Journal:A blockchain-driven IoT-based food quality traceability system for dairy products using a deep learning model

Full article title	A blockchain-driven IoT-based food quality traceability system for dairy products using a deep learning model
Journal	High-Confidence Computing
Author(s)	Manisha, Noothi; Jagadeeshwar, Madiraju
Author affiliation(s)	Chaitanya Deemed to be University
Primary contact	noothimanisha6 at gmail dot com
Year published	2023
Volume and issue	3(3)
Article #	100121
DOI	10.1016/j.hcc.2023.100121
ISSN	2667-2952
Distribution license	Attribution-NonCommercial-NoDerivs 4.0 International
Website	https://www.sciencedirect.com/science/article/pii/S2667295223000193
Download	https://www.sciencedirect.com/science/article/pii/S2667295223000193/pdfft (PDF)

This article contains rendered mathematical formulae. You may require the TeX All the Things plugin for Chrome or the Native MathML add-on and fonts for Firefox if they don't render properly for you.

This article should be considered a work in progress and incomplete. Consider this article incomplete until this notice is removed.

Abstract

Food traceability is a critical factor that can ensure food safety while enhancing the credibility of the manufactured product, thus achieving heightened user satisfaction and loyalty. The perishable food supply chain (PFSC) requires paramount care for ensuring quality owing to the limited product life. The PFSC comprises of multiple organizations with varied interests and is more likely to be hesitant in sharing the traceability details among one another owing to a lack of trust, which can be overcome by using blockchain. In this research, an efficient scheme using a blockchain-enabled deep residual network (BC-DRN) is developed to provide food traceability for dairy products. Here, food traceability is determined by using various modular tools, like the internet of things (IoT), blockchain data management, food traceability blockchain architecture, and DRN-based food quality evaluation tools. The devised BC-DRN-based food quality traceability system is examined based on performance metrics such as sensitivity, response time, and testing accuracy, and it has attained better values of 0.939, 109.564 s, and 0.931, respectively.

Keywords: perishable food supply chain, blockchain, internet of things, food traceability, deep learning

Introduction

In supply chains, the process of perishable products management is extremely complex. This process differs from other products based on various aspects, like deterioration rate, cold storage requirement, and shelf life, from the production unit to the end user. [1][2] Among the various foods in the perishable food supply chain (PFSC), milk and dairy products are the major sources of nutrients. However, poor work practices, animal feed, and the environment may introduce chemical components and contaminants in those milk and dairy products, putting consumers at risk. [3] Increasingly, consumers have become more attentive towards the quality, source, transport [4][5], and shipping regulations surrounding purchasable food items. [6] Depending on the scenario, there is a varying necessity to trace those food items for achieving better supply chain management (SCM). [7] While globalization has complicated SCM and the traceability of food items [8], blockchain, a type of distributed ledger technology (DLT), may have a role to play in solving such problems.

Blockchain is considered to be a significant DLT technology impacting commerce, wherein information can be stored in a database and shared among organizations, which have no mutual trust. [9] Blockchain is an underdeveloped technology [10], where few of the applications require trust, which is already devised. [11][12] The data in blockchain cannot be falsified [13] and has the capability for providing trustworthy, secure, and transparent data in public as well as private domains. [14] It is founded on a distributed ledger that is not controlled or owned by a single user. The major benefit of using blockchain is that it can be employed in the PFSC for enhancing the integrity of data gathered from multiple entities in the supply chain. [15]

Food traceability is component of PFSC logistics, where the data regarding any food item at every stage is captured, stored, and transmitted across the supply chain for enabling quality checks of the product at various points. Moreover, it also provides a way for tracking the product at any required time. [3][16] The primary step for providing traceability is the adequate depiction of traceable resource units (TRUs) for identifying the logistic and production units, traceability of products, and so on for managing the PFSC. [17]

Various studies have focused on developing traceability solutions in the PFSC for various products. For instance, Bumblauskas et al. proposed a blockchain-oriented traceability scheme considering the eggs, wherein the eggs were tracked from farms to the end users. [18] Cao et al. described a traceability scheme for strengthening the trust in the existing supply chain between China and Australia. [6] Similarly, Khan et al. proposed a food provenance system for handling data on the traceability of meat products using blockchain, internet of things (IoT), and sophisticated deep learning approaches [19], each with their own complexities. [20–23] Other similar approaches have considered a variety of perishable products, like beef, fish, and other agricultural products. [24] A dairy case study was considered by Casino et al., wherein a distributed trustless and secure scheme was devised for tracing the dairy products of a company. Additionally, Cocco et al. developed a blockchain-based model for managing the supply chain of a traditional bakery. Finally, the major critical aspects affecting the safety and quality of food, while storing and distributing the products, demand monitoring the humidity and temperature. The utilization of radio frequency identification (RFID) traceability schemes can be highly effective during monitoring of these and other processes. [27][28]

This research work mainly focuses on the development of a food quality traceability system using blockchain, IoT, and deep learning architecture. Here, the devised food traceability scheme is implemented using various modular tools, like the internet of things (IoT), blockchain data management, food traceability blockchain architecture, and deep residual network (DRN)-based food quality evaluation tools. Data from dairy product manufacturing and its associated supply chains are collected using IoT technologies. The collected data is then transformed into a sequence of blocks in the blockchain network, wherein each node represents the entity involved. The food traceability blockchain is developed by considering the consumers’ purchases of dairy products, product IDs, etc. This blockchain-driven IoT-based deep learning system for improving the food traceability process is designed to monitor the effects of dairy product supply chains, wherein DRN techniques are employed for grading the milk.

The remaining portions of this work are as such. The next section details the related works in food traceability, while the third section elaborates on the system model of IoT-based blockchain-driven food traceability schemes, and the fourth section discusses the methodology used to develop those schemes. The results and discussion of the developed method are discussed in the penultimate section, followed by conclusions about the work.

Literature review

Though several studies have been carried out for developing effective food traceability systems for edible goods, only eight works are considered for review here. Mondal et al. [15] used an RFID-based sensor with blockchain for producing a more transparent PFSC. Here, a consensus algorithm named Proof-of-Object (PoO) was applied, which made use of the variation found in cryptocurrency as well as supply chain for avoiding needless transactions added to the blockchain. This technique was highly efficient in making the blockchain immune to cyber attacks. However, this method suffered from having inadequate security. Bechtsis et al. [29] developed a containerized food supply chain framework for securing the information shared in the PFSC. Here, a double-staged containerized food supply chain was implemented by considering the Hyperledger Fabric open-source framework. This system offered enhanced traceability and process control, and prevented possible risks. However, the huge amount of information exchanged resulted in high costs. Hao et al. [12] developed a visual analysis technique for detecting the food safety risks in the PFSC. This scheme utilized the combination of visualization and blockchain technology, wherein features such as tamper-resistance and distribution were utilized for ensuring the authenticity of data. This scheme provided a scientific platform for managing the PFSC with the minimization of potential risks, but it suffered from low scalability. Gao et al. [30] proposed a Hyperledger Fabric-based traceability system in the PFSC using blockchain technology. Governors were introduced by the system for strengthening the regulation, and a market was introduced for providing a platform to search for information regarding food and transactions. This system was effective in achieving high throughput, but it achieved poor system functionality and efficiency.

Behnke and Janssen [16] devised a blockchain-driven technology for providing an effective solution for food traceability. The approach utilized a set of boundary criteria for providing agriculture and food traceability, thereby assuring information sharing. This method was successful in achieving high transparency, thus gaining user trust; however, this method needed to find boundary criteria for determining the changes in blockchain ahead of its adoption and had high complexity. Rambhia et al. [31] introduced Agrichain, a system for managing the PFSC using blockchain. Here, SCM activities were streamlined by using smart contracts and the Ethereum blockchain. Moreover, the interaction among different entities was captured, thus guaranteeing authenticity, traceability, and transparency. This approach was effective in ensuring authenticity by disallowing illegal changes, but it failed in handling real-time applications.

Shahbazi and Byun [7] developed a machine learning (ML) technology and fuzzy logic for tracing the products in the PFSC. This system was implemented using three parts: prediction of expiry date by using ML approaches, blockchain-based data sharing, and fuzzy-oriented food quality evaluation. The scheme was highly efficient in improving the shelf life in the PFSC. However, the supply chain formed by this scheme was insecure and unreliable, thus affecting the effectiveness of the approach. Greater reliability was provided by Ehsan et al. [32], wherein they developed a conceptual model for developing a centralized scheme for ensuring the traceability of the supply chain. Here, a blockchain-oriented scheme was developed, wherein the supply chain was effectively managed using a multi-agent system. This approach was utilized for optimizing any supply chain with high efficiency and security. However, this method was ineffective when trying to handle distinct supply chains in different formats.

System model

This section contains a detailed description of the system model of the devised food traceability system, specifically for milk and other dairy products. The system model is depicted in Fig. 1. The traceability system model is devised for ensuring the products are delivered with good quality and safety in the PFSC. The milk and dairy products collected from the dairy manufacturers are tagged with RFID tags, which enable end-to-end tracing of the products. The RFID tag comprises of data concerning the product in read-only or read–write format. The electromagnetic (EM) energy triggers the RFID tags, when they enter the range of the scanning antenna, and the data is sent in the form of radio waves. The antenna picks up the radio wave and it is forwarded to the RFID reader, where the waves are converted into digital data. The data acquired by the RFID reader/gateway is gathered by the IoT data collector. The collected data is converted into the block at the blockchain data management module, wherein the blocks comprise of records for all the activities that occurred during its creation. Moreover, every block is provided with a field containing the hash value of the prior blockhead, thus ensuring that all blocks can effectively point to their prior blocks. The generated blockchain data is fed to the food traceability blockchain architecture, wherein the multiple entities through which the transfer of products happen is considered. Every time the product goes through a new entity in the supply chain, a block is added to the blockchain, thereby providing a way to trace the product. After effective food traceability is determined, a food quality evaluation is performed for grading the milk to determine its quality.

Figure 1. System model of the food quality traceability module.

Blockchain-based food traceability prediction

The data collected by the IoT data collector is converted into the blockchain format, wherein the initial block contains the information related to the product along with the timestamp and hash value. The blockchain assists in enabling the flexibility of the system. The milk or dairy products are initially supplied by the manufacturer. The products are assigned with a batch ID [7] in the genesis block, and the tracking information stored in the cloud is contained in the supply chain process. The product can be traced by considering batch/ID as well as container ID. The blockchain data is updated throughout every stage in the supply chain, and this data is stored in the database. The supply chain forms initially send the food traceability data to the regulatory bodies, which in turn send the qualified check-up certificate. Once the certificate is received, the traceability data along with the hash value is forwarded to the blockchain nodes, which maintain the shared ledger. The nodes verify the credibility of the data received and based on the credibility, blockchain nodes transmit a reject or accept message to the supply chain. If an accepted message is received, then the original traceability matrix is added to the database. Later, the shipment process is notified as waiting, ready, transporting, or received. As the blockchain is updated at each and every event, the consumers can back trace the products by using the blockchain.

Food traceability system

A basic supply chain is considered, which comprises of producers, distributors, consumers, and transporters. [27] The milk products are transferred from the manufacturer to the end users by means of various supply chain partners. Initially, the manufacturer produces dairy products, which are packed in a special box containing the passive RFID tag. The tagged goods are transported into the cold storage ahead of delivery through the RFID gate, wherein the reader collects a unique identifier (UID) and electronic product code (EPC). These details are then conveyed to the host computer, which generates product information, like business steps (e.g., status of the product), event time, business location, and EPC. All these details are later forwarded to the server-side EPC Information Service (EPCIS). When the goods are ready to be shifted, the transporter transfers the goods to the distributor. The cold storage facilities are available at the producer and the distributor units, which utilize fixed RFID readers, whereas the transporter employs a hand-held device. At the distributor, the tagged goods are scanned by the RFID reader before storage and, finally, the goods are delivered to the user by another transporter. Every time the products are moved, the traceability system records the tagged goods. The producer and distributor use a tag direction module for determining the direction of the tag. The tag direction module identifies the direction like moving in or out of the facility using the ML model, which enables easy identification of the received and shipped products. Further, producers, distributors, and transporters utilize IoT-based sensors for monitoring humidity and temperature. The IoT sensor data as well as RFID data are then transmitted to the web service for storage in the database. Thus, real-time traceability of products can be achieved and the safety, as well as quality of products, can be ensured throughout the supply chain. Fig. 2 depicts a block diagram of the food traceability system.

Figure 2. Block diagram of the food traceability system.

Food quality evaluation using the introduced BC-DRN

In this section, the deep learning-based food quality evaluation system is detailed. The devised BC-DRN food traceability scheme is implemented with several modular tools, like IoT, blockchain data management, food traceability blockchain architecture, and DRN-based food quality evaluation tools. Initially, in the IoT module, IoT nodes are simulated, wherein the product’s RFID data is collected from the dairy product manufacturing chains and supply chains [27][33] and it is transmitted. The collected data is then converted into the blockchain format, which enables tracking of the product through the supply chain and ensures security as the blockchain cannot be tampered with. The blockchain is kept updated every time an event happens in the supply chain, thus ensuring the traceability of the dairy products. Once food traceability is determined, the DRN-based food quality evaluation module is employed for assessing the quality of milk by performing grading. The IoT module, blockchain data management module, and food traceability blockchain architecture are already detailed in the prior section.

The devised DRN-based food quality evaluation is implemented using the following steps. The data is initially pre-processed using Z-score normalization [34] for normalizing the data. From the normalized data, the prominent features are selected by using the Canberra distance. [35] The selected features are then subjected to the DRN [36], which performs grading of the milk as 1 (Good), 0.5 (Moderate), or 0 (Bad). In Fig. 3, the devised food quality evaluation scheme is portrayed, and the process is elaborated in the following subsections.

Figure 3. Block diagram of the DRN-based food quality evaluation scheme.

Data acquisition

Consider the dataset A containing n number of milk samples, where each sample corresponds to the color, turbidity, fat, odor, taste, temperature, and pH of the milk. The dataset can be represented as:

$A=\left\{{A_{1},A_{2},\ldots ,A_{l},\ldots ,A_{n}}\right\}$ ,

where ${A_{1}}$ denotes the ${l}$ th sample, which is subjected to the normalization process.

Data pre-processing

The acquired data ${A_{l}}$ is forwarded to the pre-processing phase, where pre-processing is carried out by using Z-score normalization. [34] Z-Score is normally used as the primary process in an artificial neural network. It offers the advantage of considering the variability and mean of the raw data and high computational capability. Here, normalization is performed by considering the standard deviation and mean of the input variables, in such a way that the mean and the standard deviation of all features are equated to 0 and 1 respectively. Z-score normalization is given by:

$b_{l}={\frac {A_{l}-\mu _{l}}{\sigma _{l}}}$ .

Here, $\mu _{l}$ and $\sigma _{l}$ denote the arithmetic mean and variance of ${A_{l}}$ , and $b_{l}$ indicates the normalized data which is forwarded to feature selection.

Feature selection

The normalized data $b_{l}$ is then subjected c to the feature selection process, for extracting the most significant features from pre-processed data. Here, Canberra distance [35] is used for extracting the features by finding the distance between a feature and a class label. Canberra distance offers high robustness to outliers. The following expression gives the Canberra distance:

$CD(g,h)=\sum \limits _{j=1}^{m}{\frac {\left|g_{j}-h_{j}\right|}{\left|g_{j}+h_{j}\right|}}$ .

Here, $g$ indicates the candidate feature and $h$ denotes the target, and $m$ specifies the dimension of the vectors. Consider the features chosen to be represented by $fs$ , which are given as:

$fs=\left\{{f_{1},f_{2},\ldots ,f_{k}}\right\}\quad k<m$ .

The selected features are given as input to the DRN for food quality evaluation.

Food quality evaluation using DRN

Food quality is evaluated using DRN, wherein the selected features are passed to the DRN for grading the milk into three categories: 1 (Good), 0.5 (Moderate), and 0 (Bad). The grading of the milk is done for evaluating the quality of milk, wherein grading is done based on the turbidity, fat, odor, and taste.

DRN architecture

DRN [36] is employed for evaluating food quality as it demonstrates the capability of categorizing the milk with minimum error and superior accuracy. Moreover, DRN effectively avoids overfitting problems and provides high computation efficiency. The classification accuracy can be enhanced by increasing the layer count, but doing so results in an explosion and gradient disappearance. This issue can be effectively overcome by using the residual blocks, which also accelerate the training process. Here, high performance is achieved by enhancing the depth of the network rather than the width. The DRN comprises of a number of layers, such as the convolutional (conv) layer, pooling layer, batch normalization layer, and activation layer.

Conv layer

This is the most significant layer in the DRN, wherein the input data is scanned by the convolution kernel and later convolution is performed for extracting the feature details. Here, kernel denotes the series of filters that are employed for processing the input data. Convolution operation can be specified by:

$R_{k}^{z+1}=v\left({\sum \limits _{y\in E_{k}}r_{k}^{z}\ast p_{y,k}^{z+1}+w_{k}^{z}}\right)$

Here, the activation function is indicated by $v$ , with $R_{k}^{z+1}$ designating the $k$ th feature map of the $(z+1)$ th convolution layer, which is supplied with the input $r_{k}^{z}$ . Here also, $p$ specifies the convolution kernel, $E$ represents the feature map set, and the bias is denoted by $w$ . The activation function is used to obtain a new feature map.

Pooling layer

This layer performs downsampling for minimizing the dimension of the feature map, thereby avoiding overfitting issues. The max pooling is the most commonly used model owing to its high efficiency and simplicity of operation.

Residual blocks

Residual block forms the core of DRN, which enables direct connection of input and output blocks in case the dimension of both input and output is identical. The below-given expression represents the residual block:

$q=V\left({r,\left\{U_{i}\right\}}\right)+r$
$V=U_{2}\beta \left({U_{1}r}\right)=q-r$

Here, $q$ indicates the output and $r$ is the input of the present layer. $V\left({r,\left\{U_{i}\right\}}\right)$ indicates the residual mapping that has to be learned, $\beta$ represents the Rectified Linear Unit (ReLU) function, $U_{1}$ and $U_{2}$ specify the weight of layer 1 and 2, correspondingly. The target output is made to attain $V+r$ by the process of the learning in the sub-modules with the shortcut connections in the residual blocks.

Batch normalization (BN)

After every conv layer, the BN layer is added for normalizing the outputs by making the variance and mean of the output as 1 and 0. BN effectively minimizes the training epochs and makes the learning process stable. The process of BN is represented by:

References

Notes

This presentation is faithful to the original, with only a few minor changes to presentation and updates to spelling and grammar (including to the title). In some cases important information was missing from the references, and that information was added. Some citations were vague as to how they contributed, or some statements didn't match well with their citations; minor tweaks were made to have the text and citations match up better. The key contribution portion of the introduction was removed for this version, as it was redundant and simply repeated the prior paragraph. No other changes were made in accordance with the "No-Derivatives" portion of the distribution license.