NL2023646B1 - Battery less sensor network - Google Patents

Abstract

Battery—less sensor network comprising a plurality of nodes, each node representing a sensor, provided with means to harvest energy from the environment and store said energy for later use by the respective nodes, wherein the on/off times of the nodes are randomized. Wherein the nodes are incapable of communication with other nodes in the network, randomization of the on/off times of the nodes is self—inflicted on an indi— vidual level of the nodes.

Description

Battery less sensor network Field of the invention The invention relates to a battery-less sensor net- work comprising a plurality of nodes, each node representing at least a sensor or a group of sensors, provided with means to harvest energy from the environment and store said energy for later use by the respective nodes.

Background of the invention The vision of smart cities, through the use of Inter- net of Things technologies, requires billions of sensors providing the necessary context to aid people in their daily lives. For example, cars will no longer need to wait in front of traffic lights for non-existing pedestrians to cross the road; doors, upon leaving, will provide people with the latest weather forecast; and jackets will adjust air circulation based on body temperature. Unfortunately, batteries do not provide a viable so- lution to power all the needed sensors in smart cities. Bat- teries can be bulky, for example, 32% of a 1 g sensing "back- pack" fixed on a cyborg insect is a battery [9]; hazardous; and expensive—they will deplete and require maintenance lead- ing to service disruptions and heightened subscription fees. Moreover, the raw materials for making batteries are also lim- ited. Therefore, future sensors must leave batteries behind and rely on green, perpetual energy sources. Natural energy sources such as light, vibration, and heat can power tiny sensors directly [14-16, 30]. Tiny energy harvesters, however, can only scavenge a very limited power from such energy sources [25]. Therefore, an energy-harvesting sensor operates intermittently. An intermittent sensor starts by harvesting a certain amount of energy in its buffer (i.e., a super-capacitor). Subsequently, it triggers operation which depletes the buffered energy quickly, as the power consumption rate tends to be much higher than the power accumulation rate. Once the energy is below a certain level, the sensor experi- ence a complete power-down, the cycle of charging and operat- ing repeats [8].

Intermittent devices trade off a reliable energy source (the battery) for a sustainable energy source, i.e., ambient energy. This switch to energy harvesting generates many challenges [26]. For example, preserving computation pro- gress under frequent power interrupts, enabling timely opera- tions with indeterminate power-down duration, and the fact that nodes operate intermittently. Researchers continue to investigate these challenges. For example, [2, 7, 26, 36, 37] studied the intermittent com- putation problem, which is concerned with the preservation of an application progress and data consistency under frequent power failures; [18] investigated the timely operation chal- lenge, which is concerned with data freshness after a power interrupt; [28] introduced a system design for peripherals states preservation for intermittently powered sensors; and

[47] introduced event-driven execution for the intermittent domain, which deals with input and output operations under ar- bitrarily-timed power loss. Despite significant progress achieved in the inter- mittent domain, the system availability problem has not been addressed. A monitoring sensor that has a very low probability to be available when an external event occurs is not worth de- ploying. A sensor that is capable of capturing only very short events has a limited number of potential applications. For ex- ample, a voice-controlled light-switch capable of only accept- ing short (single-word) commands has its limitations. Using "on" to turn on the lights might turn on other devices as well. Using "lights" does not allow the specification of "on" or "off". Consequently, intermittent sensors have not gained widespread adoption. For example, a voice-controlled room tem- perature interface capable of only accepting short (single- word) commands has its limitations. Using "up" to raise the temperature of the room may also cause the curtains to go up. Using "temperature" does not allow the specification of "up" or "down". As a result, intermittent sensors with such charac- teristics cannot be considered for controlling the temperature in a room or for similar applications. The invention has as objective to tackle the paradox of continuous sensing on intermittent sensors.

Summary of the invention A battery less sensor network is proposed in accord- ance with one or more features of the appended claims.

The invention proposes a new type of sensor that this specification refers to as Coalesced Intermittent Sensor (CIS). The CIS is defined as the abstraction of a group of en- ergy-aware intermittent nodes providing the collective sense of being always on.

In a first aspect of the invention the on/off times of the nodes are randomized to preserve continuous availabil- ity. Randomization may need to be enforced in two scenarios: (i) In idealized harvesting conditions (e.g. in a specialized lab), the power cycles of the energy-harvesting sensors might become correlated and thus adding more sensors would not even- tually result in continuous sensing; (ii) in favorable energy conditions, energy-harvesting sensors can lose their inherent randomization and become always on. In scenario (i) the power cycles are randomized and in scenario (ii) the response to ex- ternal events is randomized.

The randomness may be inherent, but it may be re- quired that a CIS may need to introduce artificial randomness. Doing so, however, is non-trivial as knowledge is needed about the number of (charged) nodes, which depends on a number of factors including environmental conditions regarding the power source (e.g., light or RF power intensities). This invention therefore also proposes an estimator based on local measure- ments of a node's duty cycle. The key to success is to exploit the properties (i.e. randomness) of the ambient energy source to arrive at a uniform spreading of the awake times of the in- dividual senor nodes to achieve the maximum coalesced availa- bility.

Further details of the invention Inherently the nodes are incapable of communication with other nodes in the network. It is therefore an aspect of the invention that randomization of the on/off times of the nodes is self-inflicted on an individual level of the nodes.

With no information being exchanged between intermit- tent nodes, the best a CIS can do is to uniformly distribute its node’s on-times and maintaining this distribution over time. The key observation to uniformly distribute the nodes’ on-times is to ensure that their power cycles are different. This can be achieved by forcing intermittent nodes to go into low-power-mode upon power-ups. The length of this mode is ran- domly generated in accordance with uniform distribution for each node. This will change the length of the nodes on-times and, consequently, alter their power cycles. A CIS can experience a wide range of ambient power intensities. For example, a solar-powered CIS may harvest no energy at night, modest energy from artificial light, and abundant energy from direct sunlight. Generally, we can iden- tify four different CIS powering states: + Targeted power state — These are the powering conditions that the CIS is designed for. In these conditions, the CIS nodes should operate intermittently and their on- times should be uniformly distributed. Further, the CIS's collective on-time should meet the desired availability percentage. In general, the targeted powering conditions should be near worst energy harvesting conditions to en- sure that the system is properly functioning for the ma- jority of the time.

e Under targeted power state — Ultimately, the ambient en- ergy is an uncontrollable power source, and it is not hard to imagine scenarios where a CIS will be under- powered or even comes to complete and long power down {for example, a solar CIS will come to a perpetual power down in the darkness}. In general, for under-targeted en- ergy conditions, the CIS behaviour can be considered as undefined.

* Hibernating power state — In event-based sensing scenari- os, the intermittent nodes of a CIS sleep in low-power mode waiting for an external event to wake them up. If the energy conditions are relatively higher than the tar- geted conditions, the nodes may node die and sustain their sleeping power consumption. This will cause them to synchronize their wakeups on the first incoming event and their power down as the event capturing process depletes their energy buffers quickly. Consequently, the CIS may miss the next incoming events (especially if the events happen to arrive in bursts) causing it to sense intermit- tently instead of continuously * Continuous power state — Under direct mid-noon sun even a 5 tiny solar panel can continuously power a sensor.

In such conditions, the CIS will sense continuously without the need for randomization.

Therefore, the job of a single node will be repeated N times, and instead of sending a single message to a battery-powered or tethered sink-to push the data to the internet-N identical messages will be sent which waste a lot of energy.

The inefficiencies highlighted in the Hibernating and continuous power states can be mitigated by enforcing randomization on the response of the intermittent nodes: when a node is woken up by an external event it responds to that event with a certain probability.

However, 1f the randomized response is enforced all the time, then the CIS will have a lower probability of catching events during the targeted energy conditions.

Therefore, randomization of the nodes is ena- bled only during the Hibernating power state wherein in- termittent nodes of the network sleep in low-power mode waiting for an external event to wake them up, and in a continuous power state.

Another aspect of the invention is that each individ- ual node estimates the on/off times of other nodes in the net- work based on its own measured on/off time behaviour.

The in- vention proposes to use its own duty cycle, which depends on the ambient power source.

That duty cycle can effectively be used to estimate the number of active neighbours, which in turn decides if a node should back off to avoid duplicate event detection and availability interruptions (implicit syn- chronization in favourable harvesting conditions). The on-time of an intermittent node can be measured using the built-in timers in the microcontroller.

More specifically and detailed the following implementation is proposed.

In order for an energy-harvesting intermittent node to estimate the number of its active neighbours, it needs three ingredients: (i) the total number of its neighbours,

(ii) the on/off cycles of its neighbours, and (iii) how the awake times of the sensors are distributed.

The first ingredient, it is assumed to be provided before the deployment time. For the second ingredient, it is assumed that the sensor nodes are equipped with the same ener- gy harvester, including the buffer size, and they are in a close proximity. Based on these assumptions, a node can meas- ure its own power cycle (on/off cycle) and use it to estimate neighbours’ power cycles. Figure 1 shows eight nodes’ duty cy- cles under different light intensities. In general, it can be concluded that a node duty cycle is a good metric to estimate the duty cycle of other nodes.

To calculate the off-time of an energy harvesting sensor, the incoming power (or charging rate) needs to be measured and the size of the power-buffer needs to be known. The power-buffer size is fixed and can be pre-loaded to the device memory. To measure the incoming power, it is needed to determine the amount of energy harvested (Ehar) while the de- vice is executing, and measure the time of harvesting (see Figure 5). Ehar can be determined through measuring the time in two different scenarios on a fixed load. First, the time needed to discharge the energy buffer is measured (after being fully charged) while harvesting is disabled ta (Line 4, Algo- rithm 1 depicted below). ta is a constant value that needs to be measured or calculated only once and saved into the device memory. To measure the time needed to discharge the buffer while the device is harvesting, ton, a persistent timer is needed (Line 9, Algorithm 1), as the micro-controller built-in timers are volatile: meaning, they lose their values on a pow- er failure. A simple way to have a persistent timer is to use a software non-volatile counter.

The power consumption during time measurement should be fixed to establish a linear relationship between time and energy consumption (Line 8, Algorithm 1). Using the obtained values an energy harvesting sensor can run (Lines 12-14, Algo- rithm 1) to calculate the off-time.

Algorithm 1 - off-time estimation 1: freboot {u} = u++ . power reboot counter 2: Ebuf . Size of energy bu er 3: ta . time of discharging Ebuf at load a, no harvesting 4: ti © x . # power cycles 5: 1 ® freboot (i) . 1 is a persistent variable 6: if (i mod ti) = O then 7:1 =20 8: fload{a) . set node load to a 9: toff flag = True 10: ton © tpers() . persistent infinite loop 11: end if 12: if toff flag then 13: t = ton - ta . time difference due to charg- ing 14: Ehar © (Ebuf./.ta)*delta t . harvested energy 15: Pin © Ehar ./. ton . incoming power 16: toff ® Ebuf ./. Pin 17: toff flag = False 18: end if Since, calculating the off-time requires constant load, the sensor cannot run arbitrary code during time meas- urement.

Therefore, the sensor needs to sacrifice a certain percentage of its power-ups for time measurement, Line 2-9, Algorithm 1.

Once the on-time and off-time are found the power cycle is de- termined.

For the third ingredient, ideally the energy harvest- ing sensors! awake times would have to be perfectly aligned next to each other, achieving maximum availability. However, to approximate maximum span, a form of communication for coor- dination is needed. Communication for coordination may intro- duce significant overhead and further tighten the energy budg- et for other duties. Alternatively, a probabilistic approach can be applied. Instead of trying to specify the probability distribution of the awake times of the sensors, they are en- forced to be uniformly distributed. The uniform distribution leads to maximum randomness which, in turn, leads to maximum awake time probabilistic spreading. In other words, the inven- tion breaks any correlation resulting from drawing energy from the same source, having the same characteristics, and/or being in a close proximate.

Uniform distribution can be approximated by ensuring that the power cycles of the energy harvesting sensors are slightly different. The delta-t difference between the length of the on/off cycles will cause them to shift relative to each other after each cycle. This dynamic behaviour makes the awake times of the sensors to be uniformly distributed overtime. Figure 6 illustrates the scenario of two intermittent nodes with different power cycles. Node 1 has a power cycle of 6 units of time and an on/off cycle of 1/3. Node 2 has a power cycle of 5 units of time and an on/off cycle of 1/5. Following the time axis from the left, the position of the on-time of Node 2 is shifted by 1 unit of time after each power cycle of Node 2. This implies that the on-times of the two nodes are 1/3 of the time cluster together and 2/3 of the time they are apart. If the previous scenario is extended to three or more nodes then the on-time of the resulting CIS can be described with the following formula, ton (N) = ton(N-1)+ (toff(N-1))/(toff(N-1)+ton(N-1))+ton(l); wherein N € Natural numbers and ton (N) is the on-time of a CIS with N intermittent nodes. For the initial case where N = 1 we define ton (0):= 0 and toff (0) := 1.

Detailed evaluation of the invention The invention will hereinafter be further elucidated with reference to the an exemplary embodiment and evaluation of the invention, which is not limiting as to the appended claims.

In the drawing: —-figure 1 shows the average duty cycles of eight in- termittently powered nodes for different light intensities; -figure 2 shows the number of detected events by coa- lesced intermittent command recognizer with eight intermittent 40 nodes;

-figure 3 shows capturing a burst of events without randomized response, wherein a majority of the nodes react to the first event of a burst and powers down, missing the rest of the burst; -figure 4 shows response randomization enabling a CIS to capture an entire burst of events with high capturing rates; -figure 5 shows an energy buffer discharge time when a sensor node harvests energy while executing and when it does not. Ehar is the amount of energy harvested during the dis- charge time of the device (when the device is operating); and -figure 6 shows a Coalesced Intermittent Sensor's availability is the emerging collective on-time of its inter- mittent nodes' on-times. The difference between the power cy- cles leads to a constant relative shift between the nodes duty cycles. This, in turn, causes their on-times to be uniformly distributed on the overall power cycle. The bars indicate a minimum CIS time span-CIS's nodes are overlapping- and show the maximum time span of the CIS.

Evaluation of the invention To evaluate the performance of the Coalesced Inter- mittent Sensor, several experiments were conducted in differ- ent energy conditions and with different types of events.

A prototype of a coalesced intermittent command rec- ognizer (CICR) was used: an instant of a Coalesced Intermit- tent Sensor. The CICR consists of eight battery less intermit- tent nodes. Each node is capable of performing isolated words recognition. The reason behind developing a CICR is threefold: {i) voice is a natural and convenient way for human to inter- act with miniaturized devices; {ii} demonstrating the world’s first battery less intermittently powered command recognizer, which shades light on the potential of battery less intermit- tent systems; and (iii) facilitating testing with different sensing strategies and different type of external events arri- val (i.e., regular or burst).

A CICR node consists of three main parts: a micro- phone, a microcontroller, and a harvester. MSP430RF5994 [43],

an ultra-low-power microcontroller, is used for data acquisi- tion and processing. This microcontroller has a 16-bit RISC processor running on 1 MHz, SKB of SRAM (volatile), 256KB of FRAM (non-volatile), and a 12-bit analogue to digital convert- er (ADC). It also features a Low Energy Accelerator (LEA), which offloads the main CPU for specific operations, such as FFT. For recording a PMM-3738-VM1010-R piezoelectric MEMS mi- crophone is used, which features Wake on Sound and Zero Power listening technologies [35], allowing both the microcontroller and the microphone to sleep in a low-power mode until a sound wave is detected. The microcontroller and microphone are pow- ered by a BQZ5570 solar power harvester [42] connected to an IXYS SLMD121H04L solar cell [21] and a super-capacitor of 470 UF. For debugging the Saleae logic analyzer [39] was used.

The CICR has a power interrupts immune command recog- nizer. The recognizer is capable of recognizing isolated-word type of speech. The main parts of the recognizer are explained below: Data acquisition, The Wake-on-Sound feature of the microphone triggers the data acquisition process once the energy level in the sound signal crosses a certain level. The ADC, then, samples the output of the microphone at 8 kHz. This sampling rate is sufficient to cover most of the frequency range of the human voice. To determine the end of the recording the characteris- tics of the targeted vocabulary are relied upon. In particu- lar, experimentally the minimum effective recording length was identified, which is 285 ms for the chosen set of words. By exploiting the Wake-on-Sound feature and using the minimum ef- fective recording length, the need was eliminated for an end- point detection algorithm, greatly improving the processing time and system efficiency from an energy perspective.

Feature Extraction.

Once a recording has finished, framing and data pro- cessing begins. CICR divides the digitized signal into non- overlapping frames of 256 samples (x 33 milliseconds). This size is beneficial for doing a Fast Fourier Transform and short enough for the voice-features to be considered constant inside a frame.

To extract the spectral features of a frame, CICR di- vides the frequency of interest into 12 bands (as in [20]).

The first five bands has a bandwidth of 200 Hz. The next three has a bandwidth of 300 Hz which is followed by two bands of 500 Hz. Finally, the last two bands has a 600 Hz bandwidth. This division is motivated by how the energy is concentrated in human speech [20]. Then, CICR computes the 256-point Fast Fourier Transform for each frame. The resulting feature vector contains the amount of energy concentrated in each frequency band defined earlier. This feature vector forms the basis for the words identifying process once it is normalized.

The feature vectors are normalized to minimize detec- tion errors that result from differences in the amplitude of the speech input. To normalize a feature vector, CICR computes the binary logarithm of each entry of that vector. Then it computes the mean of the resulting vector. Finally, it sub- tracts the computed mean from each entry of the resulting vec- tor.

Feature Matching.

Feature matching is achieved by computing the dis- tances between the normalized feature vectors of the recorded word and the feature vectors of the words stored during the training phase (templates).

Once the recorded word has been compared to all CICR template words, the template with the smallest distance to the recorded word is considered the correct word. However, if the smallest distance is bigger the garbage threshold which we ex- perimentally set, then the CICR will return "undefined word”.

It should be emphasized that in linear distance matching (LDM) the feature vectors of two words are compared successively, not accounting for differences in pronunciation speed. This is sufficient for our case as we are targeting isolated words and speaker dependent speech recognition type.

Power Failure Protection.

In order to preserve the progress state and to pro- tect CICR data against randomly timed power failures, the CICR program is split into 19 atomic regions. It is ensured that each of these regions requires less energy then what the ener- gy buffer can provide with a single charge. The program pro- gress state is saved in the non-volatile memory (FRAM) on the transition between these regions. This prevents the program from falling back to its starting point after each power fail- ure. Data in the non-volatile memory with Write-After-Read de- pendency is double buffered to ensure data integrity when the power supply is interrupted. Experiment setup After validating on natural light and office artifi- cial light, a testbed was designed with controllable light in- tensity for clarity and reproducible results. To this end, un- controllable light sources were blocked with a box of 60x40 cm. On the ceiling of the box, a light strip of 2.5 m was at- tached with 150 LEDs that can produce 15 different light in- tensities. On the bottom of the box, a coalesced intermittent command recognizer of 8 intermittent nodes was placed with the hardware as described above. The events in the experiments are spoken words. Different patterns of isolated words were rec- orded to emulate the arrival of bursts or individual events with varying inter-event and inter-burst timing. A Bluetooth speaker [22] was used to replay a certain record. The data were collected using logic analyzer [39] and processed on a laptop running Ubuntu 16.04 LTS.

Events detection rate The experiments show the total number of detected events and the number of uniquely detected ones with and with- out randomization. They also were conducted for a different type of events. Regular events arrival, Figure 2 shows the percentage of the total captured events and the uniquely captured ones. In this experiment the light intensity varied from 300 lux to 1400 lux and the interevent time from 1 sec to 6 sec.

The experiments show a positive correlation between light intensity and the number of detected events. In particu- lar it shows that the number of duplicated detected events rises dramatically when light intensity increases, demonstrat- ing the overpowering problem. Moreover, increasing the inter- event arrival time also surges the number of duplicated events. The reason for this phenomenon is that when the time between events increases, the intermittent nodes sleep longer in low-power mode, and this reduces the inherent randomization of the intermittent nodes and leads them to the hibernating power state.

The results also show how a CIS can achieve a much higher duty cycle than its individual intermittent nodes— Figure 1 shows that with a light intensity of 800 lux an in- termittent node is active with a duty cycle of 30% while Fig- ure 2 shows that a CIS of 8 nodes captures 100% of the unique events when the time between them is © s.

Bursty events arrival - without randomization.

Figure 3 shows the capturing behaviour of a CIS when the events arrive in a burst. A burst of four events with one second between the individual events was fired every 20 sec- onds. Each burst was repeated 10 times and for four different light intensities. The nodes sleep in low-power mode when they finish processing, waiting or the next event.

In general, the intermittent nodes react to the first event of a burst and power down shortly after missing other events in the burst.

Regular and Bursty events arrival - with randomization Regular events. Table 1 below shows the number of detected events for three different scenarios. We see that randomized response reduces duplicated events by an average of ~50%, while only marginally lowers the number of the uniquely de- tected events. The intermittent nodes were responding to events with a probability of 66% for the scenaric of 800 lux and 6 seconds arrival time and the scenario of 1400 lux and 4 seconds arrival time. However, for the highest energy level and the longest inter-event arrival time a responding proba- bility of 29% was used.

(lux, sec) (800, 6) (1400, 4) (1400, 6) randomization 205/432 236/675 223/493 no randomization 240/831 240/938 240/1802 Table 1: Randomized response reduces the number of duplicated detected events, when the CIS is overpowered, by 50% while losing only 7% of the unique events. The results are presented in the following format unique/total detected events.

Bursty events. Figure 4 shows that a CIS with randomized re- sponse spreads its resources-as compared to Figure 3 - and captures the entire burst with high probability of =95%. It also shows a positive impact of randomized response when the system is under-powered (500 lux). To randomize during bursty events (which number is known in advance), a node reacts with a certain probability on an event. This probability is differ- ent for each event since the node become active after the last recharge. In order to spread the nodes over the events, the probabilities need to increase for subsequent events, since some nodes have reacted already on previcus events, and there- fore the number of nodes still available is smaller after each event. A node reacts with a probability of 40% on the first event, with 50% on the second event, 70% on the third event and 100% on the fourth event.

Although the invention has been discussed in the foregoing with reference to an exemplary embodiment of the in- vention, the invention is not restricted to this particular embodiment which can be varied in many ways without departing from the invention. The discussed exemplary embodiment shall therefore not be used to construe the appended claims strictly in accordance therewith. On the contrary the embodiment is merely intended to explain the wording of the appended claims without intent to limit the claims to this exemplary embodi-

ment. The scope of protection of the invention shall therefore be construed in accordance with the appended claims only, wherein a possible ambiguity in the wording of the claims shall be resolved using this exemplary embodiment.

Aspects of the invention are itemized in the follow- ing section.

1. Battery less sensor network comprising a plurality of nodes, each node representing at least a sensor or group of sensors, provided with means to harvest energy from the environment and store said energy for later use by the re- spective nodes, characterized in that the on/off times of the nodes are randomized.

2. Battery less sensor network according to claim 1, wherein the nodes are incapable of communication with other nodes in the network, characterized in that randomization of the on/off times of the nodes is self-inflicted on an individ- ual level of the nodes.

3. Battery less sensor network according to claim 1 or 2, characterized in that randomization of the nodes is ena- bled only during the Hibernating power state wherein in- termittent nodes of the network sleep in low-power mode waiting for an external event to wake them up, as well as in a continuous power state.

4. Battery less sensor network according to any one of claims 1 - 3, characterized in that each individual node esti- mates the on/off times of other nodes in the network based on its own measured on/off time behaviour.

5. Battery less sensor network according to claim 4, charac- terized in that in the randomization of its own on/off time behaviour each node takes into account the total num- ber of nodes in the network.

6. Battery less sensor network according to any one of claims 1 - 5, characterized in that the power cycles during which the nodes receive energy are different from each other.

7. Battery less sensor network according to any one of claims 1 - 6, characterized in that during use intermittent nodes are forced to go into low-power-mode when energy to power the node is available.

8. Battery less sensor network according to claim 7, charac- terized in that a duration of the low-power mode is ran- domly chosen for each node.

REFERENCES

[1] Domenico Balsamo, Alex S Weddell, Anup Das, Alberto Rodriguez Arrecla, Davide Brunelli, Bashir M Al-Hashimi, Geoff V Merrett, and Luca Benini. 2016. Hibernus++: a self-calibrating and adaptive system for transiently- powered embedded devices. IEEE Transactions on Computer- Aided Design of Integrated Circuits and Systems 35, 12 (2016), 1968-1980.

[2] Domenico Balsamo, Alex S. Weddell, Geoff V. Merrett, Bashir M. Al-Hashimi, Davide Brunelli, and Luca Benini.

2015. Hibernus: Sustaining Computation During Intermit- tent Supply for Energy-harvesting Systems. IEEE Embedded Syst. Lett. 7, 1 (March 2015), 15-18.

[3] C. Bernal-Ruiz, F.E. Garcia-Tapias, B. Martin-del Brio, A. Bono-Nuez, and N.J. Medrano-Marques. 2005. Microcon- troller Implementation of a Voice Command Recognition System for Human-Machine Interface in Embedded Systems. 2005 IEEE Conference on Emerging Technologies and Facto- ry Automation 1 (2005), 587-591. https://doi.org/10.1109/ETFA.2005.1612576

[4] Naveed Anwar Bhatti and Luca Mottola. 2017. HarvOS: Ef- ficient code instrumentation for transiently-powered em- bedded sensing. In Information Processing in Sensor Net- works (IPSN), 2017 16th ACM/IEEE International Confer- ence on. IEEE, 209-220.

[5] Michael Buettner, Ben Greenstein, and David Wetherall.

2011. Dewdrop: an Energy-aware Runtime for Computational RFID. In Proc. NSDI. USENIX, Boston, MA, USA, 197-210.

[6] Alexei Colin, Graham Harvey, Brandon Lucia, and Alanson P Sample. 2016. An energy-interference-free hardware- software debugger for intermittent energy harvesting systems. ACM SIGPLAN Notices 51, 4 (2016), 577-589.

[7] Alexei Colin and Brandon Lucia. 2016. Chain: tasks and channels for reliable intermittent programs. ACM SIGPLAN Notices 51, 10 (2016), 514-530.

[8] Alexei Colin, Emily Ruppel, and Brandon Lucia. 2018. A reconfigurable energy storage architecture for energy- harvesting devices. In ACM SIGPLAN Notices, Vol. 53. ACM, 767-781.

[2] Denis C Daly, Patrick P Mercier, Manish Bhardwaj, Alice L Stone, Zane N Aldworth, Thomas L Daniel, Joel Voldman, John G Hildebrand, and Anantha P Chandrakasan. 2010. A pulsed UWB receiver SoC for insect motion control. IEEE Journal of solid-state circuits 45, 1 (2010), 153-166.

[10] Brian Delaney, Nikil Jayant, Mat Hans, Tajana Simunic, and Andrea Acquaviva. 2002. A low-power, fixed-point, front-end feature extraction for a distributed speech recognition system. In Acoustics, Speech, and Signal Processing (ICASSP), 2002 IEEE International Conference on, Vol. 1, IEEE, I—793.

[11] Brian Delaney, N Jayant, and T Simunic. 2005. Energy- aware distributed speech recognition for wireless mobile devices. IEEE design & test of computers 22, 1 (2005), 39-49,

[12] Santosh K Gaikwad, Bharti W Gawali, and Pravin Yannawar.

2010. A review on speech recognition technique. Interna- tional Journal of Computer Applications 10, 3 (2010), 16-24.

[13] Graham Gobieski, Brandon Lucia, and Nathan Beckmann.

2019. Intelligence Beyond the Edge: Inference on Inter- mittent Embedded Systems. In Proceedings of the Twenty- Fourth International Conference on Architectural Support for Programming Languages and Operating Systems. ACM, 199-213.

[14] Shyamnath Gollakota, Matthew S Reynolds, Joshua R Smith, and David JWetherall. 2014. The emergence of RF-powered computing. Computer 47, 1 (2014), 32-39.

[15] Maria Gorlatova, Peter Kinget, Ioannis Kymissis, Dan Rubenstein, Xiaodong Wang, and Gil Zussman. 2010. Energy harvesting active networked tags (En-HANTs) for ubiqui- tous object networking. IEEE Wireless Communications 1%, 6 (2010), 18-25.

[16] Maria Gorlatova, John Sarik, Guy Grebla, Mina Cong, Ioannis Kymissis, and Gil Zussman. 2014. Movers and shakers: Kinetic energy harvesting for the internet of things. In ACM SIGMETRICS Performance Evaluation Review, Vol. 42. ACM, 407-419.

[17] Josiah Hester and Jacob Sorber. 2017. Flicker: Rapid Prototyping for the Batteryless Internet-of-Things. In Proc. SenSys. ACM, Delft, The Netherlands.

[18] Josiah Hester, Kevin Storer, and Jacob Sorber. 2017. Timely Execution on Intermittently Powered Batteryless Sensors. In Proceedings of the 15th ACM Conference on Embedded Network Sensor Systems. ACM, 17.

[198] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. 2012. Deep neural networks for acoustic modelling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine 29, 6 (2012), 82-97,

[20] Greg Hopper and Reza Adhami. 1992. An FFT-based Speech Recognition System. Journal of the Franklin Institute 329, 3 (1992), 555-562.

[21] IXYS Corporation. 2018. IXOLAR™ High Efficiency SLMD121H04L Solar Module. http://ixapps.ixys.com/DataSheet/SLMD121H04L Novl6.pdf

[22] JBL. 2019. JBL Go+ bluetooth speaker.

https://www.jbl.com/

[23] Frederick Jelinek. 1997. Statistical methods for speech recognition. MIT press.

[24] Jiangtao Li, Angli Liu, Guobin Shen, Liqun Li, Chao Sun, and Feng Zhao. 2015. Retro-VLC: enabling battery-free duplex visible light communication for mobile and IoT applications. In Proceedings of the 16th International Workshop on Mobile Computing Systems and Applications. ACM, 21-26.

[25] Vincent Liu, Aaron Parks, Vamsi Talla, Shyamnath Gollakota, David Wetherall, and Joshua R Smith. 2013. Ambient backscatter: wireless communication out of thin air. In ACM SIGCOMM Computer Communication Review, Vol.

43. ACM, 39-50.

[26] Brandon Lucia, Vignesh Balaji, Alexei Colin, Kiwan Maeng, and Emily Ruppel. 2017. Intermittent Computing: Challenges and Opportunities. In LIPIcs-Leibniz Interna- tional Proceedings in Informatics, Vol. 71. Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik.

[27] Brandon Lucia and Benjamin Ransford. 2015. A simpler, safer programming and execution model for intermittent systems. ACM SIGPLAN Notices 50, 6 (2015), 575-585.

[28] Kiwan Maeng and Brandon Lucia. 2019. Supporting Peripherals in Intermittent Systems with Just-In-Time Checkpoints. In Proc. PLDI. ACM, Phoenix, AZ, USA.

[29] Amjad Yousef Majid, Michel Jansen, Guillermo Ortas Delgado, Kasim Sinan Yildirim, and Przemystaw Pawetczak.

2019. Multi-hop Backscatter Tag-to-Tag Networks. arXiv preprint arXiv:1901.10274 (2019).

[30] Robert Margolies, Guy Grebla, Tingjun Chen, Dan Rubenstein, and Gil Zussman. 2016. Panda: Neighbor dis- covery on a power harvesting budget. IEEE Journal on Se- lected Areas in Communications 34, 12 (2016), 3606-3619.

[31] Monsoon Solutions Inc. 2018. High Voltage Power Monitor. htipe://www, mscon. com/hvpm-product-documentation

[32] Saman Naderiparizi, Aaron N. Parks, Zerina Kapetanovic, Benjamin Ransford, and Joshua R. Smith. 2015. WISPCam: A Battery-Free RFID Camera. In Proc. RFID. IEEE, San Die- go, CA, USA, 166-173.

[33] Michael Price, James Glass, and Anantha P Chandrakasan.

2018. A Low-Power Speech Recognizer and Voice Activity Detector Using Deep Neural Networks. IEEE Journal of Solid-State Circuits 53, 1 (2018), 66-75.

[34] Michael Price, James R Glass, and Anantha P Chandrakasan. 2015. A 6 mW, 5, 000-Word Real-Time Speech Recognizer Using WFST Models. J. Solid-State Circuits 50, 1 (2015), 102-112.

[35] pui audio, vesper. 2019, PMM-3738-VM1010-R: A ZeroPower Listening piezoelectric MEMS microphone. http://www.puiaudio.com/pdf/PMM-3738-VM1010-R.pdf

[36] Benjamin Ransford and Brandon Lucia. 2014. Nonvolatile memory is a broken time machine. In Proc. MSPC. ACM, Edinburgh, United Kingdom.

[37] Benjamin Ransford, Jacob Sorber, and Kevin Fu. 2012. Mementos: System support for long-running computation on RFID-scale devices. In Proc. ASPLOS. ACM, Newport Beach, CA, USA, 159-170.

[38] VS Rao. 2017. Ambient-Energy Powered Multi-Hop Internet of Things. (2017).

[39] Saleae. 2017. Logic 16 Analyzer. http://www.saleae.com. Last accessed: Jul. 28, 2017.

[40] Joshua R. Smith, Alanson P. Sample, Pauline S. Powledge, Sumit Roy, and Alexander Mamishev. 2006. A Wirelessly- Powered Platform for Sensing and Computation. In Proc. UbiComp. ACM, Orange County, CA, USA.

[41] Vamsi Talla, Bryce Kellogg, Shyamnath Gollakota, and Joshua R Smith. 2017. Battery-free cellphone. Proceed- ings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 1, 2 (2017), 25.

[42] Texas Instruments. 2018. Ultra Low Power Management IC, Boost Charger Nanopowered Buck Converter Evaluation Mod- ule. http://www.ti.com/too1/BQ025570EVM-206

[43] Texas Instruments. 2019. MSP430FR5994 16MHz Ultra-Low- Power Microcontroller Product Page. http: //www.ti.com/product/MSP430FR5994

[44] Joel Van DerWoude and Matthew Hicks. 2016. Intermittent Computation without Hardware Support or Programmer In- tervention. In OSDI. 17-32.

[45] Joel Van DerWoude and MatthewHicks. 2016. Intermittent Computation Without Hardware Support or Programmer In- tervention. In Proc. OSDI. ACM, 17-32.

[46] T.K. Vintsyuk. 1972. Speech discrimination by dynamic programming. Cybernetics 4, 1 (1972), 52-57.

[47] Kasim Sinan Yildirim, Amjad Yousef Majid, Dimitris Patoukas, Koen Schaper, Przemyslaw Pawelczak, and Josiah Hester. 2018. Ink: Reactive kernel for tiny battery less sensors. In Proceedings of the 16th ACM Conference on Embedded Networked Sensor Systems. ACM, 41-53.

[48] Hong Zhang, Jeremy Gummeson, Benjamin Ransford, and Kevin Fu. 2011. Moo: A Batteryless Computational RFID and Sensing Platform. Technical Report UMCS- 2011-020. UMass Amherst.

[49] Pengyu Zhang, Deepak Ganesan, and Boyan Lu. 2013. Quarkos: Pushing the operating limits of micro-powered sensors. In Proceedings of the 14th USENIX conference on Hot Topics in Operating Systems. USENIX Association, 7-7,

[50] Yi Zhao, Joshua R Smith, and Alanson Sample. 2015. NFC-WISP: A sensing and computationally enhanced near- field RFID platform. In RFID (RFID), 2015 IEEE Interna- tional Conference on. IEEE, 174-181.

Claims (8)

CONCLUSIONS A battery-free sensor network comprising a plurality of nodes, each node representing at least one sensor or group of sensors, provided with means for harvesting energy from the environment and storing said energy for later use by the respective nodes, with the feature that the on / off times of the nodes are randomized. The battery-free sensor network according to claim 1, wherein the nodes are not able to communicate with other nodes in the network, characterized in that randomization of the on / off times of the nodes is autonomous at a given time. level of the nodes, each separately. Battery-less sensor network according to claim 1 or 2, characterized in that randomization of the nodes is only possible during the sleep mode in which periodically active nodes of the network are in sleep mode in a low power mode while waiting for an external event for activation, as well as in a continuous power state. Battery-less sensor network according to any one of claims 1 to 3, characterized in that each individual node estimates the on / off times of other nodes in the network based on its own measured on / off time behavior. Battery-free sensor network according to claim 4, characterized in that in the randomization of its own on / off time behavior, each node takes into account the total number of nodes in the network. Battery-less sensor network according to any one of claims 1 to 5, characterized in that the power cycles during which the nodes receive energy differ from each other. A battery-free sensor network according to any one of claims 1 to 6, characterized in that periodically activated nodes during use are forced to switch to the low power mode when energy to power the node is available. Battery-less sensor network according to claim 7, characterized in that a duration of the low-power mode is arbitrarily selected for each node.