Radio frequency identification (RFID) technology is rapidly emerging as one of key technologies for Industrial Internet of Things (IIoT) [1]. RFID system provide the ability to automatically identify and track objects and persons in a non-contact, non-line-of-sight manner. This enables the development of very different automated item management frameworks and thus provides a compelling business case for the rapid adoption of RFID systems. Many research results prove that RFID technology can be applied to various industrial applications such as factory [2], inventory management [3,4], medical privacy protection [5] and agriculture [6,7]. The technological level is also constantly innovating. Passive technology in RFID has matured. Its equipment does not require power supply and maintenance, and is being increasingly used. The components of RFID systems are multiple tags, a reader and a back-end server. When multiple RFID tags communication with a singleton reader simultaneously, interference will occur between the signals, resulting in reader to fail to fail to obtain the tag ID and thus affects the stability and settlement efficiency of the RFID system. In order to solve the tag collision problem in a dense environment, an anti-collision mechanism is needed to coordinate the communication between the reader and tags, which is named tag identification or anti-collision algorithm.

The rapid multi-tag identification is the first type of issue that attracted the attention of researchers in the RFID community. Generally speaking, mainstream tag identification algorithms are divided into two categories according to the functional characteristics: Aloha-based algorithms [8–13], deterministic algorithms [14–17] contains query tree-based and binary search algorithms. The Aloha-based algorithms are intuitive solution but they do not appear to be scalable. The efficiency of the protocols is highly affected by the cardinality. An advantage of the Aloha-based protocol is that it is simple to implement. The disadvantage of them is the slow throughput under high-traffic loads; also they are sensitive to changes in the number of tags, that is, when the frame length setting is not appropriate, the performance will drop sharply. In contrast, the performance of query tree-based algorithms [15–17] will not be affected by the tag cardinality, i.e., its performance almost maintain a constant value as the number of tags increases. One disadvantage of these protocols is that the length of probe command of such algorithm is not fixed, it will change with the change of the query prefix length during the tag identification process. In addition, query tree-based algorithm consumes more transmitted bits. Therefore, many researchers proposed hybrid algorithms [18,19]. The literature [18] shows us three multi-tag identification algorithms based on binary tree slotted Aloha (BTSA). Among them, the Splitting BTSA algorithm with the best performance has a throughput of 0.425. The literature [19] proposed a new kind of tree slotted Aloha protocol, in which the whole tag set can be divided into many smaller groups. Then the reader identify the tags group by group via tree splitting strategy. However, because both the Aloha-based algorithm and tree splitting algorithm are random access algorithm in nature, thus the algorithm proposed in [19] cannot guarantee that all tags are completely identified.

In this paper, we designed and implemented a hybrid architecture of tag identification, and proposed two hybrid tag identification algorithms. These two algorithms inherit the advantages of Aloha-based and deterministic algorithms, and abandon mutual shortcomings. In particular, our proposed algorithms do not need to provide an accurate estimation of the number of tags in collision phase, thereby avoiding the negative impact of estimation errors on performance, and reducing the computational complexity. Moreover, their efficiency will not be affected by the number of tags and tag IDs distribution.

The remainder of this paper is organized as follows. In Section 2, we reviewed the existing the anti-collision algorithms. Section 3 presents the system model and our proposed hybrid algorithm. In Section 4, mathematical analysis and performance evaluation through simulations is conducted. Section 5 concludes the whole paper.

As mentioned above, existing anti-collision algorithm mainly consists of two types: Aloha-based and deterministic algorithm. The deterministic algorithm can be further divided into two types: Query tree [14,15,17] and tree splitting [16] algorithms. In deterministic algorithm, the reader uses coding number carried by the tag to allow the tag to perform reading priority sorting. After sorting, the reader identifies the tags start from the bottom of the fork. Law [1] firstly proposed query tree (QT) anti-collision algorithm. In QT algorithm, the reader maintains a stack to store the query prefix. If the tags’ sequence number in a certain bit are the same, the reader continues to push new query prefixes into the stack. For the prefix, the reader recognizes the tags sequentially according to the order. This is the earliest QT algorithm. At present, there are many algorithms that have improved the QT classic algorithm, mainly through the eigenvalues of collision bits and the addition of additional queries. In [14], the authors proposed an improved QT algorithm to enhance the identification performance. The literature classified a large number of tag prefixes according to similarity, just like a big tree is divided into many branches, each type of the same prefix represents a tag groups, reader probes each group separately to reduce collisions and the transmissions of empty slots. However, such algorithm adds some extra queries, which increases the additional query time communication complexity in downlink. Some authors [20] proposed a query method based on a 4-ary tree to avoid unnecessary collisions and improve the efficiency of traverse. Specifically, the reader uses the characteristics of the highest two collision bits. The number of forks in the query tree can be adjusted to reduce unnecessary collision slots. The advantage of the deterministic algorithm is that it can ensure that the tag group is completely identified. Wang [21] proposed a binary search tree algorithm (PNBA) based on physical layer network coding, which pushed the conflicting signal information of multiple labels to a stack and discarded it by the traditional anti-collision algorithm. Dong [22] proposed an improved binary search conflict prevention protocol, BRTP, which allows two groups of tags to respond to a reader in the same slot in a dual-response mechanism.

In Aloha-based algorithm, each tag randomly picks up a slot to respond to the reader after extracting the parameters including frame size and communication rate from the reader’s command. The representative is the adaptive Q-algorithm based on UHF RFID standard EPC C1 Gen2 protocol. According to [23], there are three types of probe commands: Query, QueryRep, and QueryAdj. An identification round is defined as the time duration between successive probe commands issued by the reader and therefore only one Query command exists in each identification round. Therefore, the reader issues many Queryadj or QueryRep commands in order to identify tags during each identification round. In particular, QueryAdj command carries the value of parameter Q and instructs all tags to adjust their value of Q. All tags receiving the new value of Q reselect their slot counter based on the new Q (between 0 and 2^Q −1); here, Q could be incremented by 1, decremented by 1, or have no change according to the adaptive Q algorithm. There are three types of access events: collision, empty, and success. The reader can send either a QueryAdj or QueryRep in the case of a collision; it can send a Query, QueryAdj, or QueryRep in the case of empty; and it can send either a QueryAdj or QueryRep in the case of success. There are four assumptions determining when to use which probe command, as follows.

The advantage of the Aloha-based algorithms represented by the Q-algorithm is that they are easy to implement on the reader side. However, the random nature of such type of algorithm will cause the performance to fluctuate with the number of tags.

Our proposed Adaptive M-ary tree slotted Aloha (AMTS) algorithm starts a round of identification process by broadcasting a probe command with a key parameter named frame size [24]. Each tag that receives the probe command will randomly picks up a time slot to respond to the reader and return its own ID information. Once a collision is detected, the reader will count the current value of slot counter (SC) and push it onto the stack. When the reader reads a frame, it will count all slot statistics and obtain the values corresponding to slot counter. Unlike the conventional Aloha-based anti-collision strategies, the reader does not use the number of slot statistics to estimate the number of unread tags and attempts to identify the collided tags in the following frames. In AMTS, the reader will use the query tree algorithm to directly identify the involved tags in each collision slots, which is viewed as an independent identification process of collision slots. Only when all collisions are resolved, the entire process will end. Fig. 1 depicts an example of using AMTS algorithm to identify ten tags. As shown in the figure, there are 2 collision slots in the entire frame, which are located in the 1-st slot and 5-th slot, respectively. We can see that the reader uses a 2-ary tree method to resolve three tags in the first collision slot and uses a 4-ary tree method to resolve five tags in the second collision slot. Through this example, we can intuitively know the advantages of AMTS algorithm. Specifically, we designed two AMTS algorithms, which are based on dynamic frame size adjustment (CF-AMTS) and adaptive m-ary selection (MF-AMTS). The difference between them is that the former uses a collision factor to adjusts the m-ary search and the latter uses mapping table to estimate the number of tags involved in collision slots.

In what follows, we elaborate on the principles and operating mechanisms of these two algorithms.

(1) CF-AMTS: Algorithm description

In this paper, we define a variable called collision factor, which is expressed as

(1) μ=kn

In the above formula, k represents the number of collision bits in the current time slot, and n is the tag ID length. We make the following assumption that the current slot is a collision slot and there are m tags to be identified. For any bit in a string responded by the tag, the probability that it will not collide can be expressed as 1/2^{m −1}, Eq. (1) can be rewritten as

(2) μ=n[1-(1/2)m-1]n

Eq. (2) implies that the larger the value of m, the higher the collision factor. We assume that in the entire traversal tree, the highest query depth is N, which is the longest path required from the root of the traversal tree to the leaf nodes. The probability that a tag can be successfully recognized at the first level of traversal tree is expressed as P1=(1-1/N)m-1 and the probability that the tag can be successfully recognized at the r-th level of traversal tree is expressed as Pr=P1(1-P1)r-1. Then the expectation of r is calculated as

(3) E(r)=∑r=1∞rPr=∑ r=1∞rP1(1-P1)r-1=-P1d∑r=1∞[(1-P1)r]d(P1)=-P1d[1-P1P1]d(P1)=-P1(-P1-(1-P1)P12)=1(1-1/N)m-1

Correspondingly, we can derive the average number of slots spent by the reader to identify m tags.

(4) TN–ary=NE(r)=N(1-1/N)m-1

According to Eq. (4), we can get the following judgment conditions.

(5) {T2-ary≥T4-ary,m≥3T2-ary<T4-ary,m<3

Combining formula (2) and (5), the reader knows that it should use 2-ary tree to resolve the collided tags when m ≥ 3. If m is less than 3, the reader should use a 2-ary tree to resolve the collided tags. The collision factor is an indicator that reflects the number of tags in the collision slot. However, such indicator is not completely reliable because it is easily affected by the distribution of tag IDs. Therefore, if only the collision factor is used to estimate the number of remaining tags in a collision slot, a large error may be caused, thereby affecting the identification performance. To solve this problem, we introduce a new custom command QueryRP. It is worth noting that the EPC C1 Gen2 standard allows users to use custom commands, so this approach does not cause additional costs. The specific operation is as follows. When the collision factor is greater than 0.75, the reader will sends a QueryRP command to let the tags return a 4-bit string, which implies 2-bit collision bits information. The process of generating the 4-bit data returned by the tag is as follows. The tag executes the bitwise and operation of the first 4-bit of its ID with (1100), and then converts the highest 2-bit into a decimal value x, and then generates a new 4-bit string in which the x-th bit is 1, the rest bits are 0. As shown in Fig. 2, the reader uses CF-AMTS algorithm to identify two tags in a collision slot. From the figure we can see that the tag IDs are 11010001 and 11101010, respectively. When they collide in a time slot, the collision factor can be calculated to be 0.75. At this time, if the reader uses a 4-ary tree to resolve them, additional empty slots will be introduced. By making full use of the QueryRP command, the reader can parse out the tag ID prefix implicit in the collided string in the current slot, thereby avoiding the occurrence of empty slots.

(2) MF-AMTS: Algorithm description

In the next, we elaborate on the second AMTS algorithm, namely MF-AMTS algorithm. The salient feature of this algorithm is to use the mapping table to estimate the tag cardinality in a collision slot.

Tab. 1 is a mapping table that maps 2-bit original data to 4-bit string. In MF-AMTS, the tag will extract its own high 2-bit ID data according to the probe command sent by the reader, and map it into a 4-bit string based on the given mapping table. It is noted that the time slot for the tag to respond to the reader is no longer randomly selected. The tag will use its own ID to match the prefix parameter in the probe command sent by the reader and determine whether it responds or not. Assume that in a 4-bit mapped data, the number of collision bits (a collided bit is denoted as “x”) is t. Then the reader will use a 4-ary tree to resolve the collided tags when t is greater than or equal to 3, otherwise it will use a 2-ary tree. We still reuses the example in Fig. 2. The index of the collision slot selected by the tags is stored in the reader’s stack. The corresponding communication procedure is described in Tab. 2. We can observe that two tags have a common prefix of 11 in slot 1, so their responding data are (01→0010) and (10→0100), respectively. On the reader side, the received data string is 0xx0. According to the principle of MF-AMTS, the new generated prefix is 110 because t is 2.

By comparing Fig. 2 and Tab. 2, we can find that the number of slots consumed by the reader to identify the same batch of tags using MF-AMTS algorithm is less than CF-AMTS algorithm. The reason is given as follows. The CF-AMTS algorithm uses QueryRP command to produce the extra slots. Fig. 3 depicts the detailed flowchart of the AMTS algorithm proposed in this paper.

Where N_e, N_r and N_c are the number of idle slots, successful slots and collision slots, respectively. With the expected ratios of N_e, N_r and N_c which equals to 2:1:7/24 under the condition L = 2n according to the analysis of [7]. We can derive the rule of the adjustment of Q value as

(6) {Q=Q-1,Ne-6Nc>TthresholdQ=Q,otherwise

If Ne-6Nc>Tthreshold during the identification process, it presumes that L≥2n. To avoid too many idle slots and improve efficiency, the value of Q should be decreased by 1. In other case, Q keeps unchanged. From the above description, it is known that we design two tag identification algorithms based on hybrid architecture. The similarity between the two algorithms is that they both contain the AMS module. The key difference is that the criteria for CF-AMTS and MF-AMTS to enter the AMS module are different. The former is based on collision factor and custom commands. And the latter is based on mapping table.

In this paper, we present 2 multi-tag identification algorithms based on hybrid architecture to improve the reading efficiency in tag dense RFID scenarios. The contributions are concluded as follows. Compared with existing anti-collision algorithms, our proposed approaches in this paper not only have higher system throughput, but also do not require precise cardinality estimation of the entire tag set, and hence they can be easily implemented on low-cost RFID readers. Both theoretical analysis and various experimental results verify that our proposed algorithms are superior to prior art in terms of system throughput, the total number of slots and energy efficiency.