Exploring design methodologies for next-generation IoT sensors

The paradigm of ‘big data’ stems from an increasingly connected world through the rise of social media and business transactions. The prominence of the Internet of Things (IoT) will continue to strain the resources of the internet by a convergence of wireless sensor networks that generate massive amounts of data. It is projected that by the year 2020 there will be 20 billion devices connected to the internet with each individual having an average of 6.58 connected devices [1].

The IoT sensor backplane is increasingly expected to monitor the system under test on a real-time basis. This is true for IoT sensor solutions monitoring body area networks, safety and security solutions, industrial factory and process automation solutions, and building automation solutions to name a few. This gives rise to a new paradigm tied to the data collected by the connected devices, that of 'big data sensing.'

Big data sensing drives a rethinking of the way this data is managed. The concept of edge computing tries to address these issues by processing the data at the point where the connected device uploads the data to the network. This fails to consider the system as a whole where in addition to minimizing the amount of data on the network, the overall power consumption of the wireless sensor network needs to be minimized in order to maintain acceptable battery life. In industrial IoT solutions for example, battery life of 10 years is typically expected for the connected device. Requiring the connected device to stream data real time to the network drives resources from the end node, which reduces the battery life of the device.

A more power efficient approach would be to process the data at the point of collection. Here, the IoT sensor will have to act autonomously from the gateway, initiate data collection on statistically significant events, operate with minimal power consumption, drive efficient means of extracting data, and only initiate transfer of data under instances deemed to be statistically significant. Minimizing the occurrence of these data transfer events to the gateway reduces the amount of allocated resources by the network and leads to the most efficient solution. In this paper, we take a holistic view of the IoT sensor solution and discuss design methodologies that address the system and module requirements which enable the connected device to operate autonomously with the lowest power consumption for real-time monitoring of the system under test.

Connected Device Power Efficiency
We begin by analyzing the power efficiency of a typical connected device in an industrial application. The device wirelessly monitors flow of liquid in a typical industrial process control solution. The block diagram shown in Figure 1 consists of a low power microcontroller for processing of the data and resource management, an RF transceiver, a data logger, sensing module and an LCD display. Using a subsystem energy consumption model, the total energy consumed by the connected device is given by the following equation:

ETot = EMCU total + Esensor + Elisten + Et + Er + Esleep + Eswitch + ELCD    (Eqn 1)

where EMCU total represents the total energy consumed by the microcontroller during active and sleep modes, Esensor represents the energy consumed during sensing and ELCD is the energy consumed by the display. The overall energy consumption during RF communication is the sum of the energy required during transmission (Et ) , the energy required to receive data from the gateway or adjacent nodes (Er ) , the switching energy going from idle and active states (Eswitch ) and the listening energy and energy required to resolve anti-collision during transmissions (Elisten ). The IEEE 802.15.4 standard [2] MAC and PHY layers call for a standard CSMA (carrier sense multiple access) procedure for resolving anti-collisions. For this analysis we only consider the RF energy consumption associated with the MAC and PHY layer and do not account for additional overhead driven by the upper OSI layers of the RF protocol. In addition, we do not account for beacon events called out in the MAC layer.

Figure 1. Industrial Flow Meter block diagram (Source: Texas Instruments)

The transmit energy is required to transmit a packet of data with the associated control overhead on the radio. The overhead ensures the proper handshake between the transmitting and receiving entities. In this case we assume an overhead of 16 bytes. The transmission energy consumption is expressed as

Et = Psent x Plength x TB x Itx x N x V    (Eqn 2)

where Psent is the number of packets sent, Plength is the length of a packet in bytes, It is the current draw for the radio during transmission, TB is the time for transmitting 1 byte of data and V is the voltage of operation for the system. The values for EMCU total , Esensor , Elisten , Et , Er , Esleep , ELCD are determined by the current draw for each corresponding module and the amount of time spent during each operation. The value for Eswitch is determined by the amount of time taken by the transceiver in going from a sleep state to an active state.

Next page >>

System Performance at low RF Duty Cycles
Using this subsystem energy consumption model, we run an analysis of the industrial flow meter shown in figure 1. Typical parametric values for each of the modules are shown in table 1. We see that the current values for the transceiver (Itx , Irx ) are orders of magnitude larger compared to the current values for the microcontroller (Iactive , Iidle ) thus minimizing RF transfer of data leads to the most power efficient solution. This is illustrated in Figure 2 where we calculate the battery lifetime of a connected device as a function of RF duty cycle.

Figure 2. Simulated battery lifetime as a function of RF duty cycle. (Source: Texas Instruments)

The battery lifetime is calculated by the following equation:

Lnode lifetime = [ Cbatt x V ] / Ecomposite       (Eqn 3)

where Cbatt is the battery capacity used in the system. In this analysis we assume a Carbon-Zinc battery with an estimated battery capacity of 650 mAHr.

Table 1. Simulation Parameters used in the system analysis.

While at higher duty cycles the RF component of the system drives the overall power consumption, at lower duty cycles the other components of the system take a more prominent role. This is illustrated in Figure 2 (above) by comparing the battery lifetime of MCU1 to MCU2. In this case, MCU1 runs at a lower active and sleep current compared to MCU2 and translates to a significant improvement in the battery lifetime of the system at the lowest RF duty cycles. Thus special attention must be taking in selecting a microcontroller with the lowest active and idle currents when the system is designed to minimize communication with the network.

Autonomous Connected Device Performance
We now consider the case where the collected sensor data is stored in non-volatile memory enabling autonomous behavior by the connected device. We look at the impact of new non-volatile memory technology that significantly reduces the energy requirements required for data logging the sensor collected data.

Non-volatile ferroelectric RAM (FRAM) technology drives the rethinking of the IoT system design discussed earlier by exploiting the high capacity and highly energy efficient storage capabilities. This memory technology allows for a larger amount of sensor data to be stored locally prior to transmission to the gateway. By taking advantage of the aggregation of this larger amount of data, the overall system energy efficiency is additionally optimized by enabling a reduction of the RF duty cycle.

While Flash programming occurs through a tunneling mechanism, FRAM programming relies on a ferroelectric effect to induce polarization in a dipolar molecule. The ferroelectric effect occurs due to the electrical dipole formed by Zirconium (Zr) and Oxygen (O) atoms in the ceramic Lead-Zirkonate-Titanate (PZT) crystal of the FRAM cell. The electric field causes a polarization hysteresis effect as it moves the Zi-atom within the PZT crystal with increasing field strength. This means that it wears down far less if at all for each memory operation, and consequently lasts over one billion times longer than Flash. In addition, FRAM does not need a pre-erase cycle and the molecule polarizes in one or two nanoseconds, so the write operation is about 1000x faster than the previously mentioned nonvolatile counterparts.

The speed of FRAM is equivalent to embedded static RAM in many microcontrollers and provides dynamic accessibility and non-volatility; it is commonly referred to as a Universal Memory. This means it can function as the data memory or the program memory at any given time in its life. This gives designers the freedom to create embedded software that relies heavily on data processing or not at all depending on their specific needs without worrying about the limitations of the microcontroller. No other embedded memory can claim this feature.

The lower energy costs associated with data logging on FRAM technology enable the use of in-network storage to reduce the overall communication requirements. In this case, the power consumption of in-network storage needs to be compared to the energy needs associated with the RF communications. As shown earlier, communication costs are quite high at the higher RF duty cycles and drive the battery lifetime of the system. By taking advantage of the lower energy costs associated with in-network storage and related computational costs of FRAM, the communications energy requirements can be significantly reduced.

The overall reduction in communications energy costs can be achieved by enabling the sensor to carry out adaptive changes to its data collection based on the historical data collected over time. The sensor may choose to disregard data if it does not detect any significant changes based on the trends observed over a predefined sampling period. The system can build predictive time series models taking advantage of the low computational energy requirements associated with the algorithm [3]. These time-series models require a significant amount of historical data to build the required accuracy and can only be achieved with a low energy consumption memory technology such as FRAM to take advantage of the local storage. The reduction in energy costs comes from a reduction of the amount of data that is transferred over RF. The lower energy costs associated with FRAM also allow for aggregation of the sensor-collected data that rely on hash tables to perform duplicate packet suppression [4]. These hash tables are typically too large to be carried out in RAM. FRAM-based data management schemes can be used to store these hash tables with low energy costs, thereby improving the performance of the system.

System benefits of embedded special-function system solutions
We now consider the benefits of embedded special-function analog solutions based on analog integration on the microcontroller. Analog integration on embedded microcontrollers has resulted in unique SoC solutions that drive extremely efficient system solutions. Analog modules include analog-to-digital converters (ADCs), programmable internal reference sources, digital-to-analog converters (DACs), comparators, analog switch matrix and integrated LCD drivers. These analog modules can be used to directly sense the incoming analog signal, buffer the signal or carry out signal conditioning prior to conversion to a digital bit stream. These analog modules can also be combined to facilitate special system function solutions. These modules manage the entire system solution, which may include excitation of the sensor, data collection, signal conditioning, data logging, data processing, system management and communication with the internal host processor. Typical examples of special-function analog solutions include charge-time measuring units, various motor control solutions, RF conversion and special-function solutions to measure physical parameters such as pressure, temperature, flow and position. These SoC solutions enable the connected device to operate autonomously at the lowest possible power efficiency.

Analog integration results in better system power consumption as functionality is distributed in a single chip solution. This also leads to an increase in throughput with faster system switching times and a reduction in system noise. Inherently these mixed signal designs also lead to simplified system designs with pre-verified aspects of the component integration. Signal and power integrity conformance to device specifications is verified through specification-based functional validation. In addition, special function system solutions incorporate analog components that can be fully controlled by software through intelligent connections enabled internally by the IC. This functionality also provides for a programmable internal reference source for the analog modules. Overall the single chip solution results in a reduction of the bill of materials (BOM) and PCB board size, which translates to a significant potential system cost savings.

The future of the IoT will put a strain on resources of the cloud with a convergence of wireless sensor networks and the large amount of data generated by each connected device. The paradigm of ‘big sensor data’ calls for the connected device to operate autonomously from the cloud in order to minimize resources from the network and the connected device. Instead of having the connected device stream data to the network where it may be processed, the data will need to be processed at the point of collection. This requires design methodologies that address the system and module requirements which enable real time monitoring of the system under test, adaptive data management techniques, innovative sensor front ends that enable efficient data collection, new embedded technologies that lead to the lowest power implementation, techniques for optimizing the system architecture and innovative SoC solutions that manage the entire solution on a single chip.


[1] J. Gao, L. Lei, S. Yu; Big Data Sensing and Service: A Tutorial, 2015 IEEE First International Conference on Big Data Computing Service and Applications

[2] IEEE 802.15.4 MAC/Phy Standard for Low-Rate Wireless Personal Area Networks (LR-WPAN’s), IEEE , http://www.ieee802.or/15/pub/TG4.html, 2010

[3] P. Desnoyers, D. Ganesan, H. Li, and P. Shenoy. PRESTO: A predictive storage architecture for sensor networks. In Tenth Workshop on Hot Topics in Operating Systems (HotOS X)., June 2005

[4] J. Heidemann, F. Silva, C. Intanagonwiwat, R. Govindan, D. Estrin, and D. Ganesan. Building efficient wireless sensor networks with low-level naming. In Proceedings of the Symposium on Operating Systems Principles, pages 146-159, Banff, Alberta, Canada, October 2001. ACM.

Rafael Mena, Ph.D. , is a systems/applications manager in TI’s Microcontroller Organization and has over 20 years of experience in the design of electronic and system solutions. Rafael holds a Ph.D. in Electrical Engineering from The Ohio State University.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.