The accurate identification of meat sources, particularly pork, is crucial for consumers due to concerns about allergic reactions and religious dietary restrictions. Current methods like DSC, ELISA, NMR, and PCR are often laboratory-bound, expensive, and require complex sample preparation. This necessitates the development of a portable, rapid, and reagent-free method for meat origin authentication. The volatility of meat products, determined by their volatile organic compound (VOC) composition, influences their aroma. These VOCs (organic acids, alcohols, esters, aldehydes, hydrocarbons, etc.) provide valuable information for identifying meat floss origin. Electronic noses (e-noses), comprising gas sensor arrays and artificial intelligence, offer a potential solution. E-noses have been used successfully in various VOC-based applications, including food authenticity assessment. Compared to other nondestructive technologies, e-noses offer advantages such as high sensitivity and accuracy, rapid results, simple sample preparation, economical procedure, low power consumption, and portability. However, enhancing pattern recognition models in e-noses is crucial for accurate data analysis. This study focuses on developing a compact, portable e-nose integrated with supervised machine learning models to classify different meat floss origins (beef, chicken, and pork). The accuracy was enhanced using a window time slicing method during feature extraction, and the performance was evaluated using LDA, QDA, k-NN, and RF models. FTIR spectroscopy and GC-MS were used to validate the e-nose results and elucidate the detected VOCs.

Several studies have explored methods for identifying pork in meat products, primarily using biochemical techniques. These techniques, while accurate, are often limited by laboratory requirements, costs, and complexity. The use of e-noses for food authenticity is gaining traction. Research demonstrates e-noses’ ability to discriminate between various meat types (chicken, pork, beef, etc.) based on their distinct VOC profiles. Different machine learning models, including LDA, QDA, k-NN, and RF, have been employed for data analysis in e-nose applications, each with its strengths and weaknesses. PCA is often used for dimensionality reduction and feature extraction in e-nose data analysis, but supervised learning methods are often necessary for accurate classification.

A portable e-nose system comprising eight metal-oxide semiconductor (MOS) gas sensors was developed. The sensors, listed in Table 1, exhibited cross-sensitivity, responding to multiple gases. The e-nose system included a sampling chamber, a sensing chamber, and a data acquisition (DAQ) system (Fig. 1). Meat floss samples (2g each) were placed in the sampling chamber, and their VOCs were carried to the sensing chamber by a flow of air. Measurements were taken over 260s (20s delay, 120s sampling, 120s purging). The raw sensor signals were normalized using baseline correction (Equation 1). A time window slicing method was applied to extract features (maximum, minimum, mean, median) from the sensor responses (Fig. 1). PCA was initially used for unsupervised clustering. Four supervised learning models (LDA, QDA, k-NN, RF) were used for classification. The dataset of 300 samples (100 per meat type) was split into training (75%) and testing (25%) sets. 10-fold cross-validation was used for model evaluation. The chemical composition of the samples was analyzed using FTIR spectroscopy and GC-MS. FTIR spectra were recorded from 4000 to 400 cm⁻¹ (Fig. 5). GC-MS was used to identify VOCs (Table 4).

The LDA model with five-window-extracted features achieved the highest accuracy (>99%) for both validation and testing data in discriminating beef, chicken, and pork flosses (Table 2). Other supervised learning models (QDA, k-NN, RF) were also evaluated, with LDA consistently outperforming the others (Table 3). PCA analysis showed that pork samples clustered separately from beef and chicken samples, although some overlap existed between beef and chicken (Fig. 2). FTIR analysis revealed similar absorption spectra for all three samples, indicating similar functional groups, which is consistent with the PCA results. GC-MS analysis identified various VOCs, including hydrocarbons, alcohols, ethers, and aldehydes (Table 4). Beef and chicken shared many volatile compounds, explaining the overlap in PCA and LDA analysis. Aldehydes, particularly dodecanal and 9-octadecanal, were dominant in pork samples. Increasing the number of windows in the time slicing method improved cluster separation, reducing overlap between beef and chicken and increasing the distance between clusters, as shown by Euclidean distance calculations (Fig. 4).

The high accuracy of the LDA model with five windows demonstrates the e-nose system’s potential for rapid and accurate meat floss origin authentication. The combination of the e-nose with supervised learning, particularly LDA, effectively classified the samples based on their distinct VOC profiles. The findings highlight the importance of optimizing feature extraction methods, such as the window time slicing method, for enhancing classification accuracy. The agreement between e-nose results and FTIR and GC-MS data validates the e-nose's ability to detect relevant volatile compounds. The distinct volatile compound profiles identified in this study can be used to develop databases for future e-nose applications. The e-nose’s portability and reagent-free operation offer significant advantages over traditional methods. The results indicate that e-noses are a promising alternative for food authenticity testing, especially for rapidly discriminating pork from non-pork meat products.

This study successfully developed a compact, portable e-nose system capable of accurately identifying the origin of meat floss. The LDA model, coupled with the five-window time slicing method, achieved exceptionally high accuracy. This system offers a rapid, reagent-free, and cost-effective alternative to traditional methods for food authenticity testing. Future research could explore the application of this system to a broader range of meat products and investigate other machine learning models for further performance optimization.

The study was conducted using meat floss samples from a single geographical location and limited range of animal breeds. The generalizability of the findings to other regions and breeds may need further investigation. The sensor cross-sensitivity, while typical of MOS sensors, might affect the interpretation of results. Further work is needed to confirm the model’s robustness under varying environmental conditions and with a wider range of meat floss samples and processing methods. The study does not examine the potential influence of food additives or other factors which may influence the VOC profile. The current sample size, while adequate for the study, could be increased for more robust statistical analysis.