key: cord-0028895-zx2d6753 authors: Amani, Mohammad Amin; Sarkodie, Samuel Asumadu title: Mitigating spread of contamination in meat supply chain management using deep learning date: 2022-03-23 journal: Sci Rep DOI: 10.1038/s41598-022-08993-5 sha: 013210c3a1c5ee3dcedc5d5ca57917a34e7ab63e doc_id: 28895 cord_uid: zx2d6753 Industry 4.0 recommends a paradigm shift from traditional manufacturing to automated industrial practices, especially in different parts of supply chain management. Besides, the Sustainable Development Goal (SDG) 12 underscores the urgency of ensuring a sustainable supply chain with novel technologies including Artificial Intelligence to decrease food loss, which has the potential of mitigating food waste. These new technologies can increase productivity, especially in perishable products of the supply chain by reducing expenses, increasing the accuracy of operations, accelerating processes, and decreasing the carbon footprint of food. Artificial intelligence techniques such as deep learning can be utilized in various sections of meat supply chain management––where highly perishable products like spoiled meat need to be separated from wholesome ones to prevent cross-contamination with food-borne pathogens. Therefore, to automate this process and prevent meat spoilage and/or improve meat shelf life which is crucial to consumer meat preferences and sustainable consumption, a classification model was trained by the DCNN and PSO algorithms with 100% accuracy, which discerns wholesome meat from spoiled ones. www.nature.com/scientificreports/ thermosphacta, Carnobacterium spp., Enterobacteriaceae, Lactococcus lactis, Lactobacillus spp., Leuconostoc spp., Pseudomonas spp, and Shewanella putrefaciens) can grow and spread to the meat by extrinsic factors, processes (transporting and packaging) and environmental conditions (such as changing humidity and temperature) 18, 19 . Therefore, controlling and monitoring meat production during the various stages of the supply chain to manage this risk is crucial to achieving food safety 20 . Thus, new intelligent technologies such as smart containers, artificial intelligence, or IoT methods are required to mitigate post-harvest losses, decrease lead times, and control the perishability dynamics in meat supply chain management 12 . Industry 4.0 proposes a paradigm shift from traditional manufacturing to automated industrial operations by utilizing novel technologies, including artificial intelligence 21 . The provision of food to meet demand due to the ever-increasing population has become a global issue 22 . However, traditional agricultural supply chains may fail to meet global demand unless they are optimized by embedding intelligent technologies in the production function 6 . Enterprises can address clients' new demands and supply challenges while maintaining expectations in efficiency such as inter alia, online-enabled transparency, easy access to a multitude of options, and constant changes in stock-keeping unit (SKU) portfolio. The efficiency of a supply chain is amplified by automating both physical tasks and planning 23 . Machine learning is a method widely implemented to find patterns and linear and non-linear relationships between different variables, and it has various subcategories such as Classification, Regression, or Clustering [24] [25] [26] , which can be employed to analyze and help make decisions 27 . Machine learning and deep learning, which are subcategories of artificial intelligence, facilitate the generation of actionable intelligence by processing gathered data to improve manufacturing productivity without substantially altering recommended resources 28 . The use of artificial intelligence in the management of perishable products in the supply chain has received much attention. Shahbazi and Byun 29 employed novel technologies, including blockchain, machine learning, and fuzzy logic to develop a better traceability system in the supply chain that addresses several factors of the perishable food supply chain, including evaporation, weight, warehouse transactions, or shipping time, which enhances the shelf life of perishable foods. Alfian, Syafrudin 30 proposed the radio frequency identification (RFID) technology for traceability of perishable foods, machine learning to detect the direction of passive RFID tags, and IoT to control temperature and humidity during storage and transportation, which this integration between these technologies enhance the efficiency of the traceability system. Barbon, Barbin 31 used machine learning algorithms to predict the quality of chicken based on near-infrared (NIR) spectra data by analyzing chicken. Deep learning techniques such as deep convolutional neural networks are utilized to automate food manufacturing and supply chain management tasks based on the industry 4.0 paradigm. Al-Sarayreh 32 proposed an integrated system of hyperspectral imaging and deep learning techniques to assess the quality of various food products such as meat. Zhang 33 presented a convolutional neural networks (CNN) model for vibrational spectral analysis, which measures specific chemical bonds of atoms and molecules. The estimated model achieved 99.01% accuracy (better performance than other theoretical techniques) on a meat dataset that consists of chicken, pork, and turkey. Al-Sarayreh 34 employed the CNN algorithm to build a model for detecting lamb, beef, and pork meat adulteration (i.e., adding another type of meat that has a lower price to the meat pack). The model classified meats with 94.4% accuracy. Liu 35 utilized a CNN model to detect and analyze the complex matrices of various foods such as meat products, aquatic products, cereal products, fruits, and vegetables. In contrast, existing literature that detects and mitigates the potential spread of contamination among perishable foods is limited. Here, due to the effect of bacteria-driven meat spoilage in spreading from rotten meat into healthy meats and leading to contamination, this study employs artificial intelligence based on industry 4.0 and SDG12 paradigms to automate the controlling and monitoring process of meat production by detecting spoiled meats at various stages of the meat supply chain management. This is useful for improving the meat shelf life, reducing harm and health-related risk to consumers, and mitigating food and economic waste. Therefore, a deep convolutional neural network is employed as a classifier to increase productivity and reduce costs, whereas the PSO algorithm tunes the classifier hyperparameters. The remaining sections of this paper are organized as follows-"Methodology " proposes a deep convolutional neural network model and a PSO algorithm utilized for hyperparameter tuning. The experimental results are presented in "Data acquisition", whereas the discussion and conclusion are presented in "Particle swarm optimization algorithm" and "Deep convolutional neural networks", respectively. This section describes a popular optimization algorithm, which is employed to adjust the hyperparameters in the deep learning model. Moreover, the deep convolutional neural network and its layers are presented herein. Data acquisition. The dataset utilized in this study was adopted from previous research 36 . This dataset contains two classes, wholesome and spoiled red meat samples collected from a supermarket in Turkey. The dataset has 1896 images in total, with 948 per class gathered via an IP camera with an image resolution of 1280 × 720. Particle swarm optimization algorithm. Particle Swarm Optimization (PSO) is a population-based stochastic optimization algorithm that is inspired by the bird's swarm social behavior. This algorithm is a member of the Swarm Intelligence (SI) that arises from research based on creatures living as a group. This group members have little or no insight but can operate complicated works by interacting with each other 37 . In this algorithm, a population of random particles is first determined, and each of the individuals is assigned a velocity and position. The historical behavior of each particle and its neighbors, while they move through the search area, adjusts the paces. The new position is updated by the accelerations of the next step and the current situation. Therefore, the particles move towards the suitable search space along the searching process www.nature.com/scientificreports/ and consequently closer to the optimum spot. Each particle that has its position, acceleration, and fitness value displays a solution in search space. Every particle goes to its best place and best position of particle swarm at each iteration. The movement of particles depends on the speed change calculated using Eq. (1), expressed as: is particle i speed on the jth dimension at t iteration, − → x ij (t) is particle i position on the jth dimension at t iteration, r 1 and r 2 are random numbers between (0,1), W is a constant, ϕ 1 and ϕ 2 are constants that are known as velocity coefficients, − → p ij (t) is the best position of a particle, and − → p gi (t) is the best position of the particle swarm. The new position of a particle is calculated using Eq. (2). Deep convolutional neural networks. Deep learning is a subcategory of machine learning algorithms with various architectures, such as, inter alia, deep neural networks (DNNs), CNNs, and recurrent neural networks (RNNs). CNN is one of the common algorithms utilized for image classification and recognition. Input, hidden, and fully connected layers are elements of a CNN's construction. Figure 2 illustrates the deep convolutional neural networks (DCNN) image classification process utilized herein. The hidden layers comprise convolutional and pooling layers. In this algorithm, the pixel values of the image are turned to an array as input for CNN. The convolutional layers are the basis of CNN, which are employed as the feature extractor in this algorithm that these features discern images from each other. The neurons in these layers are arranged into feature maps. Neurons have a receptive field in a feature map connected to a neurons' neighborhood in the previous layer 38 . Figure 3 illustrates the outputs of the first convolutional layer in the proposed DCNN model by extracting the features of the image, leading to training an appropriate DCNN model. Rectified linear unit (ReLU) is an activation function that applies non-linearity to the network, which this non-linearity helps produce the non-linearity boundaries. The introduction of this non-linearity to networks makes CNN an accurate algorithm. ReLU activation function is formulated in Eq. (3), and the output of a convolutional layer and ReLU is calculated using Eq. (4). www.nature.com/scientificreports/ where Y n is the nth output of feature map, f is activation function, which is ReLU here, x is the input image, W n is the convolutional filter related to the nth feature map, and the multiplication sign refers to the 2D convolutional operator. Pooling layers have various types--decreasing the network's parameters and computations by reducing the network's spatial extent is the purpose of these layers types. The Batch-Normalization layer applies the batch normalization method, which converts the inputs to a mean of zero and a standard deviation of one 39 . The DCCN model training process is accelerated by this method. Dropout is a technique in which a percentage of randomly selected neurons are ignored. In other words, the contribution of these neurons is removed temporally on the forward propagation, and weight updating is not applied on the backward pass 40 . This technique helps avoid overfitting (i.e., the predictive model learns well but cannot predict correctly). The flatten layer transforms a three-dimensional input into a one-dimensional vector to prepare it for the fully connected layer. The fully connected is placed after the several convolutional and pooling layers to interpret and represent the features that have been extracted 41 . There are two classes in the binary classification problem; therefore, the Sigmoid activation function is utilized. The sigmoid function is formulated in Eq. (5) as: Several optimization algorithms were developed based on the stochastic gradient descent (SGD) algorithm, such as root means square propagation (RMSProp), adaptive moment estimation (Adam), and adaptive gradient algorithm (AdaGrad). In other words, these algorithms are extensions of the SGD algorithm. Adam is an optimization algorithm roughly combining AdaGrad and RMSProp algorithms 42 . Adam takes both algorithms superior, which uses the squared gradients to scale the learning rate like RMSProp and utilize the moving average of the gradient like AdaGrad. Adam updates the weights in a way formulated in Eq. (6) . www.nature.com/scientificreports/ where ǫ is a small scalar, m is the first moment (i.e., m is mean), β 1 and β 2 are hyperparameters, v is the second moment (i.e., v is uncentered variance), w is model weight, η is the learning rate (step size), and C is the cost function. Adam optimizer helps to strengthen the training efficiency and learning progress 43 . sification. The process starts by resizing the image to 100 × 100 × 3, which 100 × 100 represents the height and width whereas × 3 refers to the number of color channels (i.e., Red, Green, and Blue). Figure 4 illustrates some samples of the dataset. The dataset contains two classes, which are fresh and rotten meat, which are extracted from existing dataset (i.e., Ulucan, Karakaya 36 ). The dataset is split into the training set (90%), validation set (5%), and test set (5%), so the training set contains 1706 images, whereas the test set and validation set included 95 photos each. Table 1 shows the number of samples assigned to each class used in training. Here, the amounts are unbalanced; therefore, each weight class is calculated and employed in the DCNN training process. The previous section showed the proposed layers that make the DCNN. Table 2 shows the proposed configuration of the model. The PSO optimization process selects the number of filters in convolutional layers and the learning rate. Evaluation metrics. There are various criteria to evaluate the proposed model, such as Precision, Recall, F1-score, and Accuracy. The Precision, Recall, F1-score, and Accuracy metrics are formulated in Eqs. (7)- (10) as: www.nature.com/scientificreports/ where True Positive (TP) represents the predictive model predicted positive, and the primary value is positive, True Negative (TN) indicates the predictive model predicted negative, and the primary value is negative, False Positive (FP) refers to the predictive model predicted positive, but the primary value is negative (Type 1 error), and False Negative (FN) demonstrates the predictive model predicted negative, but the primary value is positive (Type 2 error). In this section, the PSO algorithm adjusts two critical hyperparameters, namely learning rate and number of filters in the convolutional layer. The DCNN model is trained, and several evaluation metrics are utilized to assess the model performance. Learning rate is a significant hyperparameter in deep learning models that controls the model changes in response to the estimation error of updating model weights 44 . If the amount of the learning rate selected is too small, the training process gets extended or even get stuck. In contrast, a too large amount will lead to an unstable training process and unproperly weight updating 44 . The number of different ways of extracting features from an image is determined based on the number of filters in the convolutional layers 45 . The more filters, the more features can be extracted, but this rule is not always proper for CNN models; therefore, the number of filters must be adjusted. Thus, the PSO algorithm, a popular population-based optimization method, is utilized to adjust the number of filters in convolutional layers and the learning rate in the proposed DCNN model. The detailed configuration of the proposed DCNN model is denoted in Table 3 , in which the PSO algorithm achieved the number of filters for the convolutional layers, and the learning rate was earned 0.001 among 1, 0.1, 0.01, and 0.001 by the PSO too. Evaluation of the proposed DCNN model. The classifier model is built based on the configuration mentioned in Table 3 , the number of filters and the learning rate that the PSO achieved. Figure 5 shows the training and validation process of the DCNN model. www.nature.com/scientificreports/ Some evaluation metrics such as the confusion matrix, Precision, Recall, F1-score, and Accuracy are employed to assess the predictive model. Figure 6 illustrates the results of the confusion matrix, whereas other metrics are presented in Table 4 . A comparison between the output of this study and existing literature, viz. Ulucan, Karakaya 36 is outlined in Table 5 , which shows the results achieved in this paper have better performance than previous literature. Our model proposed an intelligence method that is more efficient in reducing human error and losses while increasing the accuracy and availability in various sections of the meat supply chain than traditional monitoring and control systems that are considered in previous studies [46] [47] [48] [49] . Moreover, the model has better performance than the previous research 36 on deep learning model by utilizing PSO algorithm and proper model architecture. This research provided a DCNN and PSO algorithm to train an image classifier system to control and distinguish www.nature.com/scientificreports/ wholesome meats from spoiled ones. The PSO algorithm was utilized to tune two critical features (i.e., learning rate and the number of filters in convolutional layers) that significantly affect the performance of DCNN models. One of the common problems in image classification problems is overfitting (i.e., model is trained well but cannot predict appropriately); therefore, by using the correct architecture and Dropout method, the DCNN model, as shown in Fig. 5 , appears more robust without this problem. Several metrics, namely accuracy, Precision, Recall, and F1-score--which are calculated based on the Confusion matrix, were applied to assess the DCNN model performance and indicate the model efficiency. Based on the results outlined in Table 4 , the DCNN model with 100% accuracy has remarkable performance. The industry 4.0 and SDG12 paradigms recommend utilizing new technologies such as artificial intelligence in manufacturing and supply chain management to increase productivity and decrease expenses and losses. In this paper, the deep learning model was employed to automate the controlling and separation process of wholesome meats from spoiled ones. This system can be utilized in transportation, storage, and retail sections in the meat supply chain to control and monitor the meats. This process is necessary to reduce bacteria effects in spoiled meat that may contaminate wholesome meats, hence, improving shelf life and saving final customers from harmful risks due to food spoilage. This system displaces manual monitoring and controlling; hence, it can be active most often and mitigate human errors, leading to enhanced shelf life while decreasing losses and increasing productivity, which are the goals of industry 4.0 and SDG12 paradigms. This paper proposed the application of artificial intelligence in meat supply chain management based on industry 4.0 and SDG12 paradigms. A classifier was trained by the DCNN and PSO algorithms with 100% accuracy, which distinguishes wholesome meats from spoiled ones. This model is utilized in various steps of the meat supply chain, which increases productivity, reduces cost, and avoids the bacteria effects of rotten meats on healthy ones by automating the separation process. Besides, by enhancing the meat shelf life, consumer confidence and preferences for meat can increase, hence, increasing economic productivity. For future research, this system can Fresh meat 100 100 100 100 Rotten meat 100 100 100 100 www.nature.com/scientificreports/ Food Waste is Everyone's Problem ~: text= The% 20Meat% 20of% 20the% 20Mat ter& text= Animal% 20pro ducts% 20may% 20only% 20acc ount Ways to Save Meat Before It Gets Wasted Logistics Work, ergonomics and social sustainability: Empirical musculoskeletal system strain assessment in retail intralogistics Implementation of artificial intelligence in agriculture for optimisation of irrigation and application of pesticides and herbicides The consumer footprint: Monitoring sustainable development goal 12 with process-based life cycle assessment Ensure Sustainable Consumption and Production Patterns Moving Forward on Food Loss and Waste Reduction 2-13 (FAO Risk propagation mechanisms and risk management strategies for a sustainable perishable products supply chain Carbon and water footprint of pork supply chain in Catalonia: From feed to final products Risk mitigation in a meat supply chain with options of redirection Is Meat from Stressed Animals Unhealthy? The European Union control strategy for Campylobacter spp. in the broiler meat chain Sustainable packaging development for fresh food supply chains Quality tracing in meat supply chains A proactive model in sustainable food supply chain: Insight from a case study Meat spoilage: A critical review of a neglected alteration due to ropy slime producing bacteria Bacterial spoilage of meat and cured meat products Meat Safety for the Consumer Recent advancements in smart manufacturing technology for modern industrial revolution: a survey Can organic sources of nutrients increase crop yields to meet global food demand Gaining Supply Chain Efficiency With Artificial Intelligence A machine learning-based model for the estimation of the temperature-dependent moduli of graphene oxide reinforced nanocomposites and its application in a thermally affected buckling analysis An integrating machine learning algorithm and simulation method for improving Software Project Management: A real case study Targeted vaccination for Covid-19 based on machine learning model: A case study of Jobs' prioritization A deep learning-based model to reduce costs and increase productivity in the case of small datasets: A case study in cotton cultivation Machine Learning in Manufacturing and Industry 4.0 Applications A procedure for tracing supply chains for perishable food based on blockchain, machine learning and fuzzy logic Improving efficiency of RFID-based traceability system for perishable food by utilizing IoT sensors and machine learning model Machine learning applied to near-infrared spectra for chicken meat classification Hyperspectral Imaging and Deep Learning for Food Safety Assessment Understanding the learning mechanism of convolutional neural networks in spectral analysis Detection of red-meat adulteration by deep spectral-spatial features in hyperspectral images Efficient extraction of deep image features using convolutional neural network (CNN) for applications in detecting and analysing complex food matrices Meat quality assessment based on deep learning Particle swarm optimization (PSO) for the constrained portfolio optimization problem Deep learning Batch normalization: Accelerating deep network training by reducing internal covariate shift Dropout: A simple way to prevent neural networks from overfitting Improving Neural Networks by Preventing Co-Adaptation of Feature Detectors A Method for Stochastic Optimization Machine learning and deep learning based computational techniques in automatic agricultural diseases detection: Methodologies, applications, and challenges Understand the impact of learning rate on neural network performance How do convolutional layers work in deep learning neural networks? Investments in tracking and tracing systems in the food industry: A PLS analysis Perspectives on food traceability: A systematic literature review. Supply Chain Manage What do we think we eat? Single tracing method across foodstuff of animal origin found in Greek market Wild meat is still on the menu: Progress in wild meat research, policy, and practice from 2002 to 2020 This research received no external funding. The authors declare no competing interests. Correspondence and requests for materials should be addressed to M.A.A. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.