A Robust Reliable Low-Power High-Throughput Data Collection Wireless Sensor Network

A robust data collection protocol is proposed in this article for wireless sensor networks (WSNs) applications, which require high throughput, reliability, and low power. In terms of high throughput, time-division multiple access (TDMA) is more suitable than carrier sense multiple access (CSMA). Efficient TDMA protocol requires solid synchronization and more optimized time slot scheduling. Reliable data collection requires end-to-end packet retransmission. These factors inhibit the implementation of comprehensive data collection protocols. This article provides a heuristic TDMA scheduling algorithm for optimal time slot scheduling of multichannel communication. The combination of TDMA and multichannel communication enables to maintain high-throughput and low-power communication. Besides, constructive interference (CI)-based flooding technique is utilized in this article to deliver robust synchronization and efficient downward communication. Thus, we can combine CI-based flooding with packet retransmission for collection reliability. The proposed protocol is implemented on self-made sensor nodes with the Contiki operating system. The experimental results from a deployed network of 29 self-made sensor nodes show that the protocol can sustain persistent operation and provide a throughput of up to 14.74 kB/s with 100% data collection.


I. INTRODUCTION
W ITH the development of the Internet of Things (IoT), wireless sensor networks (WSNs) have attracted a lot of attention in recent years.WSNs are capable of capturing fine-grained information about the physical world [1], [2], so it is widely used in civil monitoring, agriculture, industry, health surveillance, and other fields [3], [4], [5], [6].However, sensor equipment for data acquisition and wireless communication transmission technology are not complete [7].
This article aims to achieve high throughput, reliability, and energy efficiency in data collection for WSNs.These capabilities are particularly crucial in applications such as earthquake monitoring, where large amounts of accelerator data need to be collected rapidly after an earthquake occurs [8].However, there exists a tradeoff between high throughput and energy efficiency [9], [10], [11].Most existing efforts focused on high-throughput data collection lack sufficient energy effi-Manuscript received 20 June 2024; accepted 10 July 2024.Date of publication 25 July 2024; date of current version 1 September 2024.The associate editor coordinating the review of this article and approving it for publication was Prof. Fangqing Wen.(Corresponding author: Jinzhi Liu.) The authors are with the College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266061, China (e-mail: jinzhi_liu@qust.edu.cn;Jzh928@outlook.com;linkaien001@gmail.com).
Digital Object Identifier 10.1109/JSEN.2024.3431017ciency.Balancing high throughput against power consumption is essential; while high throughput tends to increase the radio's ON-OFF duty cycle, energy efficiency requires minimizing radio usage.Carrier sense multiple access (CSMA)-based carrier sensing is unable to fully exploit the collection capability, whereas time-division multiple access (TDMA) allows for parallel transmission to avoid collisions [12], [13].Due to the half-duplex property of the sensor radio and the broadcast nature of wireless medium, effective data transmission in WSNs is limited by potential interference and limited bandwidth [14].Multichannel communication is a cost-effective solution to meet the higher bandwidth demands of limited bandwidth in WSNs [7], [15].
Data fidelity is the degree to which data retain its original information during transmission or processing.The fidelity of the data is usually achieved by accurate channel estimation.Channel estimation, which is the process of estimating the model parameters of a certain channel model assumed from the received data, has been used in urban traffic scene and the Industrial Internet of Things (IIoT) [7], [16].Meanwhile, a basic prerequisite for data fidelity is end-to-end reliability [17], in which packet retransmission commands from sink node to source node are required.In multihop WSNs, there is a bottleneck of inefficient downward communication from sink node to source node.
Concurrent transmission, i.e., constructive interference (CI)based flooding, has attracted intense interest since the invention of Glossy in recent years [18], [19].Ferrari et al. [18] demonstrated a successful achievement of concurrent transmission on an IEEE 802.15.4 radio module via Glossy.Turning multihop radio networks into shared network infrastructures was achieved through CI-based flooding.Inspired by the innovative flooding primitives, this article utilizes concurrent transmission to solve the bottleneck problem of downward communication in multihop WSNs.Downward transmission of packets is paved by flooding activity from sink nodes to source nodes, which piggybacks on negative acknowledge (NACK) and enables end-to-end packet retransmission.CIbased flooding also provides global synchronization in an implicit way [20].
To the authors' knowledge, we are the first to achieve a comprehensive protocol that simultaneously features optimized throughput, reliability, and low power consumption.The proposed protocol is implemented on self-made sensor nodes through the Contiki operating system.Experiments are performed on nine-node and 29-node testbeds.The protocol can provide up to 14.66-kB/s throughput with 100% data collection rate and minimum power consumption.
The rest of this article is organized as follows.Section II describes the related work.Section III outlines the protocol design.Section IV gives the theoretical analysis.Section V gives the implementation methodology.Section VI presents the evaluation, and Section VII concludes this article.

II. RELATED WORKS
CSMA-based protocols such as Mintroute [21] and CTP [22] are suitable for low-power, light data traffic load patterns.Staffetta [23] is an opportunistic data collection protocol that also capitalizes on the benefits of opportunistic routing and adaptive duty cycle.All of them are best-effort data collection protocols which do not achieve end-to-end retransmission but rely on hop-by-hop retransmission to minimize packet loss.
RACNet [24] is suitable for large-scale densely deployed sensor networks.RACNet clusters the network into spanning trees and assigns multiple channels to different spanning trees to improve throughput.It encapsulates downstream data requests within source-routed packets to enable end-to-end packet loss recovery.Its limitation is that RACNet is only suitable for densely deployed networks containing spanning tree clusters.
TDMA performs parallel transmission with collision avoidance [12].Funneling media access control (MAC) [25] is a hybrid TDMA/CSMA MAC protocol that mitigates the funneling effect that is associated with many-to-one collection in WSNs.It employs localized TDMA algorithms on a network-wide basis and overlays the funnel area near the sink node.RLPL [26] is also a hybrid MAC protocol using a typical time slot super-frame technique.However, the scheduling method in it does not fully utilize the bandwidth capacity.
The collection capability of current TDMA mechanisms varies depending on the scheduling method [27], [28].Song et al. [29] presented a localized TDMA MAC protocol called TreeMAC, where TreeMAC allocates time slots based on the required data rate of each subtree, ensuring collision avoidance within two hops.As scheduling is determined locally by the parent node, TreeMAC strives to provide topology adaptation.However, it limits the throughput of TreeMAC to one-third of the maximum collection throughput.Lee and Cho [30] proposed Tree TDMA MAC with the time slot allocation algorithm and a frequency slot allocation algorithm to solve the problems afflicting prevalent MAC protocols in tree topology networks.The method outperformed CSMA/collision avoid (CA) and TreeMAC in terms of throughput and network delay.
PIP [31] is a high-throughput multichannel TDMA protocol targeted at applications such as structural health monitoring.PIP utilizes a connection-oriented mechanism to provide reliable data collection.It uses a mechanism similar to Flush [14] to collect data with a single stream.Flow setup and tearing down can cause collection delays.In 2011, Ferrari et al. [18] implemented Glossy, a fast network flooding and time synchronization protocol based on concurrent transmissions.They delivered the essential implementation of CI based on the IEEE 802.15.4 standard [18].
Aiming to overcome the problem of packet conflicts that affect network throughput as the number of long-range radio (LoRa) deployments increases, an interference cancellation technique called concurrent interference cancellation (CIC) has been proposed by Shahid et al. [32].CIC performs concurrent decoding of multiple conflicting LoRa packets by selecting the optimal subset of symbols to eliminate the interfering symbols.Experiments show that the CIC is robust to signal-to-noise ratio variations in LoRa devices and therefore suitable for practical LoRa deployments.Likewise, Xu et al. [33] proposed a window matching scheme called Cantor.It is also an application of concurrent transmission to the LoRa standard.Cantor constructed a concurrent transmission to explore the correlation between downlink packet reception rate (PRR) and transmission parameters such as duty cycle and receiving delay, and used a regression model to derive the actual downlink PRR under different network settings.Results of the experiments show that Cantor is robust in all scenarios and can improve throughput and reduce energy consumption.Jia et al. [34] integrated interference graph estimation (IGE) with concurrent transmission and validated its feasibility in real-world scenarios experimentally.
Batta et al. [35] proposed improved-distributed randomized (I-DRAND), which aims to provide a TDMA schedule for tree-based networks.I-DRAND guarantees collision-free slot assignment for any network with message complexity.Due to the random nature, a node may lose the lottery several times before being allowed to send a request.Each node may have to go through several trials (rounds), which causes higher running time and more energy consumption.Another treebased distributed TDMA scheduling algorithm is presented by Batta et al. [36].This algorithm works in a bottom-up manner.However, the algorithm does not provide an acknowledge mechanism for request messages from child nodes to their parent; thus, they may have to send requests continuously every round until receiving a schedule from the parent, leading to more generated message overhead.
Distributed and concurrent link scheduling algorithm (DICSA) is a concurrent link scheduling algorithm [37], which introduces a primary and secondary dual-state machine, allowing each node and its neighbors to participate in each other's slot reservation process.The algorithm must maintain a common list of prohibited slots to ensure conflict-free scheduling.Although DICSA does not require time synchronization, its performance may be varied as time differences among nodes can cause slot overlaps and collisions.Frey et al. [38] have developed optimal collision/conflict-free distance-2 coloring algorithms for tree networks.First, they have proposed a sequential algorithm.In this algorithm, one node is chosen to be the root of the tree.From the root node, a depthfirst tree traversal-like algorithm is launched.Suffering from long running time, the authors improved this algorithm to a parallel one.Although it is itself collision and conflict-free, the proposed sequential algorithm requires a long running time.On the other hand, a parallel algorithm cannot be easy to implement without causing collision [39].
There has been an intense interest in research on data collection for WSNs [40], [41], [42], but most of the existing researches in recent years are based on mathematical modeling.In contrast, related works with advantageous features such as implementation on real sensor nodes and experimentation on deployed testbeds are becoming scarce.Experiments involving the implementation of protocol primitives on physical sensor nodes require heavy embedded programming and distributed debugging tasks, as is the case in this article.Moreover, existing works face the dilemma of achieving optimized throughput, reliability, and low power consumption at the same time.This article proposes a novel collection method that achieves robust and reliable high throughput by combining concurrent transmissions with multichannel TDMA.CI-based flooding provides global synchronization, which avoids time slot overlap and conflict due to time difference between nodes.In addition, flooding can piggyback on NACK, which ensures the reliability of the protocol.The heuristic scheduling algorithm ensures that the node only turns on the radio during a predetermined time slot and CI process, thereby reducing radio on-time.The algorithm achieves an approximate maximum throughput [43] with no collision of communications.

III. DESIGN OVERVIEW
Time slot scheduling is a crucial part of TDMA in multihop WSNs.In order to achieve optimized scheduling that maximizes energy efficiency while providing optimized throughput, a heuristic technique is presented in this study.A vital part of high-fidelity data collecting is reliable data collection.Reliable data collection requires end-to-end packet retransmission from source nodes to sink nodes.Moreover, this study makes use of CI-based flooding to facilitate the process of providing source nodes with packet retransmission information.

A. Heuristic Scheduling
Combining TDMA and multichannel communication is an effective method for achieving parallel data rendezvous in many-to-one WSNs.Thus, as shown in Algorithm 1, we design a heuristic TDMA scheduling algorithm specifically for manyto-one multihop network topologies.The technique is intended for networks with homogeneous data traffic, in which the rate at which sensor nodes produce data is constant.It is assumed that half-duplex single radio transceivers that comply with IEEE 802.15.4 are installed in sensor nodes.
The algorithm's key idea has two folds: 1) to schedule transmissions simultaneously along multiple branches of the tree and 2) to maximize the sink's reception of data packets across as many time slots as possible.Since the sink can only receive from one root of a subtree at most in any given time slot, we must determine which subtree should be activated.It is assumed that the sink knows the number of nodes in each subtree.Each source node maintains a buffer and its associated state, which can be either full or empty depending on whether it contains a data packet.And for ease of explanation, it is assumed that the sink's buffer is always full.Choose node i ′ unscheduled child node j, whose top-subtree contains the largest greater than zero number of bu f f er == 1 nodes.end for 29: end while The first block of Algorithm 1 from line 1 to line 9 is an initialization process.node [i].buffer = 1 means there is one packet that needs to be transmitted.T slot is initialized to be 0. The remaining lines of the algorithm are the time slot scheduling process.As shown in line 10, the scheduling process continues when there are uncollected packets in the network.In each round of scheduling, the current time slot is allocated to the corresponding node x, if there exists a node x that satisfies two conditions: 1) node [x].buffer == 0 and 2) there are at least one of its child nodes y and node y's topsubtree contains the largest number of buffer == 1 nodes.In a tree-based topology, the top-subtree of node y consists of node y and all its descendants.In this case, the present time slot is allocated to node x for receiving and to node y for transmitting, as shown in lines 12-15.The unscheduled node is implicitly aware that it should turn off the radio to improve energy efficiency in the current time slot.The other lines are reserved for managing scheduling operations per scheduling cycle.Note that for sink nodes, node [i].buffer == 0 represents each scheduling cycle.
The scheduling algorithm does not take into account the interference among the sensor nodes.In order to mitigate intrapath interference and interpath interference, we fully utilize multiple channels.The IEEE 802.15.4 standard divides its band into 16 orthogonal channels from 11th to 26th.The given scheduling method ensures that only one sensor node transmits within the same hop count, so the hop-based channel assignment strategy is an adequate solution for the interference problem.
For ease of understanding, we illustrate with the nine-node network topology shown in Fig. 1(a).Channels 25, 24, and 23 are used in this instance.We use packet A to represent sensor data from node 2, packet B to represent sensor data from node 3, packet C to represent sensor data from node 4, and so on.With a heuristic algorithm, one round of data collection consists of eight time slots.
This shows that in time slot 1, nodes 2, 4, and 7 can send their packets to the sink node.The node with the largest uncollected packet in the top-subtree will transmit the packet first.In this case, node 2 and its child node 3 have packets A and B, respectively, node 4 and its children have packets C-E, and node 7 has packets F-H in its top-subtree.Both node 4 and node 7 have the maximum number of packets to send, so the protocol randomly selects node 4 to send packet C on channel 25

B. Media Access Control
Media access uses a super-frame-based structure.The protocol divides time into super-frame, as shown in Fig. 3.Each super-frame consists of a co-channel time phase and a multichannel time domain.The common channel phase is used to transmit synchronization messages based on CI, while the multichannel time slot is used for data collection.The basic synchronization philosophy of the common channel phase is the same as Glossy [18].As shown in Fig. 3, data collection consists of several consecutive rounds of data collection during the multichannel time.The packet transmission in the data collection rounds is unicast communication, so the maximum packet size is 128 B. The length of the super-frame is governed by the interval between cycle synchronization message transmissions.Typically, the super-frame length is set to be less than 2 s.

C. Reliability
CI-based flooding begins with the sink node initiating transmission, while all source nodes are in the listening state.Initially, only the sink node is transmitting.When one or more source nodes receive the packet, they increment the forwarding counter in the packet after a brief delay and then retransmit the packet.Due to precise synchronization between nodes and minor time differences, CI-based flooding of radio waves is formed.Some nodes need multiple hops to receive the packet, so the packet needs to be sent and received ten times in a row before CI-based flooding stops.
CI-based flooding ensures efficient downward communication from the sink node to the source node.Therefore, the information of lost packets can be piggybacked on the flooding information in order to require packet retransmission from the source node to the sink node to achieve end-to-end reliability.
The structural design of the broadcast packet is shown in Fig. 4. The synchronization information occupies 10 B of each packet.Then, the remaining payload available for ACK piggyback is equal to the packet size minus 10.Although the maximum packet size can reach 128 B, CI is prone to fail when the packet is too large.Therefore, due to the limited payload, the sink node endeavors to piggyback ACK information in a more efficient way.As shown in Fig. 4, one bit is used to indicate whether a packet is received or not.
The ACK message is simply loaded in the order of node identity document (ID).That is, the first bit indicates the first round of data collection for node 2, the second bit indicates the first round of data collection for node 3, and so on.Since a super-frame consists of multiple rounds of data collection, ACK messages are also loaded in the order of the collection rounds.After receiving the flooding message based on phaselength interference, the sensor node retransmits the lost packets in the next data collection rounds.
The size of the payload required for ACK piggyback depends on different networks.As illustrated in Fig. 4, the minimum integer value n for payload is given by where r is the number of rounds of one super-frame, and N is the total number of source nodes.In fact, the packet size of CI-based flooding can be flexibly adjusted.For example, for a nine-node network with seven rounds of collection in a super-frame, each interval of CI-based flooding allows the sink node to receive 56 packets.Therefore, a flooded packet requires 56 bits (i.e., 7 B) to encapsulate the ACK of these 56 sensor packets.If node 5's packet is lost in the fourth round, the 28th bit will be set to "0." IV. THEORETICAL ANALYSIS To achieve high-throughput data collection, the scheduling policy should aim to keep the sink node in a reception state as much as possible.For uniform data traffic, we prove that the heuristic algorithm can provide a lower limit of the collection time, also called as optimized scheduling.

A. Collection Time Lower Limit
The lower limit of the whole network data collection time is determined by the routing topology.Specifically, it is up to the relationship among top-subtrees in terms of subtree size.We get the following theorem concerning lower limit collection time.
Theorem 1: The lower limit time required to collect data from all source nodes is given by where N is the total number of source nodes, M is the number of nodes in the largest top-subtree, and T is the time required for data collection from all source nodes.
Proof: Theorem 1 illustrates that regardless of the topology, collection time can be deemed as a function of N and M: 1) collection time lower limit depends on the total number of source nodes.Obviously, the time required for data collection from all source nodes is at least N T slot and 2) collection time lower limit depends on the value of M. For a top-subtree, the minimum collection time is (2m−1)T slot due to the half-duplex radio transceiver, where T slot is the one-hop transmission time of a data packet, and m is the number of nodes in a top-subtree.In the case that a top-subtree contains more than half of the source nodes N , the time required for data collection from all nodes, regardless of the topology, is at least the shortest collection time of the largest subtree, i.e., (2M − 1)T slot .

B. Scheduling Application
We demonstrate that Algorithm 1 can achieve the lower limit specified in Theorem 1.To this end, we first derive recursive formulas concerning the number of cycles and the remaining data for each subtree as well, and then, we derive recursive formulas for the number of nodes and collection time.In the deduction process, we define two time slots as one cycle.
Proposition 1: Based on Algorithm 1, a node with an empty transmission buffer in any slot can receive data from its subtree nodes with nonzero remaining data in the following cycle.
Proof: 1) At the beginning of data collection, the transmission buffer is full for all nodes except the sink node.Then, Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
in the first time slot, data transmission happens between the sink node and one of its child nodes.And the transmission buffer of this child node becomes empty in time slot 2. 2) Assume the proposition holds for a certain slot b.Any node i with an empty transmission buffer receives data from its child node j, so the transmission buffer of node j is empty at slot b + 1.In this case, we consider the transmission buffer of the child node l of node j in slot b.When the transmission buffer is empty, node l can receive data in slot b according to the assumption of induction.When the transmission buffer is not empty, node l is in a waiting state because node j is communicating in slot b.Therefore, data can be received from all children of node j in slot b + 1.Using the mathematical induction method, the proposition is proved by letting: 1) be the basis and 2) be the inductive step.
Proposition 1 demonstrates that in every time slot, data packets in nonempty subtrees will move toward the root node, ensuring that the sink node can possibly receive data packets in each time slot.Then, we give a recursive formula concerning the remaining data counts of each top-subtree.The recursive formula is derived based on Proposition 1, where each node becomes capable of transmission after one cycle has passed.
x } denote the set of the remaining data counts of top-subtrees at the beginning of a cycle c, and S c is sorted in descending order.P c j denotes the remaining data counts of an arbitrary top-subtree, where j is the child node of the sink node and x is the number of top-subtrees.
Proposition 2: Based on Algorithm 1, for any cycle c, given S c = {P c 1 , P c 2 , . . ., P c x }. 1) In the case of P c 2) In the case of P c ) Proof: Regardless of the topology, the relationship between P c 1 and P c 2 can only be divided into two scenarios: 3) and (4), respectively.
1) At the beginning of data collection, the sink node receives data from the top-subtree that contains the largest number of data counts and then receives data from the next top-subtree with the second largest number of data counts in the second slot.As a result, in cycle 2, we get S 2 = {P 1 1 − 1, P 1 2 − 1, . . ., P 1 x }. 2) Assume the proposition holds for a certain cycle c.According to Proposition 1, in the case of (3), since P c 1 ≥ P c 2 > 0, the sink node receives data from the top-subtree that contains the largest number of remaining data counts in the first slot of cycle c and then receives data from the next top-subtree with the largest number of remaining data counts in the following time slot of cycle c.As a result, in cycle c + 1, we get the sink node can receive from the subtree with the maximum remaining data in {P c 1 −1, P c 3 , . . ., P c x } in the first time slot of the cycle c + 1.Similarly, in the second slot of cycle c + 1, the sink node receives from the subtree with the most remaining data besides the node sent in the previous slot.
Using the mathematical induction method, the proposition is proved by letting: 1) be the basis and 2) be the inductive step.Similarly, it also shows that (4) holds.
Theorem 2: In case Algorithm 1 is adopted, for any topology, the collection time is where N is the total number of source nodes, and M is the number of nodes in the largest top-subtree.
Proof: 1) Equation ( 5) holds when N = 1, T (1, 1) = 1T slot ; when N = 2, T (2, 1) = 2T slot ; and when N = 2, T (2, 2) = 3T slot .2) Assume Theorem 2 establishes for any topology that has N = k, where k is a positive integer, and let denote the size of its top-subtrees sorted in descending order; that is, s 1 = M.Then, consider any topology that has The number of nodes in the largest subtree becomes M ′ = M + 1.According to Proposition 2, after one cycle, data collection time S c−1 k+2 equals S c k , i.e., for this topology T = T (k + 2, M + 1), and it can be expressed as Equation ( 6) holds for any topology that has N = k + 2 and the maximum top-subtree As for M ′ = 1, it corresponds to the case where all nodes are 1 hop away, so T (k +2, 1) = N T solt .As for M ′ = k+2, it corresponds to the case where there is only one top-subtree, so Using the mathematical induction method, Theorem 2 is proved for any topology by letting: 1) be the basis and 2) be the inductive step.
Theorem 2 illustrates that the heuristic algorithm can achieve the lowest time limit of data collection time.

V. IMPLEMENTATION
This work is realized on a sensor node designed based on the MSP430F1611 microcontroller unit (MCU) and CC2420 radio chip.The software platform is the Contiki operating system.The CC2420 hardware module is directly controlled by the MCU serial peripheral interface (SPI) interface.As a result, the MCU can read data from the CC2420 buffer at the same time as it receives the data, which greatly improves the throughput, as shown in Fig. 5.In this work, no ACK-based packet transmission is used at the link layer.
Synchronization is the most important task in the implementation.Time synchronization is critical in a TDMA-based network [44].All nodes need to be kept globally synchronized  to enable time alignment.However, the heterogeneity of the sensor nodes makes it possible for clock drift to cause synchronization errors, especially when the crystal oscillator varies too much.Synchronization errors may induce confusion by causing sensor nodes to lose synchronization.
The synchronization philosophy rests on the fact that all source nodes use the time of the sink node as a reference, while the source node calculates the difference between the reference time and the actual local time.Locally, the source node tries to predict the clock drift value at the end of the current super-frame by counting the clock drift values of the last few super-frames.In fact, in order to further minimize the effect of clock jitter to some extent, the moving average of the latest eight super-frames is taken into account when calculating the predicted clock drift value.
Since super-frame starts with the synchronization process, the synchronization error at the beginning of the super-frame is minimal.However, when entering the data collection process, the accumulated clock drift causes the synchronization error to grow larger.Therefore, clock drift needs to be compensated for in subsequent collection rounds to improve the stability of the protocol.First, we need to calculate the theoretical time difference between the current super-frame start time and the current time slot start time using the following equation: (7) where D ttd is the theoretical time difference between the start of the current super-frame and the start time of the current time slot; D ci is the duration of the common channel time phase; y is the sequence of the current collection round; f is the sequence of the current time slot; D ts is the duration of a time slot; N ts is the number of time slots per collection round; and D rp is the interval between collection rounds, preserved for the software process, as illustrated in Fig. 6.
Assuming that the clock drift rate is constant during a super-frame, the clock drift value can be calculated for each communication time slot of the data collection using the following equation: where T sst is the start time of the current time slot that is captured by MCU; T sp is the start time of the current super-frame; T cd is the predicted clock drift for the current super-frame; D sf is the duration of the current super-frame; and other variables have the same meaning as in (7).Note that T sst is a real clock rather than a theoretical value.Therefore, D ttd − (T sst − T sp ) is a very small value, namely, delta, which indicates the time error caused by software execution.Experiments show that the combination of moving average-based clock drift prediction and time slot-based clock drift compensation can limit the synchronization error to less than 1 ms, thus ensuring the stable operation of the protocol.In any case, each time slot should allow enough protection time to cope with the jitter between different nodes.As shown in Fig. 5, a time slot cycle is composed of processing time, protection time, transmit time, and redundant time.When the packet size is fixed, the transmit time hardly changes because it is determined by the CC2420 hardware design.The transmission time for a 128-B packet is about 4.5 ms.The guard time and the duration of the entire time slot are configurable.We set the duration of each time slot to 229 ticks (6.99 ms) and make the redundant time equal to the sum of the processing time and the guard time (about 1.2 ms).Thus, the maximum tolerance of time jitter is about ±1.1 ms in each time slot.

VI. EVALUATION
In this section, we conduct a number of experiments to evaluate data collection performance.To gain insight into the evaluation statistics, experiments are performed on a nine-node testbed and a 29-node testbed, respectively.Physical sensor nodes are designed and implemented with refers to the circuit schematic of Tmote Sky.Since we endeavored to achieve optimized impedance matching of the RF circuit between the antenna and CC2420 radio chip based on HFSS as well as ADS simulation, sensor nodes can provide the same performance as that of Tmote Sky.Yet, due to various batches of crystal oscillator component across past years, the oscillating period skew among them is far larger than the same batch of components.
Table I shows sensor nodes' relative clock skew in terms of tick, measured in 5-s period.Normally, the relative clock error of homogeneous sensor nodes should be within several dozens.However, experiments observed hundreds of clock skews, posing a significant challenge to the protocol implementation.Therefore, besides of quantitative experiment data, successful experiments also imply sophisticated implementation of the protocol prototype.
Experiments have shown that the use of different channel groups has a limited impact on the protocol performance.Therefore, we use channels 15, 19, 21, and 25 that have the least overlap with the 2.4-GHz Wi-Fi channels [45].
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE I PERIOD SKEW
In the nine-node testbed experiment and the 29-node testbed experiment, we set the duration of a super-frame to 422 ms for 56 time slots.Despite the challenging testbed, we observed that the protocol was able to operate safely at full speed for tens of hours.

A. Metrics
The evaluation is designed to gather statistics on data collection capacity, successful packet delivery rate, and power consumption.Therefore, key performance metrics include throughput, PRR, and radio duty cycle.Network throughput is calculated according to the following equation: where y is the collection round; p is the source node; r is the number of rounds; N is the number of source nodes; and PacketPayload is the one packet's payload that is used in computing throughput.To keep in accordance with other high throughput works, a physical layer payload is used here.t is the duration of one super-frame.f (y, p) indicates whether a packet is received or not, as given in the following formula: Note that if f (y, p) = 1 represents all collection rounds and all source nodes, then the result of ( 9) becomes a theoretical throughput upper bound.PRR is defined as the probability that a packet will be successfully transmitted to the sink node without packet retransmission and is calculated according to the following equation: where C lost is the number of total lost packets, and C sts is the number of total scheduled time slots.PRR is calculated at the sink node.Note that PRR here denotes the probability of successful packet delivery at the link layer.Actually, in the following experiments, 100% data collection has been realized by end-to-end packet retransmission.The percentage of time that the node keeps the radio on is calculated as a duty cycle  where T rx is the duration of radio listening and receiving, T tx is the duration of radio sending, and T MCU is the duration of MCU running.

B. Experiment
Experiments are performed on a nine-node testbed and a twenty-nine-node testbed, as illustrated in Figs.7 and 8, respectively.First, experiments are performed on a nine-node testbed.The sensor nodes are positioned at 3-m spacing in a 13.5 × 8. 2 m office, as illustrated in Fig. 7. Data collection lasted for more than a day with a 100% data collection rate.We extracted one hour of data from midday and one hour of data from midnight for analysis.
The received signal strength indication (RSSI) statistics collected are shown in Table II.For example, the RSSI of node 5 in Table II is −63 dBm, which means that the signal strength from node 5 to its parent node, node 4, is −63 dBm.It indicates that the midday RSSI is almost indistinguishable from the midnight RSSI.
Although there is no significant difference between the RSSI at midday and the RSSI at midnight, the protocol performance varies considerably.Fig. 9(a) displays the throughput at noon and midnight.The average throughput at midday is 12.56 kB/s (100.46 kb/s), which is 105.66% of TreeTDMA and 133.06% of TreeMAC, as illustrated in Table III, while midnight is 14.22 kB/s (121.77kb/s), which is 128.07% of TreeTDMA [30] and 161.28% of TreeMAC [29], as illustrated in Table III.The throughput at midnight approaches the upper theoretical throughput limit of 15.685 kB/s (125.48 kb/s).In addition, large fluctuations can be observed in the midday experiments.This is partly due to the fact that the RSSI measurements do not adequately reflect the signal quality.It seems that Wi-Fi interference has little effect on the RSSI values based on IEEE 802.15.4 packets.In fact, human activity at midday significantly increases Wi-Fi interference to IEEE 802.15.4,resulting in a significant degradation of collection performance.

TABLE III COMPARISON OF THROUGHPUT
The radio duty cycle represents the power consumption of the sensor node.The sensor node turns on the radio during a predetermined time slot and CI process.Immediately after the transmission is over, it turns the radio off and waits for the next time slot to arrive, thus minimizing the radio power consumption.Fig. 9(b) illustrates the duty cycle of the ninenode network.Nodes 3, 5, 6, and 9 have a duty cycle of about 10.71% because they are all at the top of the network topology.They only need to transmit local data once at the specified time slot for each round of collection.Comparatively, node 2 and node 8 need to forward packets from their child nodes; hence, their duty cycle values are almost double that of their child nodes.
The PRR statistics in terms of cumulative distribution function (cdf) are shown in Fig. 9(c).The network performance at midnight is significantly better than that at midday.At midnight, the average PRR of the nine-node network is 97.05% and all sensor nodes' PRR values are larger than 90%.In contrast, at midday, the average PRR of the nine-node network is 80.07%.In some collection rounds, the PRR of some sensor nodes may be as low as 52.12%.Fig. 9(d) gives the packet loss rate to realize the link layer packet transmission rate for each sensor node.Note that from a topological point of view, node 4 is associated with node 5 and node 6; node 7 is associated with node 8 and node 9; and node 2 is associated node 3. It is apparent that the packet loss rate is consistent with the network Link 4->1 has an impact link 5->4 and link 6->4; similarly, link 2->1 has an impact on link 3->2; and link 7->1 has an impact on link 8->7 and link 9->8.As for the differences between links 4->1, 7->1, and 2->1, the explanations are twofold: 1) the presence of random factors in the form of interference and human activities and the synchronization accuracy affects the link quality.As a result of these uncertainties, the statistical results of loss rates do not always conform to the theoretical analysis.For instance, the packet loss rate of node 4 is only 0.71% at midnight, but it increases to 32.46% at midday.
As illustrated in Fig. 8, the experiments are further conducted on a testbed consisting of 29 homemade sensor nodes randomly disposed in a 17.2 × 20 m office.The network topology is structured using an expected transmission count (Etx-)-based link estimator [21].The sensor nodes are actually from different production batches, resulting in poor homogeneity.This increases the difficulty of implementation.Despite the challenging task, a meticulous design effort was made to successfully run the network on a realistic testbed.Fig. 10(a) illustrates the throughput during 10 000 continu- ous super-frames with 100% data collection.The average throughput is 14.74 kB/s (117.97 kb/s), which is 124.07% of TreeTDMA [30] and 156.25% of TreeMAC [29], as illustrated in Table III, while the theoretical throughput limit is 15.685 kB/s (125.48 kb/s).The experiments demonstrate that the proposed protocol is capable of providing higher throughput than the existing protocols.
Corresponding cycles are shown in Fig. 10(b).The average duty cycle of the tip nodes is about 5.21%, ranging from 5.05% for node 15 to 5.32% for node 4; the average duty cycle of nodes with only one descendant node (e.g., node 13) is about 9.88%; the average duty cycle for node 6 with two descendants is 14.71%; average duty cycle for nodes with three descendant nodes (e.g., node 14) is about 19.03%; node 29 with four descendant nodes has an average duty cycle of 24.02%; node 10 with five descendant nodes has an average duty cycle of 28.71%; and node 16 with six progeny nodes has an average duty cycle of 33.49%.Obviously, the duty cycle is closely proportional to the number of descendant nodes because sensor nodes consume more power when forwarding packets from their descendant nodes.
Corresponding PRR in terms of cdf is shown in Fig. 10(c), and the packet loss rate is shown in Fig. 10(d).The average PRR of the 29-node network is 93.46%, ranging from 90.81% to 94.50%.In a similar way to the experimental analysis of the nine-node testbed, the packet loss rate is related to the topology of the network.For example, from a topological point of view, nodes 21, 25, 26, 27, and 28 are descendants of node 29.Since node 29 relays packets from all its descendants, the Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.packet loss rate of node 29 is expected to be less than the packet loss rate of its descendants, which is consistent with Fig. 10.The collected statistics validate the network perfor- mance and yield the same conclusions as the aforementioned experiments on the nine-node network.Furthermore, experiments were conducted on randomly selected channels 19, 21, 23, and 25.The experiments indicate that the differences attributed to the use of different channels are very limited.

VII. CONCLUSION
This article proposes a reliable, high-throughput, lowpower TDMA data collection protocol.There is a heuristic algorithm that is presented to implement the optimized time slot scheduling.High throughput is achieved by associating multiple channels with optimized time slot scheduling.We combine CI-based synchronization with end-to-end packet retransmission for reliable data collection.An implementation and validation of the protocol is carried out on a testbed deployed in an office.The experimental results verify that the protocol achieves a high throughput of up to 14.74 kB/s (117.97 kb/s) with low power consumption and 100% data collection rate.

14 :
Schedule T slot to i for receiving 15: Schedule T slot to j for sending 16: node[i].scheduled= T r ue 17: node[i].buffer= 1 18: node[ j].scheduled = T r ue . The details are shown in Figs.1(b) and 2. Similarly, in time slot 2, as shown in Figs.1(c) and 2, node 5 sends its packet D to node 4 on channel 24, and node 7 sends its packet F to node 1 on channel 25.In time slot 3, as shown in Figs.1(d) and 2, node 2 sends its packet A to node 1 on channel 25, and node 8 sends its packet G to node 7 on channel 24.In time slot 4, as shown in Figs.1(e) and 2, node 4 sends data D to node 1, node 3 sends its packet B to node 2 on channel 24, and node 9 sends its packet H to node 8 on channel 23.In time slot 5, as shown in Figs.1(f) and 2, node 6 sends its packet E to node 4 on channel 24, and node 7 sends its packet G to node 1 on channel 25.In time slot 6, as shown in Figs.1(g) and 2, node 2 sends its packet B to node 1 on channel 25, and node 8 sends packet H to node 7 on channel 24.In time slot 7, as shown in Figs.1(h) and 2, node 4 sends packet E to node 1 on channel 25.In time slot 8, as shown in Figs.1(i) and 2, node 7 sends packet H to node 1 on channel 25.

Fig. 9 .
Fig. 9. Nine-node network experimental statistics.(a) Throughput of nine-node network.(b) Radio duty cycle of nine-node network.(c) PRR of nine-node network.(d) Packet loss rate of nine-node network.