models shows that incorporating CBAM leads to clear performance gains, including higher accuracy and coefficient of determination and lower prediction error, highlighting its effectiveness in feature weighting and information selection. In contrast, the baseline model without CBAM and FPN exhibited the poorest overall performance, indicating the limitations of a single convolutional structure in capturing the complex spectral features. Overall, the ablation results demonstrated that both CBAM and FPN contributed to an improved model stability, generalization, and predictive performance.

3.3. Qualitative Identification of Adulterated Gasoline

In traditional machine learning, the classification performance largely depends on the feature representation; therefore, single-task classifiers are systematically evaluated in combination with multiple-feature selection methods. Using the same dataset and cross-validation protocol, three classifiers (SVM, KNN, and Random Forest) were trained and tested on both the original features and those selected by the VIP, CARS, and Random Frog to ensure comparability. To maintain fairness, all models were implemented under identical preprocessing pipelines and validation settings. Hyperparameters were chosen based on commonly used configurations in spectral analysis and tuned within a limited range to achieve a stable performance, avoiding extensive optimization that could bias the comparison.

As shown in Table , the feature selection has a pronounced effect on the classification performance. Overall, CARS achieves the best or near-best results across the three classifiers, particularly for SVM and KNN, where the accuracy and macroaverage metrics (MR, MF1) are improved compared with the original features and exhibit lower variance, indicating enhanced accuracy and stability. In contrast, the VIP method generally fails to improve performance and, in some cases, leads to slight degradation, suggesting its limited effectiveness in preserving discriminative information under the current data distribution. Random Frog shows intermediate performance, outperforming the original features but with a slightly lower stability than CARS. For random forest, which inherently incorporates feature selection and nonlinear modeling, the performance differences among feature selection methods are relatively small; however, CARS still provides the best overall results. These findings indicate that appropriate feature selection can improve the performance of single-task classifiers and that the consistent advantage of CARS across models confirms its effectiveness at extracting informative spectral features. By combining unified preprocessing, controlled hyperparameter tuning, and multiple feature selection strategies, this framework provides a balanced evaluation of traditional machine learning methods and supports a more transparent comparison to the proposed deep-learning approach.

4. Performance Comparison of Single-Task Classification Models.

Methods	Feature selection	ACC	MR	MF1
SVM	Original	0.9650±0.0174	0.9650±0.0174	0.9645±0.0179
	VIP	0.9525±0.0200	0.9525±0.0200	0.9520±0.0205
	CARS	0.9733±0.0114	0.9733±0.0114	0.9734±0.0113
	RandomFrog	0.9616±0.0205	0.9617±0.0205	0.9611±0.0211
KNN	Original	0.9525±0.0077	0.9525±0.0077	0.9525±0.0077
	VIP	0.9500±0.0160	0.9500±0.0160	0.9501±0.0158
	CARS	0.9716±0.0061	0.9717±0.0061	0.9717±0.0061
	RandomFrog	0.9541±0.0075	0.9542±0.0075	0.9542±0.0074
Random Forest	Original	0.9808±0.0108	0.9808±0.0107	0.9808±0.0108
	VIP	0.9700±0.0097	0.9700±0.0096	0.9701±0.0096
	CARS	0.9783±0.0133	0.9783±0.0133	0.9784±0.0133
	RandomFrog	0.9775±0.0148	0.9775±0.0148	0.9774±0.0149

Figure a-d shows the confusion matrices for classifying the four fuel types (GD, GE, GK, and GM) using four feature-processing approaches: CARS_KNN, Original_KNN, VIP_KNN, and MSAF-CNN (full data set). Overall, the GE class exhibited the highest separability and stability across all settings, with nearly no misclassification, indicating a distinct and consistent Raman spectral fingerprint. In contrast, most errors occurred between GD and GK, suggesting similarities in composition or adulteration characteristics that led to overlapping spectral features and reduced class separability.

3.

Confusion matrix classification results of CARS_KNN (a), Original_KNN (b), VIP_KNN (c), and MSAF-CNN (d).

The proposed MSAF-CNN method achieved high classification performance by effectively integrating features across the entire dataset and optimizing the modeling process. It attained an overall accuracy of 99.75% with very few misclassifications, demonstrating its ability to preserve informative spectral features and improve the accuracy and reliability of the fuel adulteration identification.

3.4. Quantitative Identification of Adulterated Gasoline

In single-task regression, performance depends on both the regression algorithm and the feature selection strategy; therefore, different models and feature selection methods were compared under consistent experimental conditions. Using the same dataset and cross-validation protocol, two representative regressors, PLSR and SVR, were evaluated using four feature sets (original, VIP, CARS, and random frog) to ensure comparability. The regression target was the adulteration rate for all samples across the different systems. A single regression head predicts continuous concentration values irrespective of the adulterant type, whereas the adulterant category is handled by a classification branch within a multitask framework.

The model performance showed a clear sensitivity to the choice of the feature selection method. For PLSR, relatively high R ² and low RMSE are achieved using the original features, whereas feature selection does not improve performance. Notably, CARS leads to a marked decline, suggesting that aggressive feature reduction may disrupt the model’s underlying linear relationships. In contrast, SVR benefits more from feature selection, particularly with VIP and random frogs, which improve R ² while maintaining a low error, with VIP yielding the best overall performance. This suggests that nonlinear models are more responsive to informative feature subsets and can benefit from an appropriate feature selection. Overall, SVR combined with suitable feature selection methods exhibits greater stability, whereas PLSR depends more on feature completeness, highlighting the interaction between the model type and feature selection strategy.

The quantitative prediction results for the gasoline adulteration concentration are presented in Figure . In the scatter plots, the diagonal dashed line represents the ideal regression (predicted = true), whereas the residual plots illustrate the error patterns and model stability across the concentration ranges. All three models capture the relationship between the adulteration ratio and Raman spectral response; however, differences in the fitting tightness and residual distributions indicate varying abilities to handle high-dimensional redundancy, peak overlap, and nonlinear effects. Notably, samples with the same nominal concentration may belong to different chemical systems owing to varying adulterant types, meaning that concentration-related spectral variations are not uniform across the categories. In the proposed framework, this issue is addressed through the joint optimization of the classification and regression. The classification branch constrains the shared feature space to retain adulterant-specific information, whereas the regression branch learns a continuous mapping from these representations to concentration values.

4.

Quantitative results and residual scatter plots of CARS_PLSR (a, d), Original_PLSR (b, e), and MSAF-CNN (c, f).

As shown in Figures a and d, the CARS_PLSR model exhibited clear underfitting and error amplification, with R ² = 0.8488 ± 0.0161and RMSE = 0.033 ± 0.002. The predicted points deviated substantially from the ideal regression line, particularly at medium-to-high adulteration levels, and the residuals were widely dispersed. This suggests that CARS-based feature selection may weaken concentration-related information and disrupt spectral continuity, thereby reducing the model’s ability to capture concentration gradients. In contrast, the Original_PLSR model (Figure b,e) showed improved fitting when the full spectral information was retained, achieving R ² = 0.9314 ± 0.0048 and RMSE = 0.022 ± 0.001. The predicted values aligned more closely with the measured values; however, some residual dispersion and fluctuations persisted at certain concentration ranges.

Figures c and f correspond to fold 3 of the five-fold cross-validation for the MSAF-CNN model and demonstrate a strong predictive consistency with well-controlled errors. The predicted points closely followed the ideal regression line, indicating a stable fitting across different adulteration ratios. This performance can be attributed to the attention mechanism and multi-scale feature fusion, which emphasize informative spectral regions and capture relationships at different scales within a multitask framework. Because the regression head is shared across all adulterant types, the model cannot rely on class labels alone to estimate concentrations. Instead, it must learn concentration-sensitive features that are consistent within each class but distinguishable across different systems. This design reduces the potential confusion between classification and regression tasks and supports a more meaningful mapping between spectral variations and concentration.

A comparative analysis was performed to evaluate the quantitative prediction performances of CARS_PLSR, Original_PLSR, and MSAF-CNN. The MSAF-CNN model achieved the best results with a coefficient of determination of R ² = 0.9886 ± 0.0046 and an RMSE of 0.009 ± 0.002. The residuals showed low variability and strong agreement between the predicted and measured values, indicating stable and accurate performance in adulteration concentration prediction.

In feature pyramid networks, the number of mapping channels directly influences feature representation and information capacity at each scale. Too few channels may limit representational ability, whereas too many channels can introduce redundancy, increase model complexity, and degrade the training stability and generalization. Therefore, selecting an appropriate channel dimension is essential for balancing performance and computational efficiency. To investigate this effect, comparative experiments were conducted with different channel configurations while keeping the network architecture and training strategy fixed. The channel sizes were set to 32, 64, 128, and 256, and the corresponding classification and regression metrics were evaluated using five-fold cross-validation.

When the number of mapping channels is small, the performance of both the classification and regression is limited. With 32 channels, the model achieved the lowest accuracy (ACC), indicating an insufficient capacity to represent multi-scale spectral features. Increasing the channel size to 64 improved the classification performance, but the regression performance declined. At 128 channels, the model achieved the best overall performance, with a high ACC, the highest mean R ² , the lowest RMSE, and low variability, indicating improved stability and generalization. In contrast, further increasing the channel size to 256 leads to performance degradation, particularly in regression, where R ² decreases and the RMSE increases. This suggests that excessive channel dimensionality introduces redundant features and complicates model optimization.

In summary, under the current experimental conditions, a channel dimension of 128 provides a suitable balance between the representational capacity and the model complexity.

3.5. Model Visualization Analysis

Visualization is performed in the feature space to intuitively assess the discriminative capability of the model at the feature-level. Principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) were used to project high-dimensional data into two dimensions, capturing the global structure and local neighborhood relationships, respectively. The visualizations were based on two types of inputs: (i) preprocessed Raman spectra (after smoothing, baseline correction, and normalization) and (ii) high-level representations extracted from the shared layers of the proposed multitask model. The former reflects the original spectral characteristics, whereas the latter corresponds to the feature embeddings learned through multi-scale fusion and attention mechanisms.

To compare the feature distributions before and after modeling, dimensionality-reduction visualization was applied to both preprocessed spectra and high-level shared features using PCA and t-SNE. Figure a-d presents the corresponding results. Before modeling, the original spectral features showed substantial interclass overlap in both spaces, with dispersed samples and unclear boundaries, indicating limited discriminative power in low-dimensional representations. After feature extraction, samples form more compact, well-separated clusters with reduced overlap and increased intra-class consistency, suggesting that the model suppresses irrelevant variations and enhances the discriminative structure. PCA preserves global variance through linear projection, whereas t-SNE emphasizes local neighborhood relationships and may distort the global distances. Therefore, these visualizations are used to illustrate relative changes in feature organization rather than to provide strict evidence of class separability. Overall, the clustering patterns indicate that the model improves the structure of feature representations in the latent space, which is consistent with the enhanced extraction of the discriminative spectral information.

5.

PCA (a, c) and t-SNE (b, d) feature visualization results.

4. Conclusion

This study proposes a multi-scale convolutional neural network, termed MSAF-CNN, which integrates the CBAM attention mechanism with a one-dimensional FPN to jointly perform adulterant classification and concentration prediction. Specifically, we constructed four binary mixture systemsgasoline–methanol (GM), gasoline–ethanol (GE), gasoline–kerosene (GK), and gasoline–diesel (GD)and 1200 Raman spectra covering adulteration ratios ranging from 5% to 30% were collected. As a result, the proposed model achieved a high classification accuracy (up to 99.92%) and strong regression performance, with R ² > 0.9886. Additionally, the effectiveness of the proposed framework was validated through comparative experiments with alternative backbones and traditional single-task models, while ablation studies further verified the contribution of each individual module. Moreover, visualization results demonstrated that the model effectively captured discriminative spectral features and consistently focused on informative wavelength regions. Compared with conventional approaches that separately address adulterant identification or concentration prediction, the proposed framework provides an integrated solution for simultaneous qualitative and quantitative gasoline adulteration analysis.

The MSAF-CNN model demonstrated high classification accuracy and stable quantitative prediction in complex gasoline adulteration systems. Specifically, the multitask framework jointly optimizes classification and regression by sharing feature representations, improving efficiency, and reducing redundant modeling. Consequently, from an analytical perspective, this work provides a unified framework for multiadulterant systems; from a modeling perspective, it highlights the effectiveness of combining attention mechanisms with multi-scale feature extraction, and from a chemical perspective, it suggests that model decisions are partially associated with Raman spectral variations linked to molecular composition. These characteristics distinguish the proposed method from previous Raman-spectroscopy-based studies that mainly focus on individual analytical tasks or conventional spectral modeling strategies. In particular, the identified sensitive regions correspond to oxygen-related vibrations and hydrocarbon skeletal or deformation modes, which is consistent with the compositional differences among the ethanol, methanol, kerosene, and diesel adulterants. Therefore, the proposed multiscale, multitask convolutional neural network demonstrates strong potential as an effective approach for gasoline quality assessment and adulteration analysis. However, this study primarily focuses on single-doped systems within the investigated concentration range, and the proposed framework exhibits a stable classification and quantitative performance in independent cross-validation at the formulation level. On the basis of these conclusions, future research will further evaluate the framework using mixed-doped systems and samples with doping concentrations below 5% to broaden its applicability. Furthermore, although the proposed MSAF-CNN demonstrates excellent performance under the controlled experimental conditions of this study, additional validation using more diverse gasoline samples from different geographical regions, refineries, production batches, and seasonal formulations will further enhance its generalization ability under real-world conditions. Therefore, future work will incorporate independent external datasets to comprehensively evaluate and further improve the model’s robustness, generalization ability, and practical applicability. Ultimately, in addition to these validation efforts, the proposed framework will be expanded and refined to support a wider range of applications, including multilabel spectral analysis and more complex fuel systems, thereby further enhancing its adaptability and application potential in practical fuel quality monitoring.

Data will be made available on request.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c05800.

• Description of the study area; MSAF-CNN model training flowchart; confusion matrices of different classification models; scatter plots and residual plots of predictions from different regression models; overall wavelength attention heatmap; comparison of FPN performance with different numbers of mapped channels; comparison of the effects of different activation functions; information on adulterated gasoline samples; regression single-task model performance comparison (PDF)

Yingling Bai: Writing – review & editing, Writing – original draft, Visualization, Project administration, Methodology, Investigation, Formal analysis, Conceptualization. Yongjiao Yuan: Writing – review & editing, Software, Methodology, Investigation, Formal analysis, Data curation. Hua Yang: Writing – review & editing, Visualization, Validation, Methodology, Investigation, Formal analysis. Yu Li: Writing – review & editing, Methodology, Investigation, Formal analysis. Yuding Zuo: Writing – review & editing, Methodology, Investigation, Formal analysis. Zhenhao Xiao: Writing – review & editing, Visualization, Validation, Methodology. Ming Kang: Writing – review & editing, Software, Methodology, Investigation, Formal analysis. Bin Tang: Writing – review & editing, Resources, Project administration, Investigation, Funding acquisition, Conceptualization.

The authors declare no competing financial interest.