Classification of Non-Frozen and Frozen–Thawed Pork with Adaptive Support Vector Machine and Electronic Nose ^†

Abstract

The quality of raw meat is important for community health as its freshness is crucial to preventing foodborne illnesses. In the United States, the related illness cases were 9.4 million cases with 55,961 hospital admissions and 1351 deaths annually. This underscores the urgent need for improved meat quality monitoring. This study aims to develop an electronic nose (E-nose) that can differentiate between frozen–thawed and fresh pork meat samples, thereby enhancing food safety. We designed the E-nose using MQ series gas sensor array with temperature and humidity sensors, and an Arduino Uno microcontroller. Sensors were calibrated for accurate data collection. An adaptive support vector machine (ASVM) was used for data classification. We evaluated the model’s accuracy using a confusion matrix. The ASVM model exhibited robust performance, achieving an accuracy of 88%. Its performance was evaluated with recall, F1 scores, and precision. To further enhance the model’s performance, future studies are mandated to integrate additional gas sensors, increase sample sizes, advance data preprocessing techniques, and explore different machine learning algorithms or ensemble methods.

1. Introduction

Raw meat quality substantially influences the well-being of the community, transcending beyond conventional considerations. The nature of raw meat, particularly its freshness or staleness, carries profound implications for public health. Statistics highlight a troubling proportion of raw meat in communities succumbing to premature spoilage and concealed bacteria, leading to severe consequences. In a study from the National Library of Medicine Database, active and passive surveillance data were collected for 31 pathogens from foodborne illnesses. In the United States, 9.4 million cases were reported with 55,961 hospitalizations and 1351 deaths annually [1]. This attests to the seriousness and accentuates the pressing need for attention to raw meat quality.

The adverse impacts of raw meat quality reverberate in community life. Beyond the individual health toll, economic strains emerge as medical expenses surge due to increased instances of foodborne illnesses. Moreover, the community’s overall productivity heightens absenteeism resulting from illnesses caused by compromised raw meat quality. These multifaceted consequences culminate in a pervasive sense of the community, eroding the trust in local food sources and disrupting everyday life. As the world becomes increasingly interconnected, the poor quality of raw meat not only raises local health concerns but also negatively impacts international trade dynamics.

Many types of E-nose have been developed with sensors and algorithms for quality control, and process monitoring [2]. The E-nose is used to detect Salmonella typhimurium contamination in fresh pork meat samples. Support vector machine regression (SVMR) in conjunction with other algorithms is used as a metaheuristic approach. Genetic algorithm-SVMR is efficient and accurate in identifying volatile compounds of the bacteria in pork meat. The limitations of the E-nose are addressed by creating a portable E-nose with elastic architecture and fault tolerance based on edge computing, ensemble learning, and sensor networks. Those are used to identify 15 volatile organic carbons (VOCs) with edge computing. The accuracy of classifying different VOCs reaches 81.1% with an additional enhancement of 20% with an individual classifier [3]. A different approach is used in Ref. [4] in the identification of fresh meat samples using an E-nose. A color sensor is used in conjunction with neural network pattern recognition. Utilizing three gas sensors and a color sensor, the E-nose successfully captures unique patterns corresponding to each freshness category. The accuracy of the meat freshness assessment reached 80% [5]. Different approaches are used to test the electronic nose in identifying chilled and frozen–thawed chicken meat in estimating their shelf life. In total, eight metal oxide semiconductor (MOS)-type sensors were used and pre-processed for training the fuzzy K-Nearest Neighbors Algorithm (F-KNN) for classification. The E-nose differentiates between fresh and frozen–thawed samples. The future trends of the electronic nose focus on the creation of better sensors. Algorithms are used for better detecting specific gases [6,7,8,9]. Different machine-learning techniques are also used, and the device setup and sensor prototypes are improved for higher accuracy [10,11].

SVM can be used for aroma-based classification. By processing aroma data on a laptop, an accuracy of at least 70% was achieved [12]. The E-nose is effective in detecting food spoilage while it is especially efficient in detecting spoilage in vegetables [13,14].

2. Method

Figure 1 shows the conceptual framework of the classification of pork meat based on the VOC profile using gas sensors. We developed an E-nose using Raspberry Pi 4 (Raspberry Pi Ltd.; Cambridge, UK) along with Arduino Uno (Arduino S.r.l., Monza, Italy), MQ sensor Array, and an LCD. The sample of non-frozen pork meat was collected from a local butcher, and the frozen meat was from a grocery. Their quality was pre-tested based on the VOC profile. The sample pork meat underwent VOC sensing. Afterward, the gathered data was preprocessed for feature extraction. Sensors were calibrated to ensure the accuracy of data collection. The adaptive support vector machine (ASVM) was used to analyze the data and extract the main features for classification. The confusion matrix was established to evaluate the system’s classification accuracy. For the output process, the system presented the VOC profile of the sample pork meat using the LCD module.

2.1. Hardware

Figure 2 shows the hardware components of the system. The proposed device is composed of a power supply, an Arduino UNO, a Raspberry Pi 4, an MQ series gas sensor array (MQ2, MQ8, MQ135, MQ138, and DHT 22), a fan, and an LCD. The Arduino UNO and the Raspberry Pi are used to collect and convert the data gathered from the gas sensor array. Once the data are transferred by the Raspberry Pi 4, then the data are pre-processed for training the machine learning model.

2.2. Software

Figure 3 shows the main flowchart of the systems training and deployment phase. The training was carried out on a personal computer, before data gathering and sensor calibration. The sensors were preheated for about 24 h and exposed to clean air. After sensor calibration, the MQ gas sensor array was used for 5 min to gather the data. The gathered data were preprocessed for feature extraction and consistency. Once the raw data were prepared, the machine learning algorithm was trained through cross-validation to assess its effectiveness. The accuracy of the machine learning algorithm was then evaluated.

Figure 4 shows the deployed model and its testing process. After training the model, the ASVM was loaded into the server. Once the pork meat sample was tested, the result was displayed for frozen or non-frozen meat samples. Before using the sensors for data gathering, the sensor was calibrated for accurate results. The MQ series gas sensor array and the DHT temperature sensor were preheated for 24 h. After preheating the sensors, they were exposed to clean air. The code was uploaded to the Arduino UNO for tests. Using the serial monitor from the Arduino IDE, the value of each sensor was monitored. If the value of the sensor did not stabilize, the load resistance from the code was implemented until the sensor values were stabilized.

The collected data underwent data cleaning. Inconsistencies, errors, or missing values in the data were corrected to ensure the data’s accuracy and reliability for further analysis. Noise filtering was used to identify and mitigate noise in the data and improve its quality for analysis and modeling. Data reduction was conducted to reduce the data volume but retain the integrity and quality to improve efficiency during analysis. Feature extraction was performed to select and transform the relevant information from the sensor data to extract important features. Lastly, data were validated for data integrity, consistency, and quality. Once the data were preprocessed, its classification was carried out. Then, the data were used for model training and classification. The classification process started to detect if meat samples were non-frozen.

2.3. Experimental Setup

The experimental setup of the E-nose is presented in Figure 5. Pork meat samples are classified as class 1: non-frozen pork meat sample and class 2: frozen–thawed pork meat sample. The non-pork meat was acquired from a market and the frozen pork meat sample was bought from the grocery stores. The frozen meat sample was stored at 3 °C in the refrigerator for 24 h. The sensor was calibrated before use. The sample was placed inside the container for 5 min under the gas sensor array. The fan was operated for 15 min to clean the container after removing the sample. The data were sent to Raspberry Pi 4 by the Arduino UNO.

Figure 6 shows the gas sensor array which comprises MQ2, MQ8, MQ135, and MQ138. The Arduino UNO and the Raspberry Pi operate as edge devices to collect and convert gathered data. Once the data were transferred to the Raspberry Pi 4, the data was preprocessed for training the machine learning model.

3. Results and Discussion

3.1. Data Acquisition

The E-nose collected data from the non-frozen and frozen–thawed samples using the MQ sensors and then trained later to distinguish the class of the meat based on the data collected.

Figure 7 shows data graphs for each sensor. Subfigures (a) shows the readings for MQ2 sensor readings, wherein it shows some separation, with “non-frozen” readings slightly lower than “Frozen Thawed.” (b) shows the MQ8 sensor readings, which significantly overlap the two classes. (c) shows the MQ138 sensor readings with moderate separation between the classes, with some overlap. (d) the MQ135 sensor readings show the most apparent separation between “Frozen Thawed” and “Non-Frozen” classes.

Table 1. Data of frozen–thawed and non-frozen–thawed meat.

Class	MQ-138	MQ-135	MQ-8	MQ-2
Frozen–thawed	120	128	75	137
Non-frozen	8	114	72	136

shows the results of the data of the sensors using the ASVM model. The testing data consisted of 10 samples for two classes of pork: frozen–thawed and non-frozen. For the frozen–thawed pork, the MQ gas sensors recorded values of 120 (MQ-138), 128 (MQ-135), 75 (MQ-8), and 137 (MQ-2). For the non-frozen pork, the MQ gas sensors recorded values of 8 (MQ-138), 114 (MQ-135), 72 (MQ-8), and 136 (MQ-2). These readings indicated differences in environmental and gas sensor data between the two classes, which are used to distinguish between frozen–thawed and non-frozen pork.

Figure 8 displays the output of the model. It accurately predicted frozen–thawed meat samples. The predictions showed the accuracy of the predicted values when compared with the actual status of the 10 samples.

3.2. Statistical Treatment

The confusion matrix in Figure 9 summarizes the performance of the classification model for distinguishing between frozen–thawed pork and non-frozen pork using 10 samples with 100 readings. The model successfully identified the frozen–thawed and non-frozen samples, achieving an accuracy of 88%. It had a precision of 89%. The recall was 88%, demonstrating that 86% of the real frozen–thawed pork samples were accurately identified by the model. The precision and recall-balancing F1 score was 88%. These metrics suggested that while the model was precise, it had room for improvement, as it misidentified several frozen–thawed samples. Other studies using SVM in electronic gas detection achieved a precision of 90–92% [15,16,17].

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(1)

Precision measures the accuracy of the positive predictions made by a model. It is the ratio of true positive predictions to the total number of positive predictions, both correct and incorrect. A high precision means that the model has a low false positive rate.

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(2)

Recall measures how well a model can identify actual positive cases. It is the ratio of true positive (TP) predictions to the total number of actual positives (true positives and false negatives (FNs)).

$$\mathrm{F}1=2\times {\displaystyle \frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$

(3)

The F1 score is the harmonic mean of precision and recall, balancing both metrics into a single number. A higher F1 score indicates better performance, especially in cases where both precision and recall are important.

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}}\%$$

(4)

The accuracy of the system is calculated using (1), wherein “predicted frozen–thawed pork” matches “actual frozen–thawed pork” as an example of TPs, which are accurate prediction of a particular class or category. False positives (FPs) are the outcomes of inaccurate class predictions. Correct forecasts of other classes are known as true negatives (TNs). Inaccurate rejections of a class are known as FNs. The model’s performance is assessed using accuracy.

$$\mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}\%$$

(5)

Sensitivity measures how effectively a model detects positive cases. It is computed by dividing the total number of FNs and TPs by TP. True positives are cases correctly identified as positive, while false negatives are positive cases missed by the model. This ratio in percentage shows the fraction of actual positives that were correctly recognized. High sensitivity means the model is good at finding positive cases.

4. Conclusions and Recommendations

The classification of non-frozen and frozen–thawed pork meat with an adaptive support vector machine and E-nose was investigated. The E-nose system effectively differentiated between frozen–thawed and non-frozen pork meat samples. The ASVM model achieved an accuracy of 88%, demonstrating that it was effectively integrated into the E-nose system. The system accurately collected data from pork meat samples and distinguished frozen–thawed and non-frozen meats. By integrating additional gas sensors, a more comprehensive VOC profile was obtained which increased accuracy. With advanced data preprocessing techniques, different machine learning algorithms, and software and hardware integration, the system enhances robustness and applicability.