Node modeling and calculation of processing power consumption of wireless sensor networks with the help of neural network

Dissertation for M.Sc

Abstract

Wireless sensor network is a network consisting of many small nodes. The node receives information about the environment through sensors. The energy consumption of the nodes is usually provided by batteries, which in most cases cannot be replaced. Therefore, the consumption power of nodes is an important issue in these networks. And it is very necessary to use accurate and fast methods of calculating power consumption in the design of low power systems.  The power estimation method is divided into 4 levels: 1) system level, 2) RTL level, 3) gate level, 4) placement level. The accuracy of power calculation at the level of gate and placement is between 70 and 95%. But the problem of power calculation at these levels is the long simulation time. Power calculation at the system level has the lowest simulation time, but its accuracy is between 40 and 75%. The low accuracy at the system level and the long simulation time at the gate and placement level have made power estimation at the RTL level more important. In this thesis, the simulation is done at the RTL level and the power consumption is predicted by the macromodel function. The main power consuming components in the wireless sensor network node are simulated in SystemC, then the designed circuit is converted into synthesizable blocks in Verilog. These blocks and input sets are given to the Power Compiler software and the power consumption is calculated. In the proposed method, coefficients are calculated for different input sets and the processing power of the system is estimated. By comparing the estimated power and the calculated power, it can be seen that this method has good accuracy, but it has problems with some input sets. A neural network has been used to find the appropriate input to carry out the design and to ensure the accuracy of the estimation made.

Power in wireless sensor networks

With the advances that have occurred in the field of electronics and wireless communication, the ability to design and manufacture sensors with low power consumption, small size, reasonable price and various applications has been provided. These small sensors are able to perform actions such as receiving information from the environment, processing and sending it. All of these factors have led to the creation and expansion of networks known as wireless sensor networks [1]WSN. A sensor network consists of a large number of sensor nodes [2] that are dispersed in an environment and collect information from the environment. The location of the sensor nodes is not predetermined. And it is possible to leave them in inaccessible places. Each sensor node has a processor and performs a series of preliminary processing on the received information and then sends the data.

Although each sensor alone has little ability, the combination of hundreds of small sensors provides impressive possibilities. In fact, the popularity of wireless sensor networks is based on the use of a large number of small nodes that can be organized together and used in cases such as simultaneous routing, monitoring of environmental conditions (such as ambient temperature, the presence of gases in tunnels), monitoring of infrastructure or equipment of a system.

In these networks, unlike wired systems, the costs of configuring and arranging the network are reduced and instead of installing thousands of meters of wire, you only need to install small devices at the desired points. placed The network is simply expanded by adding a few nodes and no special design and configuration is required. Nowadays, reducing the size and weight of sensors and increasing their sensitivity is the main goal of many research laboratories and different companies. But the shrinking of the size of the sensor nodes meant the shrinking of their energy generating batteries.

Wireless sensors are often used to receive and process remote information, so the power consumption of the nodes is an important issue in wireless sensor networks. Because in wireless sensor networks, nodes must work for a long time with a certain and limited power source. The energy consumption of the nodes is usually provided by batteries, which in most cases cannot be replaced and the life of the sensor ends when the battery is exhausted. Another important issue is that batteries often make up about 50% of the volume and weight of sensors. In general, power consumption can be saved in two ways. The first way is by making sensors with less energy consumption and the second way is by using power management methods in network software design.. For example, [4] TDMA transmission is suitable in terms of power consumption, because in the interval of each slot when the information of each sensor is not transmitted, the sensor is in sleep mode, which consumes very little energy (Lewis, 2004). This method is shown in Figure 1-1.

Figure 1-1. Periodic sleep and wake-up in TDMA method (Ye & eta1, 2002,1567-1576)

The required transmission power increases proportionally to the square of the distance between the source and the destination. So a few small jumps consume less power than one big jump. If the distance between origin and destination is equal to R, the power required for a jump is proportional to If there are nodes between the origin and the destination that contain n small jumps, the power required in each node is proportional to (Lewis, 2004).

Most hardware has several working states: off, standby (no load) and on. As a result, with power management, the components are on only at a specific time. Minimizing the number of messages is also a solution. Since sending and receiving messages consumes energy, reducing the number of connections is also a way to reduce power consumption. Correct routing reduces the number of messages sent. Scheduling the work of nodes will also reduce power consumption. Or that a small number of nodes are awake to provide the required coverage. In order to balance the energy consumption, rotation is done periodically, i.e. the place of sleeping and awake nodes is changed. Increasing the efficiency [5] after expanding the nodes maximizes the lifetime of the network (Coelho&Fiore, 2005). Topology control protocols [6] can reduce the transmission power of the network by adjusting the transmission range of each node while maintaining the essential characteristics of the network. The power-aware routing protocol[7] selects the appropriate transmission path and range of each node to reduce energy consumption. Both of the aforementioned protocols reduce power consumption when the radio interface is actively transmitting/receiving packets. While this interface consumes considerable power when it is idle. Sleep management [8] is proposed to reduce the energy consumption so that the radios are turned off when they are not used (Xing&eta1, 2005, 1-30).

The sensor node consists of a processor, radio, memory and a number of sensors. These components have different consumption power, but in the meantime, processor and radio have higher consumption power. Therefore, using methods to reduce the power consumption of these components significantly increases the life of the sensor network.

There are three sources of power loss in CMOS. The total loss power is expressed according to the following equation: In the first part, CL is the load capacitor, fclk is the clock frequency and pt is the probability of loss in transmissions, which we call this part the switching power. In the second part, Isc is the short circuit current that flows from the power supply to the ground when both NMOS and PMOS are active. In the third part, Ileakage is the leakage current. The first and second parts represent the dynamic power and the third part represents the static power. Because the first part is the dominant part in dynamic power, we consider the energy consumed in CMOS processors to be the sum of switching and leakage power (Happonen, 2004). The switching energy is proportional to the square of the supply voltage. The leakage energy is modeled as follows: where Vth is the threshold voltage and VT is the thermal equivalent voltage. Reducing the cycle time increases the leakage energy, so that the leakage energy becomes the dominant factor of the total energy consumption. Using techniques such as voltage dynamic scaling [9] and turning off the no-load components reduces the energy consumption in short duty cycle mode.

The power estimation method is divided into 4 levels of abstraction: 1) system level, 2) RTL level [10], 3) gate level, 4) placement level [11] which is shown in Figure 2-1.

Figure 1-2. Levels of power estimation

The accuracy of power calculation at the level of gate and placement is between 70 and 95%. But the problem of power calculation at these levels is the long simulation time. Power calculation at the system level has the lowest simulation time, but its accuracy is between 40 and 75%. The low accuracy at the system level and the long simulation time at the gate and placement level have made power estimation at the RTL level more important. The importance of power estimation at the RTL level stems from the fact that most digital designs are RTL-based. Gate-level design analysis is possible for designs with a complexity of 50,000-10,000 gates. At the RTL level, the simulation time for power estimation has been reduced and the amount of power consumption is determined in a shorter time, which is effective on the design.