1. Towards Optimal Operation State Scheduling
in RF-Powered Internet of Things
Songyuan Li, Shibo He, Lingkun Fu, Shuo Chen, and Jiming Chen
IEEE SECON
Hong Kong
June 2018

2. Background and Motivation

3. Internet of Things IoT has brought convenience to our daily life.
Billions of IoT devices require charging service.
The Internet of Things (IoT) has grown from concept to reality in recent years. The maturity of IoT has brought great convenience to our daily life. It is estimated that the IoT industry will consist of about 30 billion objects by Most of the existing IoT sensors are powered by batteries. The limited battery capacity leads to insufficient sensor lifetime and high network operation cost.
Radio frequency (RF) power transfer is becoming a reliable solution to providing energy for low-power electronic devices. This technology enables IoT to achieve flexible and low-cost deployment, which gives rise to RF-powered IoT.
Radio frequency (RF) power transfer is becoming a reliable solution to providing energy for low-power electronic devices.
RF power transfer enables IoT to achieve flexible and low-cost deployment, which gives rise to RF-powered IoT.

4. Key front-end component
Wireless Rechargeable Sensor Networks
Charger
Sensor node
Wireless rechargeable sensor networks is the key front-end component of RF-powered IoT. WRSNs collect information of the physical world via sensors and upload data
to servers for further fusion and decision. To achieve high operation efficiency, we need to improve the network utility of WRSNs, including deployment cost, charging latency, communication delay, etc.
Key front-end component
of RF-powered IoT
Network utility optimization (deployment cost, charging latency, communication delay)

5. Can not perform simultaneously
Motivation
Can not perform simultaneously
RF wireless power
Data flow
Operation State Scheduling
In most of the previous works, data collection and energy transfer were assumed to run simultaneously, the same way as in the battery-powered sensor networks. This assumption
does not hold for most of the commercial wireless charging and sensing platforms. For example, Powercast is one of the pioneering platforms for battery-free wireless applications. To reduce the manufacture cost, it utilizes a super capacitor as an energy buffer. Such a design deprives sensor nodes of working (discharging) and harvesting
energy (recharging) simultaneously. This makes the utility optimization problem quite different from the battery-powered sensor networks. We need to schedule them to proper operation states in different time slots

6. How to schedule sensors’ operation state optimally?
Two Challenges
How to schedule sensors’ operation state optimally?
Sensing rate control
and Routing
Operation State Switching
The challenges of operation state scheduling problem come from two parts. 1. Operation state switching. Because of the battery capacity constraint, sensor nodes should switch their states according to the energy level. This makes the network topology dynamic, which leads to a temporal coupling of optimization strategies among different time slots.2. Sensing rate control and routing. In every time slot, we should jointly optimize each sensor node’s sensing rate and routing path, which results in a spatial coupling of strategies. So we find that there is a strong spatiotemporal coupling in this category of problems. We intend to solve it in this work.
Temporal Coupling
Spatial Coupling
Towards Optimal Operation State Scheduling

7. Three Main Contributions
Operation State Scheduling Problem
We first take the hardware design constraint in WRSNs into consideration, and propose a working-recharging operation state scheduling problem to maximize the network utility.
Effective Solutions to Two Cases
We provide solutions to the primal problem in two cases. An optimal one based on geometric programming for single-hop case, and an SSA algorithm based on Lyapunov optimization for multi-hop case.
Extensive Practical Experiments and Simulations
We conduct practical experiments based on Powercast wireless charging and sensing platform and simulations. The results demonstrate that the SSA algorithm performs at least 85% of the optimal.

8. System Model

9. Energy Recharging and Consumption Models
Network Settings
A wireless rechargeable sensor network with
sensor nodes.
Sensor nodes harvest RF power from fixed wireless chargers, and upload data to sink nodes.
Each sensor node select either State R for recharging or State W for working in each slot.
In our model, we consider a wireless rechargeable sensor network with N sensor nodes. Sensor nodes harvest RF power from fixed wireless chargers, and upload data to sink nodes. Each sensor node select either State R for recharging or State W for working in each slot.
For the energy recharging model, our design is based on the Powercast wireless charging and sensing platform. We utilize the Friis’ space equation here.
For the energy consumption model, we select the radio model proposed by Heinzelman. One for transmission, one for reception.
Energy Recharging and Consumption Models
Energy charging model:
Energy consumption model:

10. Battery Capacity Constraint
Data Flow Constraint
Battery Capacity Constraint
Consumed energy in time slot t :
Received energy in time slot t :
Without loss of generality, we assume a dynamic routing scheme. That is, sensor nodes can forward data to their next hops via a different route according to current conditions. Under such a condition, each sensor node should satisfy the data flow balance, that the incoming data flow equals to the outgoing data flow plus its generating data flow. We also takes the link capacity into consideration.
Another important constraint is battery capacity. In each time slot, sensor node’s consumed energy should never exceed the sum of the accumulated energy and the received energy; and the remaining energy should never exceed the battery capacity. Here, fai is the consumed energy, psi denotes the received energy, and e is the remaining energy after that time slot. Notice that, in (8), fai and psi can not exist in the same time slot.
Remaining energy after time slot t :

11. . . . Optimization Objective Problem Hardness and Insight
To maximize the network’s time average throughput while guaranteeing fairness among sensor nodes.
Problem Hardness and Insight
Spatiotemporal coupling
Sensor node state decision
Input
Output
Energy level
Topology
Input
Output
Energy level
Topology
Input
Output
Energy level
Topology
Here comes the optimization objective. We intend to maximize the network utility. In this paper, we take the network throughput as an example. We set the utility as log(ri), since it can guarantee the fairness while maximizing the time average utility. We claim that the results can be easily extended to other utility settings. Then we have the whole formulation.
Here, let us discuss the hardness and insight of this problem. At the beginning of every time slot, we need to decide which state a sensor should switch to. So we should know the energy level of each node after the last time slot. While we can only obtain the energy level information after calculating the sensing rate and routing in the last slot. Then we need to optimize the sensing rate and routing path, here we need to know the topology, but the topology information can only be obtained after state scheduling. So we can see that the two processes are coupled together.
. . .
time slot 1
time slot 2
time slot 3
Input
Output
Topology
Energy level
Input
Output
Topology
Energy level
Input
Output
Topology
Energy level
Sensing rate control and routing

12. Solution to Small-scale Networks (Single-hop Case)
Firstly, we propose an optimal analytical solution for the single-hop case, where sensor nodes transmit data directly to sink stations without relay. This case can be broadly applied to small-scale network scenarios.

13. geometric programming
Model Simplification
Local Time Allocation Problem
In the single-hop case, a sensor node trends to select the nearest sink station greedily. So constraint (4) can be simplified as (11). And we can sum Constraint (6) in different time slots together. Notice that under the single-hop setting, each sensor node does not relay data for others. Thus, the network topology is fixed, which means, there is no spatial coupling in this case. The primal problem can be separated to each sensor node’s local sensing rate optimization problem.
This problem is not convex, since ri alpha i is not a concave function. To tackle this problem, we introduce geometric programming to convert it to a convex one. Then we can use KKT condition to obtain the optimal analytical solution.
geometric programming
KKT
condition
Convex optimization

14. Solution to Large-scale Networks (Multi-hop Case)
In larger-scale networks, with the assistance of other relay nodes, sensor nodes can transmit data to sink stations over a long distance via multi-hop routing. However, the introduction of multi-hop communication makes it impossible to decouple the primal problem directly. Thus, the optimization problem is quite different from the previous one.

15. virtual energy threshold vector
The SSA Algorithm based on Lyapunov Optimization
Lyapunov function:
virtual energy threshold vector
1-slot Lyapunov drift plus penalty:
To minimize Lyapunov drift plus penalty:
To address such a problem, we propose an SSA algorithm based on Lyapunov optimization.
We first define a virtual energy threshold vector theta, and E(t) is defined as the vector of all sensor nodes’ energy queues. So we have Lyapunov function as this. The intuition of the definitions is that to keep the value of Lyapunov function small enough, we should push ei(t) to theta i(t). By further optimizing the value of theta(t), we can ensure that the energy queue is bounded and the algorithm will achieve a near optimal solution.
Then we consider 1-slot Lyapunov drift plus penalty. Through derivation, we have the decision stragegies.
The key idea of Lyapunov optimization is to minimize the drift plus penalty, so that the energy queue can be bounded and the optimization objective can be achieved. So to minimize the drift plus penalty, that is, the right side of the inequality, we obtain the following results.
1. Virtual energy threshold
2. Operation state scheduling

16. 1. Deterministic bounded energy queue:
3. Sensing rate control and routing
Convex optimization
Performance Analysis
1. Deterministic bounded energy queue:
For , we have and
We now provide the theoretical analysis of the SSA algorithm. Firstly, in both cases, we derive the bound of energy queue.
Then
For , we have and
Then

17. 2. Performance Guarantee:
T-slot Lyapunov drift with penalty:
Then we provide the performance guarantee of SSA. We first define the T-slot Lyapunov drift with penalty.
Taking a sum of the inequality above over K and dividing both sides by VKT, we have this inequality.
Finally, we have the worst-case bound for network utility.

18. Evaluation
In this section, we conduct practical testbed experiments and numerous simulations with rational coefficient settings to verify our design.

19. Testbed Experiments 1. Testbed: 250*250cm square area
We conduct field experiments based on our Powercast testbed. The system is deployed in a 250*250cm square area, with 7 wireless rechargeable sensor nodes, 6 power transmitters and 2 sink stations connecting to 2 laptops, respectively. Chargers provide wireless power for sensor nodes with 3W transmission power. Sensor nodes communicate with each other via Zigbee protocol. We run the SSA algorithm on the testbed for 60 time slots with 1s for each, and adopt the time-average network throughput as the utility.
250*250cm square area
7 Powercast sensors integrated with Zigbee
6 TX91501 chargers with 3W transmitted power
2 sink nodes connecting to 2 laptops

20. Utility of each sensor node
2. Experimental Results:
We compare the performance of SSA with the optimal solution. We can see that, as the operation time goes, the time-average network throughput tends to be stable. The SSA performs at least 85% of the optimal. We also record each sensor node’s total sensing rate. The results show that there is little difference among sensors’ total sensing rates. This indicates that our design of objective function successfully guarantees the utility fairness.
Experimental results
Utility of each sensor node
The result performs at least 85% of the optimal, which verifies that the SSA provides near optimal solution.
The utility comparison indicates that our design of objective function guarantees the utility fairness.

21. Simulations Throughput comparison in large-scale networks
We also conduct extensive simulations to further test the algorithm.
We test the algorithms under different sensor node deployment settings. SSA improves about 10% and 20% of network utility compared with two existing methods.
We also present the performance of SSA in small-scale networks. Our design still outperforms the two baseline algorithms and achieves about 85% of the optimal.
Throughput comparison in large-scale networks
Throughput comparison in small-scale networks
The SSA improves about 10% and 20% of network utility compared with two existing methods.
The SSA achieve about 85% of the optimal in various network settings.

22. Throughput vs. length of deployment area
As the scale of IoTs is increasing rapidly, the robustness of our algorithm under different network sizes is also of great interest. As shown in this figure, although the network size is changing, our design still outperforms other mechanisms. As the network size increases, the network utility decreases faster. This indicates that the wireless charging power and energy consumption are very "energy sensitive", since they are both inversely proportional to the 2nd-4th power of the transmission distance.
Throughput vs. length of deployment area
The SSA shows robustness under different network sizes.
The fast utility decreasing rate when network size is large indicates WRSN is “energy sensitive”.

23. Energy efficiency of SSA Energy efficiency of 0.5-thr
Energy efficiency of Alt
In this part, we try to explain why the SSA performs much better than the other mechanisms. The key factor that affects the network utility is the power efficiency. The sensor nodes should fully utilize the recharged energy.
Those figures compared the energy efficiencies among SSA and other methods. The result indicates that SSA achieve 89.9% energy efficiency, which is much higher than the others. This is because the state scheduling optimization makes the network operate under dynamic topology. Dynamic topology enables sensor nodes with sufficient energy to undertake more relay tasks for the energy insufficient sensor nodes nearby. Thus, all the sensor nodes can fully utilize their recharged energy. Moreover, the global optimization of energy threshold further provides near-optimal network topology in every slot.
The SSA outperforms other algorithms because of its high energy efficiency.
Dynamic topology enables sensors to fully utilize recharged energy, which leads the network to a good balance between collecting energy flow and uploading data flow.

24. Conclusion

25. Conclusion Operation State Scheduling Problem
We first take the hardware design constraint in WRSNs into consideration, and propose a working-recharging operation state scheduling problem to maximize the network utility
Effective Solutions to Two Cases
We provide solutions to two cases. An optimal one based on geometric programming for single-hop case, and an SSA algorithm based on Lyapunov optimization for multi-hop case with a proved performance guarantee.
Extensive Practical Experiments and Simulations
We conduct practical experiments based on Powercast wireless charging and sensing platform and simulations. The results demonstrate that the SSA algorithm performs at least 85% of the optimal with high robustness.

26. Thanks for your attention ! Songyuan Li songyuanli.zju@gmail.com
IEEE SECON
Hong Kong
June 2018