ABSTRACT Collecting High-Rate Data Over Low-Rate Sensor Network Radios
Collecting High-Rate Data
Over Low-Rate Sensor Network Radios Ben Greenstein?Alex Pesterev Christopher Mar Eddie Kohler
Jack Judy Shahin Farshchi Deborah Estrin
University of California,Los Angeles
{ben,cemar,kohler,destrin}@https://www.wendangku.net/doc/7b17009981.html,,alex.p@https://www.wendangku.net/doc/7b17009981.html,,{jjudy,shahin}@https://www.wendangku.net/doc/7b17009981.html,
A BSTRACT
Embedded systems can already capture data produced at high
rates,and embedded CPU and sensor performance are still
rapidly improving.Radio technology,however,can not keep
pace,and will not in the future due to known physical lim-
its of shared communication channels.This leads to a fun-
damental gap between the data a sensor network node can
collect and the data it can transmit back for analysis.
VanGo,our software system for data collection,uses?ex-
ible transcoding to narrow this gap.To make effective use of
channel bandwidth,data reduction software must run on sen-
sor nodes.However,to calibrate how data reduction software
should run,that same software should be capable of running
on the back end on real data received from the network.In
VanGo,users decide where data processing occurs.
To show that transcoding helps,we evaluate two radi-
cally different applications:acoustic collection and the mea-
surement of neural activity.Among our?ndings is that in
bandwidth-limited environments,proactive?ltering of some
of our signal can result in collecting three times the signal
energy than we could by removing silent periods alone.
1I NTRODUCTION
Sensor networks are envisioned to provide long-lived,in-
expensive,unobtrusive and untethered collection of sensor
data(e.g.,near bird nests or on rodents’backs)[22].Low-
power platforms with limited computational and communi-
cation facilities have been developed to this end[12,14,21].
Although even the?rst deployed applications sampled
the environment at much greater spatial resolutions than pre-
viously possible,they sampled at very low rates(on the or-
der of seconds or minutes between samples)[2,18,19].
While these applications proved that it was possible to use
very simple devices to collect and transfer data over multi-
hop wireless networks,they did not test the responsiveness
and robustness of these networks in the face of intense traf?c
bursts and congestion.Moreover,for the sake of simplicity
and rapid deployment and at the urging of the their users,
these deployments[4]tend not to process data in-network,
instead opting to collect all data samples to a resource-rich
fer data processing to the sensor nodes themselves,to return as much interesting information as possible given our band-width limitations.
To tune our network,we should have exactly the same ?lters and classi?ers present on the backend as we have on our wireless sensor nodes.When we?nd the right param-eters for our system–the pass-band and appropriate order for a?nite impulse response?lter,for example–we should have a mechanism to transfer the responsibility of?ltering from the back end to deployed nodes.When our network conditions change,responsibility may transfer back to the backend,so that we may calibrate for operation under new constraints.This process may be iterative over the course of an application’s lifetime.
This paper makes three contributions:We have devel-oped the?rst acoustic collection system for motes that we know of and the?rst neural spike acquisition application capable of supporting a network of more than two nodes. Second,we have designed the software system on which these applications run.This system includes a processing li-brary for data measirement,classi?cation,?ltering and com-pression;a fast and low-jitter data acquisition system for resource-constrained motes;and a mechanism to activate and control mote and backend processing of signals.
Finally,we demonstrate through experiments the?delity tradeoffs in bandwidth limited networks.We show that on-line calibration of our processing algorithms can dramati-cally improve the yield of interesting data that is collected; and that in congested networks we should?lter our data dif-ferently than in unconstrained environments.
2A PPLICATIONS AND P LATFORM
Two target applications motivated our work,each of which collects samples data at a high rate.The Auricle system re-ports acoustic data,while the Neuromote system[5]detects and reports neural spikes generated,for example,by a rat’s neurons.Both of these applications are instances of the same TinyOS-based[16]software system,which we call VanGo.
Mote-like devices,including the relatively advanced TelosB mote we use,can’t continuously transmit at the rates these applications require.Furthermore,different deployments of either system might require different data compression strate-gies.In Neuromote,for example,different types of signals vary by an order of magnitude in frequency;the compres-sion strategy must allow the network user to choose which signal is most interesting.Nevertheless,the advantages of
a mote form factor,including size,cost,deployment?exi-bility,and low environmental impact,remain important for these applications.Our work shows how to bridge the gap:
a design and software structure that lets motes collect high-rate data,even given low-rate radios.This section describes our hardware platform and applications in more depth.2.1Sensor Platform
VanGo applications use networks of TelosB motes[21].TelosB is a state of the art platform for untethered low-power data acquisition.Each TelosB node has10KB of RAM,a250kbps radio,and a16-bit MCU running at8Mhz with no hardware divide or?oating point support.Its12-bit ADC is theoreti-cally capable of generating200kilosamples per second,but practical sampling rates are limited to about21kHz due to insuf?cient communication bandwidth.Even that presumes that data has no accompaning header information,the link protocol has no header,and that the channel is perfect and available all the time;in practice,it is dif?cult to support raw data transmission at rates above2kHz.For reasons of cost and power consumption,TelosB motes lack application-speci?c DSPs that could compress raw data in sophisticated ways.
2.2Auricle
Acoustics have heretofore been used by mote-grade plat-forms solely for the purpose of target localization;acous-tic waveforms have always been discarded after being mea-sured but before transmission.We are losing interesting data. For example,macroscopic acoustic observations,enabled by dense deployments of untethered and unobtrusive sensor nodes, could provide scientists with a deeper understanding of wild-life interactions;proposed projects for acoustic collection systems include monitoring West Coast acorn woodpeckers and marmots.Equally important applications exist in other settings,such as building monitoring and smart spaces.
Auricle goes far beyond our prior acoustic sensor sys-tems[26,28]:it actually collects acoustic data from a net-work of motes,rather than using that data to trigger.Each deployed node consists of a TelosB mote connected to an ampli?er and microphone.Nodes are tasked by,and send processed data back to,a Linux-based microserver.In gen-eral,Auricle brings the advantages of mote-based sensors—coverage and scale,low power consumption,low cost,and easy deployment—to audio monitoring.
This system is designed to collect raw acoustic wave-forms sampled at more than8kHz.(For comparison,the normal range of adult human hearing is approximately20Hz to16kHz.)
2.3Neuromote
Electrophysiological recording is a powerful tool for inves-tigating the mechanisms by which the brain creates and in-terprets signals.Recordings can help neuroscientists under-stand the brain function that accompanies emotions,such as fear and aggression,and diseases,such as epilepsy and Parkinson’s disease.Neural signals on interest range from EEG(on the order of10Hz)to fast ripples(on the order of 100Hz),which are an indication of the activity of a popula-tion of neurons[1].Single units,otherwise known as spikes, are waveforms with a period of a couple miliseconds that
represent the ion discharge of a single neuron,which nor-mally occur at a rate of6-10Hz[29].To detect neuron spikes, however,a sampling rate of at least2kHz is necessary.
Existing wired electrophysiological techniques prohibit the study of freely behaving and interacting test subjects in an enriched natural and social environment,due to the teth-ering caused by wires and harnesses.Thus,the Neuromote application,which runs on wireless TelosB sensor nodes in-terfaced with test subjects(such as rats)via implanted depth electrodes and preampli?er circuitry.A Neuromote attach-ment restricts rat movement and behavior far less than a wired tethering.
2.4Discussion
Any system that collects and transmits data at such high rates will clearly use a lot of energy.Energy-limited deploy-ments should sample at high rates only occasionally—with duty cycling,say,or triggered by some other event.We note, however,that Neuromote and similar deployments are not energy-limited as the term is conventionally understood.The mote’s small form factor and portability are important for Neuromote,but long-term disconnected operation is not.
3D ATA P ATH
Sampling at Auricle or Neuromote rates stretches the lim-its of current mote hardware and software technology,even assuming limited one-hop transmission and no congestion. Conventional interrupt-driven sampling breaks down at high rates,introducing jitter and preventing other processing,so we built a direct memory access driver for TelosB’s analog-to-digital converter.The software stack must support both interactive experimentation and long-term data collection, and both uncompressed sample formats convenient for ma-nipulation and compressed formats designed for transmis-sion.We designed and implemented VanGo,a new stack that supports both Auricle and Neuromote.Its basic abstraction, the sample set,ef?ciently supports sample processing and application-speci?c format extensions.Each deployed con-?guration looks like a single?lter chain,signi?cantly sim-plifying control messages(and therefore interactive experi-mentation)and simplifying the construction of new?lters. Figure1describes a typical VanGo software con?guration; to support tuning and experimentation,the single?lter chain spans across two platforms.
3.1Data Acquisition
The interrupt load of a sensor network application that inter-acts with an ADC,radio,and timers will induce signi?cant sampling jitter.Furthermore,the interrupt load produced by an ADC operating at a high rate will overload a system,pre-venting it from doing much else.To collect high-rate data, therefore,we make use of the DMA controller packaged with the MSP430MCU on TelosB.We have written a driver for this DMA and modi?ed existing TinyOS ADC code to use the DMA to coordinate the transfer of samples from
Figure1—Data and control paths through a VanGo application,including code running on a TelosB mote and on a Linux-based microserver.
ADC conversion registers to sequential words in RAM.As opposed to generating an interrupt after each sample con-version,the DMA generates an interrupt each time it?lls a RAM buffer with data.In order to minimize the latency in providing the DMA with a new buffer to?ll,and hence to decrease the probability that samples are missed while set-ting up the next buffer,we automatically prefetch a spare buffer that will be available the instant it is requested.
The sampler operates asynchronously,reacting to the the DMA’s interrupts.We place a queue between the sampler and the rest of the system to isolate this asynchrony.Our queues are designed to work with dynamically allocated buff-ers;internally,they maintain a circular buffer of pointers to such buffers.To avoid polling by those wishing to dequeue, the queue signals when the queue becomes non-empty.
In practice,we run the data acquisition subsystem at rates up to10kHz.If we do not intend to process or transmit this data,our acquisition subsystem can sustain rates of up to 115kHz.
For Auricle,the data is generated using a low voltage microphone preampli?er(SSM2167)from Analog Devices powered by TelosB and an omnidirectional condenser mi-crophone(WM-61A)from Panasonic,as suggested by a ref-
erence design from Moteiv.To maintain a constant signal to reference voltage ratio(as our batteries drop in voltage),we use this Vcc as our ADC’s reference voltage as well.For Neuromote,we programmed a Hewlett Packard33120A ar-bitrary waveform generator to output a prerecorded neural dataset.The output of the signal generator is AC-coupled to the input of a neural preamplifer ciruit,which is an Analog Devices AD627instrumentation ampli?er with its gain set to 200.The neural signals are acquired differentially from the signal generator(as they would be from a live subject),and the ampli?ed output is referenced to half the battery voltage via a buffered(Texas Instruments OPA234)voltage divider circuit.The DC-referenced output is applied directly to the ADC input of the TelosB mote,while the ampli?er ground is shared with the mote ground.
3.2Sample Sets
At the core of our system lies the sample set data struc-ture,which contains sensor data and metadata.Sample sets are dynamically allocated from a memory pool,and travel through the system much as messages do.Each sample set consists of one metadata buffer and one linked data buffer. The DMA engine described above writes its data directly into a sample set’s data buffer.The metadata buffer con-tains?xed slots for commonly-needed information,such as modality,channel,rate,and time,as well as an extensible scratchpad containing type-length-value tuples used by?l-ters.For example,our spike detection algorithm will anno-tate a sample set with the time,width and height of spikes found in the raw waveform,while the ADPCM codec adds its state to the scratchpad.The sample sets used in our Auri-cle application have372-byte sample sets,68bytes of which is allocated to?xed metadata?elds and the scratchpad.The resulting structure offers a tradeoff between?exibility(the scratchpad)and space ef?ciency(the?xed metadata area).
Functions are also available for marshalling and unmar-shalling sample sets into packet format,facilitating commu-nication with microservers and other PC-class hardware.To make more ef?cient use of bandwidth,the sample set con-tains a mask describing which?elds to send.Sampling and processing components coordinate with the marshalling ser-vice to ensure that sample set buffers are allocated to best ?t into packets.For example,a system with a2to1com-pression algorithm will,at runtime,determine that raw data buffers can be twice as large as maximum space available for data in a packet’s payload.
3.3Filters
The VanGo data path is designed to support?exible signal processing.For example,signal processing elements can at-tach analysis results to a sample set’s metadata,making it available for later elements to process.In practice,we re-strict VanGo-level signal processing to a single linear chain of?lters,each of which transforms a single input sample set into a single output sample set.This reduces processing gen-erality,but has several important advantages,particularly for motes.Filters are easy to compose.Performance of a linear ?lter chain is relatively easy to analyze,which is particularly helpful on hard-to-debug sensor nodes.Control messages are easily directed to the right?lter,since each?lter appears at most once in the chain.[9]
Not all signal processing algorithms should run as VanGo ?lters on motes,of course.Many of even the simplest signal processing functions are too CPU-hungry to function in real time on an8MHz16-bit processor with no hardware di-vide or?oating point support.For example,consider Fast Fourier Transform(FFT),one of the most basic transfor-mations in signal processing.When optimized for speed,a FFT[24]over a512-sample window(0.064seconds of data at8kHz runs in0.5seconds on TelosB.Furthermore,these algorithms typically use large lookup tables(on the order of2KB)for computing sine.We must seek even simpler processing elements that execute quickly,and be willing to trade off some accuracy.This accuracy loss adds,of course, to the pressure to adjust processing elements based on de-ployment needs:recovering the maximum total signal en-ergy from an event requires different compression and gat-ing steps than recovering the most signal from the closest sensor to the event(without worrying about other sensors). Section5demonstrates this in practice.
We have found that the basic signal processing tasks in our applications can be completed sequentially,where pro-cessing functions transform a single input to a single output. Structuring the datapath as a linear chain restricts processing generality,but simpli?es composition.
Our?lter elements—classi?ers,compression algorithms, and measurements—operate in the time domain directly on raw waveform data,as represented by sample set structures. The Filter interface type is used to pass sample sets from ?lter to?lter.Each?lter has an input and an output inter-face of this type;the Sampler component has a Filter out-put,and the Packetizer component(the marshaller)has a Filter input.A data path devoid of signal processing func-tionality appears as like this:
Sampler[Filter]->Packetizer;
(The syntax is from SNACK[10].)Filters are added,in an application-speci?c order,between Sampler and Packet-izer.For example,to add ADPCM compression,we write: Sampler[Filter]->Adpcm->Packetizer;
The data processing path of Auricle is:
Sampler[Filter]->Stats
->AmplitudeGate
->FrequencyGate
->Adpcm
->Packetizer;
The elements named in such a wiring are the processing elements that will be built into an application.A runtime control mechanism may be used to dynamically enable and disable them individually.
We have written?lters to analyze and annotate sample sets,to classify sample sets as being worthy of transmis-sion,and to compress sample sets into more parsimonious formats.The execution times of the?lters presented in this paper are summarized in Table1.
Measurement Filters These?lters analyze sample sets, annotating each set with its statistics(using the set’s annota-tion scratchpad).This accomplishes several https://www.wendangku.net/doc/7b17009981.html,mon analysis code is factored out of other?lters,which can sim-ply examine the analysis results calculated by a prior mea-surement?lter.Additionally,a?lter pipeline might choose to throw out the actual sample data,instead transmitting mea-sured statistics.We have implemented one measurement?l-ter.
The Stats?lter calculates the running mean,mean de-viation from the mean,and standard deviation from the mean for a stream of sample sets,and annotates each passing sam-ple set with the current values of these statistics.In its default mode,Stats works on groups of64samples at a time.It cal-culates its three statistics for each such group,then adds the three results to three separate exponentially-weighted mov-ing averages(EWMAs).A handful of summary counters are carried over from sample set to sample set,allowing these statistics to be calculated even when sets don’t contain even multiples of64samples each.Each sample set is annotated with the three EWMA values after the packet is processed. This informs downstream?lters of historical statistics for the sample stream,allowing them to detect unusual devia-tions.(The con?gurable EWMA smoothness constant is set toα=0.9375in our experiments.)
We use Stats’s running mean to detect the DC offset of the waveform,an indicator of biases in the underlying sens-ing system.For example,the TelosB ADC assigns the ratio of an input voltage to a reference voltage an unsigned12-bit value.For acoustic collection and other purely frequency-domain measurements,the center of our waveform should equal half the reference.Small hardware variations lead to imperfections in these measurements;we want to determine the actual signal bias,and do so simply by measuring the sig-nal mean.This is susceptible,of course,to aliasing effects. The components necessary to avoid aliasing would require either additional circuitry,or extensive processing dif?cult on a mote—an FFT/inverse-FFT pair or a convolution?lter requiring more than30multiplications per sample.We be-lieve a mote-resident software convolution?lter would be possible,but have not yet implemented one.Instead,our experiments deal with frequencies generally below half the sampling rate.The sampling rate is thus above the Nyquist rate,eliminating most aliasing effects;and we observed no aliasing effects in our experiments.
Classi?cation Filters These?lters classify each sample set as“interesting”or“not interesting”by modifying an an-notation in the sample set metadata.Each sample set be-gins as interesting;classi?cation?lters change uninteresting sample sets’annotations to“not interesting,”but leave other sets’annotations alone.Thus,a sample set is marked“in-teresting”at the end of a?lter bank if and only if every in-tervening classi?cation?lter thought it was interesting.The Packetizer component only transmits interesting sample sets;uninteresting sets are dropped.We have implemented three classi?ers,two generic and one designed speci?cally for Neuromote.In each case,the challenge was to imple-ment meaningful classi?cation with minimal computation—for example,to implement a lightweight frequency estima-tor precise enough to support meaningful classi?cation de-cisions.All of these?lters use the statistics generated by the Stats?lter,above.
The AmplitudeGate?lter marks sample sets uninter-esting if they have only low-amplitude data.It has two pa-rameters:the threshold above the signal mean and the num-ber of samples that must be above this threshold to consider the sample set interesting.The signal mean is read from the Stats annotation.
The FrequencyGate?lter classi?es sample sets based on their dominant frequency.It counts the number of times the signal amplitude crosses the mean and relates this count as a coarse measure of the dominant frequency of the col-lected signal[27].Of the known time-domain techniques, this is perhaps the simplest way to estimate a dominant fre-quency.This gate?lters the signal using uses two exponen-tially weighted moving averages to perform a simple multi-resolution temporal analysis.A sample set is considered in-teresting if the fast-moving,?ne-resolution EWMA is within a desired dominant frequency range.Each time we transi-tion from not interesting to interesting,we reinitialize the slow-moving,coarse-resolution EWMA to the value of the fast-moving EWMA.Subsequent sample sets are considered interesting so long as the slow-moving EWMA is still within the desired range.This technique is sensitive enough to avoid missing the beginning of an interesting event in the signal, and provides hysteresis,allowing the gating parameters to be set to a high level while avoiding both false positives and false negatives.We found empirically that for acous-tics,smoothness parameters ofα=0.5(fast-moving)and 0.96875(slow-moving)work well.
Finally,the SpikeDetector?lter was designed to de-tect single neuron activity in the form of a several-millisecond amplitude spike in the neural signal.The?lter’s single pa-rameter is the minimum spike height,measured in standard deviations above the mean;any sample above that height in-dicates a spike.A sample set that contains no spikes is not interesting.
Other classi?ers are possible,of course,as are other clas-si?er methodologies.For example,?lters can mark interest-ing sets as“interesting”(rather than marking uninteresting sets as“not interesting”);this leads to sets that are interest-ing if any(rather than all)of the classi?ers were interested. Compression Filters Finally,compression?lters reduce a sample set’s resolution,pack its samples more tightly,or
compress it using a stateful compression algorithm.The goal is simply to reduce the data that the mote must transmit.
The Format?lter alters the precision and alignment of samples within a buffer.Depending on its parameters,Format can reduce sampling precision by truncating12-bit samples to8bits each,or reduce waste by packing pairs of12-bit samples into3bytes each.Since Format’s output is a valid sample set,annotated appropriately with its precision and alignment,Format may appear before or after other?lters.
The Adpcm?lter is an adaptive pulse-code modulation compressor used in Auricle.Adaptive pulse-code modula-tion is a well-known technique for lossy compression of voice data.When compared to several other compression schemes, including LPC schemes(GSM6.10)and simple logarith-mic encoding(u-law),ADPCM has the best combination of sound quality and compression rate among the few viable for real-time compression on motes.We use a variant of the In-tel/DVI ADPCM codec,modi?ed to eliminate mulitplication and division operations.Encoding with ADPCM reduces12-bit ADC samples to4-bit values.These values are not sam-ples,and cannot be operated on until they are expanded into samples;thus,Adpcm must occur last in any?lter chain of which it is a part.ADPCM is stateful,so to ensure resiliency to packet loss(and to Packetizer’s refusal to transmit un-interesting sample sets),we include the state of the encoder in each packet as a four-byte header extension.
Finally,SpikeDetector can be con?gured to compress as well as classify.When con?gured in spike-only mode, SpikeDetector?lters out the baseline noise to produce an abridged version of the signal,containing only time-referenced spike waveforms.It also measures each spike’s height(am-plitude)and width(duration),as these help distinguish the neurons from which it was generated;the sample set is anno-tated with these parameters,which might obviate most inves-tigators’need for the raw waveform.The resulting data,like that of Adpcm,uses a special format,so a SpikeDetector in spike-only mode must occur last in the?lter chain.
4C ONTROL P ATH
Motes collect,process,and transmit data.Data reception, backend processing,data presentation,and the creation and dispactch of control messages are the responsibility of the microservers in our network.Hence,our applications are comprised of two executables:one for the deployed mote, another for the backend microserver.
To tightly intregrate these two executables,we write all of our services using a combination of nesC module de?-nitions[7]and SNACK service compositions[10].In order to execute this code on a PC,we use the EmTOS environ-ment[8].From the perspective of a mote,an EmTOS appli-cation running on a PC-class device appears to be another mote.However,from the perspective of the Linux system on which an EmTOS application is running,it appears to be a standard Linux process that may interact with other system processing using IPC.
Using nesC as the base module description language for our entire system greatly simpli?es porting data processing algorithms,control interpretation logic,and link and rout-ing code from motes to microservers and vice versa.In most cases,the port requires no coding changes.
Over the course of a user session with the application, she may alternate between invoking classi?cation elements on the back-end microserver and on the motes.Control over where and how processing occurs is determined by a user connected to the microserver.Backend processing is invoked using the same tasking syntax as is used to control deployed motes.
4.1Tasking and Control
Control of the application is exposed via socket so that a human user or controlling application may issue commands from any device with an IP https://www.wendangku.net/doc/7b17009981.html,mands are sent in ASCII and have the following syntax:
dest:cmd-name cmd-value[;cmd-name cmd-value]*
broadcast:gate-threshold200;adpcm enable true
and to transfer dominant-frequency gating responsibil-ities from the microserver to mote4,for instance,we can write:
local:dominant-frequency-enable false
4:dominant-frequency-enable true
Commands are translated into a shorter binary format be-fore being sent to the control activation service.At the time of this writing,there are about50commands de?ned in our system.
The sampler and all processing elements are initialized and controlled by a single control activation service.The control activation service dispatches signals for local pro-cessing control and marshalls control commands for trans-mission to deployed nodes.Since processing elements may exist on either the motes or their microservers,this service resides on both.
Commands are collected as type-length-value attributes. Consequently,they are stored in the same scratchpad data structure used by the data path.
In single-hop scenarios these messages are delivered di-rectly to intended recipients(either by unicast or broadcast addressessing).In multihop scenarios,for reliability control messages are?ooded using the Drip[15]dissemination pro-tocol irrespective of the destination address.Relative to the bandwidth consumed by our data traf?c,the overhead even of?ooding control messages is small.
Filter
ADPCM
0.628
with summarization
0.776
with summarization
3.790
Format
1.010
SpikeDetector
1.017
Table1—Worst-case run times of our data processing elements for152byte (typical)data buffers.The most MCU-intensive?lter(ADPCM)consumes 44%of available cycles when sampling at8kHz.
4.2Iterative Experimentation
Real applications must support interactive tuning.Best effort delivery leads to indiscriminate packet loss;packets contain-ing interesting data are just as likely to be discarded as those containing only noise.High-rate data must be classi?ed and ?ltered before transmission.
Factory calibration of an application’s data extraction pa-rameters is rarely suf?cient:
?The deployment environment may produce unpredictable noise and other distractions that should be?ltered from
the raw signal.
?Climatic and physical variation impacts RF propaga-
tion.
?Network topologies may vary in both the level and
uniformity of density,making bandwidth availability
unpredictable.
?The users of sensor networks haven’t seen spatially
dense data before,so they often don’t a priori how best
to?lter it.
To support interactive tuning,this processing chain spans from the data acquisition subsystem of a mote,to the data presentation facilities of a microserver.Packetization com-ponents obscure the physical break in the processing chain (when data is transmitted and received)to provide a single-chain abstraction.
On motes,data produced and enqueued by our sampler
is introduced into the chain.Data exiting the chain is trans-mitted.For instance,a processing chain con?guration may look like:
Sampler[Filter]->Fragment
->Statistics
->AmplitudeGate
->Packetizer;
On our microservers,data received from the network is introduced into a processing chain.Data exiting the chain
is passed to modules responsible for presentation,either by socket or Emstar device?les:
Unpacketizer[Filter]->AmplitudeGate
->FrequencyGate
->PacketDevices;
Data is passed through the chain one buffer at a time. Buffers each typically hold100to200samples.5R ESULTS
This section presents experimental results for the Auricle
and Neuromote applications.We demonstrate that it is pos-
sible to collect high rate data over low rate radios;that our
simple and coarse signal processing subsystem works well;
and that through runtime con?guration of?ltering parame-
ters,we can dramatically improve the effectiveness of our collection system.
Our experiments include tests in indoor and laboratory settings,using single-link and multihop communication ser-
vices.We evaluate the performance of several combinations
of processing components,including Adpcm,Stats,AmplitudeGate, FrequencyGate,and SpikeDetector,and measure the resulting?delity tradeoffs in our networks.
5.1Auricle:Classi?er Tuning
First,we quantify and characterize the Auricle application
using an AmplitudeGate classi?cation?lter.We deploy
12motes,a single acoustic data source,and a single sink,
and de?ne the system’s goal as collecting as much of the sig-
nal as possible at the sink.The deployment models a dense
sensor network:the source phenomenon is detected simul-taneously at all nodes,although with varying?delity due to environmental factors.If all nodes that detect a signal report
that signal,one might reasonably expect link contention to
reduce the signal recovered at the sink.Our performance hy-pothesis is that?ltering the sample stream on the motes can increase the total recovered signal power by reducing link contention.This experiment veri?es our system’s sampling, statistics,compression,and transmission functions,and its
ability to simultaneously perform basic signal processing op-erations.
Twelve Auricle motes were situated in a straight line
with each pair separated by six feet(Figure2).A loudspeaker
was placed at the end of the line,six feet from the?rst
mote.Each mote was elevated three feet off the ground.A microserver sink was deployed close to the middle of the
line,one hop away from every mote.We played two min-
utes of a recording of former president Jimmy Carter’s“Cri-
sis of Con?dence”speech,occasionally interrupted by a cell
phone ringing.We measured the total signal energy collected
at the sink when the motes were instructed to capture every-
thing,and when the motes were instructed to gate on am-plitude;amplitude gating essentially provides us with noise suppression,silence elimination,and basic event detection.
We sampled at4kHz,which is enough to clearly understand
a voice,albeit with a noticeable loss of quality.The voice
and ring amplitudes were roughly equal.Experiments were
run outdoors at night in an open-space environment.
Observed sound pressures varied throughout the speech
from78–86dBC at the closest mote to the speaker to62–
68dBC at the farthest.The peak-to-peak amplitude of the
closest mote’s signal measured around1.22V,roughly half
the range of our ADC.We varied the amplitude threshold
for AmplitudeGate from0(that is,no gating)to488mV
Figure2—Linear topology used for unicast experiments.White circles rep-resent mote sensor nodes;the grey circle is the microserver sink. above the mean.1The received power was calculated as P=
s2
i ,where s i denotes sample i’s AC-coupled value.Only
samples collected at the sink were counted.Baseline record-ings were conducted to measure the signal power each node is capable of receiving in a contention-free environment,and we normalized our recovered power readings using these baseline measurements.
Figure3shows total recovered power as a function of the amplitude gate’s threshold.Each data point represents an average of3trials.
The shape of this graph is dominated by two competing effects.At very low thresholds,little or none of the signal is?ltered before transmission.This results in channel con-tention and packet loss.Contention becomes less severe as the threshold increases because silent periods,as well as the lower-amplitude signals at distant nodes,are proactively?l-tered.This,in turn,allows a smaller number of nodes—those better positioned to detect the signal—to have improved ac-cess to the wireless channel.At very high thresholds,so much data is suppressed at the gate that not only is the chan-nel underutilized,but signi?cant features in the raw signal are removed before transmission.
Although the general trend is intuitive,we expected the maximum signal power to occur at a threshold just above ambient noise(at about50mV).The burstiness of human speech is commonly exploited to optimize voice communi-cation systems;we believed that once our gate eliminated the background noise between words and syllables,the ag-gregate received power would trend quickly upwards.In-stead,we found that even when gaps between words were ?ltered,the remaining traf?c was still great enough to induce contention-based packet loss and decreased signal https://www.wendangku.net/doc/7b17009981.html,-work effects are more pronounced than we had initially ex-pected.
In Figure3,total recovered power is maximized at a gate setting of around230mV.However,this setting does not—say—maximize the closest mote’s contribution.Fig-ure4plots the per-node signal contribution for the experi-ments of Figure3for three nodes:one close to the signal, one far away,and one in the middle.At very low gating lev-els,each mote transmits approximately the same number of packets,and thus approximately the same fraction of each node’s total theoretical signal power is received.However, since acoustics attenuate rapidly with distance,nodes farther from the source will record the audio at lower absolute am-plitudes than closer nodes,and the peak per-node contribu-tion shifts with distance from the source.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0
100
200 300 400 500 600 700 800
900
F r a c t i o n o f B a s e l i n e P o w e r
Minimum Dominant Frequency (Hz)
Total Recovered Power
Phone: Amp. Gate 245 mV Phone: Amp. Gate 120 mV
Phone: No Amp. Gate Voice: Amp. Gate 245 mV Voice: Amp. Gate 120 mV
Voice: No Amp. Gate
Figure 5—Total signal power recovered at the sink from all nodes,for inter-esting (“Phone”)and uninteresting (“V oice”)signals,as we vary the mini-mum dominant frequency to pass.Optimizing parameters for a composition of ?lters cannot be done by optimizing each ?lter independently.Recovered power is normalized to the total theoretical power for the relevant the por-tion of the signal.
ous section,but change the goal:now we want to capture as much of the cell phone signal as possible at the sink.Our performance hypothesis is that motes are capable of suf?ciently sophisticated ?ltering to discriminate classes of high-rate signals (here,based on frequency).This experi-ment shows that the dominant frequency ?lter is effective,far surpassing the amplitude gate as an event detector and classi?er for an interesting class of signals.Moreover,we show that ?lters are as effective in combination as they are alone—more effective,in fact,at reducing signal-to-noise rato.
We collected baseline measurements as before,and care-fully noted the times when the cell phone started and stopped ringing.We use these notes to determine how much recov-ered power is derived from periods when the cell phone is ringing and,conversely,when the cell phone is not ringing.Since the rings have higher frequency than the voice por-tion of the signal,we apply our dominant frequency gate to attempt to cut the voice portion of the signal without affect-ing the phone portion.The FrequencyGate classi?er is in-structed to ?lter out sample sets whose dominant frequency is below a speci?ed minimum,and vary this minimum from 0Hz to 920Hz in steps of approximately 65Hz.We also include an AmplitudeGate classi?er,varying its threshold among 0mV (off),120mV ,and 240mV .
Figure 5shows the results.The frequency ?lter is clearly effective:a sharp decrease in unwanted (“V oice”)power hap-pens between frequency gates of 200and 400Hz,regardless of threshold setting,indicating that the dominant frequency of much of the President’s speech falls in this range.In this experiment,and at the amplitude gates we tested,the com-bination of frequency and amplitude gating does not greatly improve recovered signal power compared to amplitude gat-ing alone.However,it does improve the signal-to-noise ra-tio (SNR).The best setting to remove voice while retaining phones lies around 500kHz,depending on the desired com-promise between false negative and false positive readings.
With an amplitude gating level of 245mV and a dominant frequency ?lter setting of approximately 525Hz,the ratio of interesting (ring)power to uninteresting (other)power is ap-proximately 16.6:1,giving an SNR of 12.2dB.By compari-son,when the dominant frequency ?lter is effectively off,the power ratio is 1:1.4,giving a SNR of ?1.39dB.Therefore,optimizing the dominant frequency ?lter parameters results in a 13.6dB improvement in SNR.
Amplitude gating at 245mV—approximately the opti-mal setting determined in the last section—can help or hurt signal recovery depending on the frequency threshold.For frequency gates below approximately 725Hz,amplitude gat-ing at 245mV is always better than lower gating levels;above that point,the opposite is true.Such high frequency gates remove enough of the signal that the network is un-congested;additional ?ltering only reduces received signal power.This again argues that tuning must be done in the ?eld:?lter effects and their implications on network perfor-mance make iterative,run-time tuning of ?lter parameters an effective way to improve network performance.Moreover,individual ?lter elements cannot be tuned individually and then applied together,as such an approach would likely yield sub-optimal results.
5.3Auricle:Multihop Collection
We now turn from a dense deployment to a sparse multihop deployment:one in which nodes are between one and three hops from the sink,in which sources are heard by relatively few motes,and in which there is more than one source.A single-hop deployment of Auricle already pushes the lim-its of the TelosB’s resources;multihop networking demands signi?cant additional RAM devoted to forwarding queues,and additional code space to accomodate the routing and transport algorithms.Our performance hypothesis is that the network effects that in?uence tuning levels in one-hop ex-periments will only increase in multihop networks,and in particular,that the best tuning levels for recovering signal power will vary based on the distance of a source from the sink.We ?nd that although there is a noticeable performance decrease with multihop networking,the integrated system as a whole operates well.Even in the presence of signi?cant unwanted traf?c generators,?lter parameters can be tuned to allow a signi?cant portion of theoretical power to be re-covered from one node of interest.
Fourteen TelosB motes were deployed in a 3-by-5grid,with nodes along the length separated by 25feet and nodes along the width separated by 8feet.A sink was deployed at one corner of the network,with sources six feet from the nodes at two other corners;see Figure 6.The node closest to one source had one-hop connectivity to the base station,while the node closest to the other had two-hop connectivity.We placed the nodes directly on the ground to increase RF attenuation,thus making the study of multihop effects ten-able.2To implement the multihop networking functionality,
Figure6—Grid topology used for multihop experiments.Diameter is three hops.The grey shaded node represents the data sink.
we enhanced Auricle to support multihop tree-based collec-tion and to use the MultihopLQI variant of the MintRoute code[30].In constrast to actively beaconing,as is required
to ascertain link quality on more primitive mote radios,Mul-tihopLQI incorporates a value of link quality provided di-rectly by the CC2420radio into a path cost estimate.To re-peat experiments,we froze the routes once they stabilized. To disseminate tasks we use the Drip protocol implemen-tation[15],which reliably disseminates control messages throughout the network with a delay of up to several sec-onds.As in the single hop experiments,we instruct all nodes
to collect data.
We ran two sets of experiments for amplitude gate set-tings.The nodes were divided into two groups,depending on which source(the1-hop source or the2-hop source)was physically nearest.In one set of experiments,we set the1-hop source group’s gate to275mV and varied the gate for the2-hop source group;in the other set,we did the reverse,?xing the2-hop source group’s gate and varying that of the
1-hop source group.This procedure was designed to demon-strate the impact of hop distance on optimum gating level. As usual,we compare against baselines measured once per node in the absence of contention—the baselines for the two sources are approximately equal.Each point represents the best result of?ve trials.
Figures7and8show the results.The maximum total recovered power,55%of the theoretical power,is obtained when the1-hop source group has amplitude gate275mV and the2-hop source group has amplitude gate450mV;this point is visible as the2-hop source’s peak in Figure7.The gating level for the2-hop source group is probably higher than that for the1-hop source group because network con-tention impacts multihop transmission proportionately more than single-hop transmission.This is further visible in the differences between Figure7and8at high gating levels: even when all of the1-hop source’s signal is?ltered out, the2-hop source with gating level275mV(right-hand side
of Figure8)achieves less recovered power than it does with a higher gating level in Figure7.Aggressive,topology-dependent ?ltering can thus improve the recovered power from a net-work;and we have shown that?lter tuning can lead to a system that can successfully collect high-rate signals over low-rate radios,even in a multihop network.
nodes and the sampling rate.(2)Lowering the sampling rate
will lead to post processing error in the measurement of spike heights and widths,and ultimately to missed spikes when we signi?cantly undersample.(3)Since spike signals occupy only about three percent of the raw waveform,the removal of noise between spike events and the transmission of a concatenation of spike waveforms will decrease channel contention.
Our experimental setup consists of an waveform gen-erator programmed to output pre-recorded neural signals, a neural preampli?er circuit,and eight TelosB motes.The data programmed into the waveform generator was origi-nally acquired in vivo from freely moving rats using?ve four-channel MOSFET input operational ampli?ers mounted in the cable connector to remove movement artifacts.Data were recorded wide band(0.1Hz to5kHz)and sampled at10kHz/channel(16channels)with12-bit precision.The spikes were isolated from the local?eld potentials by apply-ing a high-pass?lter with an f-3dB frequency of600Hz. The output signals from the waveform generator are applied to the neural preampli?er circuit,which ampli?es and DC references the signals,which are then applied directly to the ADC inputs of all eight TelosB motes.The dataset corre-sponds to one second of neural activity,over which there are seven spikes,each with an amplitude of1.85volts(at the ADC input of the TelosB motes),and a peak-trough duration of approximately1.34ms.
We performed two sets of tests.One test measured the percentage of spikes recovered for different(a)sampling fre-quencies and(b)number of motes communicating in the network.This test was performed in two modes:raw and spike-only.In the raw mode of operation,the entire sam-pled signal was transmitted by each node.In the spike-only mode of operation,the spike waveforms were isolated lo-cally by the motes,who transmitted an abridged version of time-referenced spike waveforms only,thus resulting in a de-crease in network traf?c.
The second test measured the extent of spike parameter variation resulting from different sampling rates on one to eight motes.The spike parameters of interest are(a)spike height,which is the voltage of the acquired signal peak,and (b)spike width,which is the time difference between the spike’s peak voltage and minimum voltage.Spikes that were lost due to packet loss are not accounted for in this test.The sample data used contained spikes from a single cell;there-fore,each spike was originally equal in amplitude.However, the the?nite resolution(12-bit),and sampling rate(10kHz) of the original neural signal acquisition apparatus has re-sulted in a deviation of108mV in the dataset that has been programmed into the waveform generator.
Figure9describes the percentage of recovered spikes as a function of the number of nodes in the network at sampling rates of2kHz and8kHz.When sending the complete raw data set,we?nd that the network has the bandwidth to sup-port one or two motes sampling at2kHz,returning nearly
P
e
r
c
e
n
t
S
p
i
k
e
s
R
e
c
o
v
e
r
e
d
Number of Motes vs. Spike Recovery in Spike Only Mode
Figure9of nodes in the network.In the continuous signal transmission mode,100% of spikes are only recovered with up to2motes,both of which must be sampling at2kHz.However,when only the waveform containing spike information is transmitted,our spike recovery rate is near100%for up to six motes sampling at2Khz and always above65%when sampling at8kHz
P
e
r
c
e
n
t
o
f
S
p
i
k
e
s
R
e
c
o
v
e
r
e
d
Rate vs. Spike Recovery in Spike Only Mode
Figure10—Percentage of recovered spikes as a function of sampling rate and node count in both continuous signal and isolated spike waveform tran-simssion modes.In mote modes,spike loss is observed below2kHz due to under-sampling.At2kHz and above,due to channel contention the spike recovery rate drops with rate,but much less signi?cantly with abridged data. Packet loss is worse for larger networks.
100%of the data pertaining to spikes.However,even at a 2kHz sampling rate,with three transmitting motes the spike delivery rate drops off signi?cantly(to under60%)and de-cays to20%when eight motes are active in the network.In terms of the total number of spikes returned by the network, when sending raw data we?nd that the maximum spike yield occurs with two motes.
In contrast,when spike information is concatenated and the noise between spikes is removed before transmission,the network performs signi?cantly better.At2kHz our yield is above90%,even for eight motes.At8kHz the network suf-fers a bit from contentions and the spike recovery rate decays linearly from100%with two motes to to under70%with eight motes.In terms of the total number of spikes received, we see that the maximimum total spikes recovered(with8 motes at8Khz)is roughly two and a half times greater than the maximum for our raw data experiments(with2motes at 2kHz).
Figure10describes the effects of sampling rates on net-
0.001
0.0015 0.002 0.0025 0.003 0.0035 0.004 0 1000 2000
3000 4000 5000 6000S p i k e W i d t h (s e c o n d s )
Rate (Hz)
Effects of Undersampling on Spike Width
median spike width
Figure 11—Median spike width as a function of sampling rate.The er-ror bars indicate the 1st and 3rd quartiles.The dotted line at 0.0013seconds represents the median width for all spikes in the input.At frequencies below 2kHz,the variation in the recovered spike widths is high due to undersam-pling.As the sampling rate is increased,the reported median spike width converges to the correct value.Recovering the median spike width with in-creased accuracy enables investigators to classify the cell from which the spike pattern originated with greater precision.
works of one,two,and eight motes.Irrespective of whether raw or abridged waveforms are sent,at below 2kHz we wit-ness considerable spike loss due to undersampling.But again in this ?gure,it is evident that the network can support two nodes worth of raw data at up to a sampling rate of 2kHz.This suggests that when sending the complete waveform the only operating point where nearly all spikes are detected is at 2kHz with exactly two motes.
Figure 10also shows us that there is a penalty to re-questing all raw waveforms,due to network congestion.This penalty is more pronounced both as the number of motes and as the sampling rate is increased.However,in the isolated spike waveform transmission mode for 1and 2nodes the re-covery rate is near 100%irrespective of sampling rate.We receive nearly 80%percent of the spikes when running eight motes in this mode.
The results for the test measuring the variation of ac-quired spike width and height as a result of different sam-pling rates are displayed in Figures 11and 12,respectively.As expected,the greatest amount of variation can be ob-served when the signal is being undersampled (1kHz).The parameter variation drops off with increasing sampling fre-quency and levels off between 2and 3kHz for both height and width variation.At 3kHz the variation of the spike heights and widths approach those of the original dataset.The vari-ation of spike heights and widths acquired by the motes at a 10kHz sampling rate are not zero due to slight distortion in the preampli?er and sigma-delta analog-to-digital converter circuits,and 8-bit signal signal quantization (as opposed to the 12-bit quantization of the original signal).
6R ELATED W ORK
The applications we have built fall into two categories:acous-tic collection and biological monitoring.In sensor networks,acoustic systems are typically factory-tuned to detect spe-
1.2
1.3 1.4 1.5 1.6 1.7 1.8 1.9 0
1000 2000
3000 4000 5000 6000
S p i k e H e i g h t V a r i a t i o n (V o l t s )
Rate (Hz)
Effects of Undersampling on Spike Height
median spike height
Figure 12—Median spike height as a function of sampling rate.The error bars indicate the 1st and 3rd quartiles.The dotted line at 1.85V represents the actual median height for all spikes in the input.At frequencies increase,the accuracy of the reported spike heights increases substantially.Accu-rate reporting of spike height enables investigators to classify the cell from which the spike pattern originated with greater precision.
ci?c events;they transmit information about these events and discard the raw waveform.The counter-sniper system [26],is one example that uses mote networks (and signi?cant com-plementary signal processing hardware)to detect gunshots and localize their sources.Likewise,VigilNet [11]uses acous-tics as part of its surveillance system.Our work considers not only local acquisition of high-rate data and event detec-tion,but also a wide array of data reduction choices to make transmission of the raw waveform tenable.
In the biological domain,the CodeBlue [17]project has developed mote-based sensors with biological interface cir-cuits similar to those used for our neural monitoring appli-cation.The CodeBlue sensors are also similar to our work in an application sense,as they are used to acquire and wire-lessly transmit biological signals.However,the previously-reported CodeBlue sensor application was not designed to acquire and ?lter data at the rates necessary,for example,to obtain and transmit neural spike activity.
Until recently,sensor network data reduction has been treated as a collaborative process among motes [13,18].Build-ing on the observations of Tenet [9],we diverge from this trend.Although we allow a fairly extensive level of runtime control over signal processing,we constrain motes to pro-cess only locally generated data and disallow mote-to-mote collaborative processing.This restriction simpli?es our sys-tem design.In the imaging domain,Cyclops [23]follows a similar architectural pattern.
The problem of reducing high-rate data to alleviate net-work congestion has been studied extensively in the context of Internet and cellular transcoding [6,20,31].We share one goal in particular with these projects,maximizing appli-cation performance given bandwidth constraints.Unlike the previous work,however,a sensor network application com-petes with itself for bandwidth.This property alone can lead to ?ner control over how data is collected and ?ltered,be-cause a sensor network application can make network-wide decisions about how to allocate and use bandwidth.Unlike
the previous work which was focused on the contract be-tween many clients and a single server,we focus on the sin-gle client of typical sensor network application and its use of many servers supplying data.Hence,we investigate con-trol across servers.There is some interesting research on dis-tibuted music that considers multiple servers providing re-lated data to a client[3,25]but unlike our project,which focuses on making effective use of bandwidth,these works focused on the dif?cult task of synchronizing a performance with low latency.
7F UTURE W ORK
The most important next steps for this project focus on the processing library and tasking language.We envision a score of simple waveform classi?ers that will detect various types of sounds and sound patterns and report when these sounds occur and measurement components that will augment our statistics component to return basic properties about sensed waveforms in addition to raw waveform data.To support short burst at very high sampling rates,we will implement Flash buffering.
Furthermore,we will expand and formally de?ne the se-mantics of our tasking language to add features necessary to command motes as groups,focus on motes sensing certain features in their waveforms,duty cycling,adding wakeup events,etc.This work will be done in the larger context of the Tenet architecture.
Real embedded systems need to be optimized.We will be focused on improving the communication subsystems, perhaps adding?ow control to multihop transport and win-dowed acknowledgements to the serial wire protocol.
8C ONCLUSION
We have designed a heterogeneous software system capa-ble of high rate data acquisition,ef?cient signal process-ing for data classi?cation and compression,and single-and multiple-hop wireless transmission.We have presented high rate collection applications in two domains:acoustics and bi-ological monitoring.We found that simple?lters can impact network performance greatly,particularly in the bandwidth limited environments.As we have shown with both applica-tions,selective?ltering before transmission yields more of the data that interests a user.Furthermore,since usage pat-terns vary and environmental dynamics in?uence our appli-cation behavior,we have found system calibration to require runtime tuning.
A CKNOWLEDGMENTS
This work was made possible by the NSF Center for Em-bedded Networked Sensing(CENS)under contract number CCR-0120778.SNACK is funded by the National Science Foundation under award number0435497.We would like to thank Alec Woo and Gilman Tolle for contributing tree-routing code,Phil Levis for the Drip dissemination protocol, and our anonymous reviewers for their valuable feedback.R EFERENCES
[1] A.Bragin,J.Engel,C.L.Wilson,I.Fried,and G.W.Math-
ern.Hippocampal and entorhinal cortex high-frequency os-cillations(100-500hz)in human epileptic brain and kainia acid-treated rates with chronic seizures.Epilepsia,40:127–137,February1999.
[2]Alberto Cerpa,Jeremy Elson,Deborah Estrin,Lewis Girod,
Michael Hamilton,and Jerry Zhao.Habitat monitoring:Ap-plication driver for wireless communications technology.In Proc.2001ACM SIGCOMM Workshop on Data Communi-cations in Latin America and the Caribbean,April2001. [3]Elaine Chew,Roger Zimmermann,Alexander A.Sawchuk,
Christos Papadopoulos Chris Kyriakakis,Alexandre R.J.
Francois,Gerry Kim,and Anja V olk.Musical interaction at
a distance:Distributed immersive performance.In4th Open
Workshop of MUSICNETWORK,2004.
[4]ESS.Extensible sensing system:An advanced network
design for micr oclimate sensing.http://www.cens.
https://www.wendangku.net/doc/7b17009981.html,.
[5]S.Farshchi,P.H.Nuyujukian,A.Pesterev,I.Mody,and J.W.
Judy.A TinyOS-based wireless neural sensing archiving and hosting system.In Proceedings of the2nd International Con-ference IEEE Engineering in Medicing and Biology Society Conference on Neural Engineering,March2005.
[6]Armando Fox,Steven D.Gribble,Yatin Chawathe,Eric A.
Brewer,and Paul Gauthier.Cluster-based scalable network services.In SOSP’97:Proceedings of the sixteenth ACM symposium on Operating systems principles,pages78–91, New York,NY,USA,1997.ACM Press.ISBN0-89791-916-5.
[7]David Gay,Philip Levis,Robert von Behren,Matt Welsh,Eric
Brewer,and David Culler.The nesC language:A holistic ap-proach to networked embedded systems.In Proc.ACM SIG-PLAN2003Conference on Programming Language Design and Implementation(PLDI),pages1–11,San Diego,Califor-nia,June2003.
[8]Lewis Girod,Thanos Stathopoulos,Nithya Ramanathan,
Jeremy Elson,Deborah Estrin,Eric Osterweil,and Tom Schoellhammer.A system for simulation,emulation,and de-ployment of heterogeneous sensor networks.In SenSys’04: Proceedings of the2nd international conference on Embed-ded networked sensor systems,pages201–213,New York, NY,USA,2004.ACM Press.ISBN1-58113-879-2.
[9]Ramesh Govindan,Eddie Kohler,Deborah Estrin,Fang Bian,
Krishna Chintalapudi,Om Gnawali,Ramakrishna Gummadi, Sumit Rangwala,and Thanos Stathopoulos.Tenet:An archi-tecture for tiered embedded networks,October172005. [10]Benjamin Greenstein,Eddie Kohler,and Deborah Estrin.A
sensor network application construction kit(SNACK).In Proc.2nd ACM Conference on Embedded Networked Sensor Systems(SenSys2004),pages69–80,Baltimore,Maryland, November2004.
[11]Tian He,Sudha Krishnamurthy,John Stankovic,Tarek Ab-
delzaher,Liqian Luo,Radu Stoleru,Ting Yan,Lin Gu,Gang Zhou,Jonathan Hui,and Bruce Krogh.Vigilnet:An in-tegrated sensor network system for energy-ef?cient surveil-lance.In SenSys’05:Proceedings of the3rd international conference on Embedded networked sensor systems,New
York,NY,USA,2005.ACM Press.
[12]Crossbow Technology Inc.https://www.wendangku.net/doc/7b17009981.html,/.
[13]Chalermek Intanagonwiwat,Ramesh Govindan,and Deborah
Estrin.Directed diffusion:a scalable and robust communica-tion paradigm for sensor networks.In Proceedings of the In-ternational Conference on Mobile Computing and Network-ing,August2000.
[14]J.M.Kahn,R.H.Katz,and K.S.J.Pister.Next century chal-
lenges:mobile networking for Smart Dust.In Proc.?fth an-nual ACM/IEEE international conference on Mobile comput-ing and networking,pages271–278,1999.
[15]P.Levis,N.Patel,D.Culler,and S.Shenker.Trickle:A self-
regulating algorithm for code propagation and maintenance in wireless sensor networks,2004.URL citeseer.ist.
https://www.wendangku.net/doc/7b17009981.html,/levis04trickle.html.
[16]Philip Levis,Sam Madden,David Gay,Joseph Polastre,
Robert Szewczyk,Alec Woo,Eric Brewer,and David Culler.
The emergence of networking abstractions and techniques in TinyOS.In Proc.1st Symposium on Networked Systems De-sign and Implementation(NSDI’04),pages1–14,San Fran-cisco,California,March2004.
[17]K.Lorincz,D.J.Malan,T.R.F.Fulford-Jones,A.Nowoj,
A.Clavel,V.Shnayder,G.Mainland,,and M.Welsh.Sensor
networks for emergency response”challenges and opportuni-ties.IEEE Pervasive Computing.
[18]Samuel Madden,Michael Franklin,Joseph Hellerstein,and
Wei Hong.TAG:a tiny aggregation service for ad-hoc sensor networks.In Proc.5th Symposium on Operating Systems De-sign and Implementation(OSDI’02),Boston,Massachusetts, December2002.
[19]Alan Mainwaring,Joseph Polastre,Robert Szewczyk,David
Culler,and John Anderson.Wireless sensor networks for habitat monitoring.In ACM International Workshop on Wire-less Sensor Networks and Applications,Atlanta,GA,Septem-ber2002.
[20]Steven McCanne,Van Jacobson,and Martin Vetterli.
Receiver-driven layered multicast.In ACM SIGCOMM,vol-ume26,4,pages117–130,New York,August1996.ACM Press.ISBN0-89791-790-1.URL citeseer.ist.psu.
edu/steven96receiverdriven.html.
[21]Joe Polastre,Cory Sharp,and Rob Szewczyk.URL http:
//https://www.wendangku.net/doc/7b17009981.html,.Moteiv website.
[22]Gregory J.Pottie and William J.Kaiser.Embedding
the internet:wireless integrated network sensors.43(5): 51–58,May2000.URL https://www.wendangku.net/doc/7b17009981.html,/ pubs/articles/journals/cacm/2000-43-5/
p51-pottie/p51%-pottie.pdf.
[23]Mohammad Rahimi,Deborah Estrin,Rick Baer,Henry
Uyeno,and Jay Warrior.Cyclops,image sensing and inter-pretation in wireless networks.In SenSys’04:Proceedings of the2nd international conference on Embedded networked sensor systems,pages311–311,New York,NY,USA,2004.
ACM Press.ISBN1-58113-879-2.
[24]Tom Roberts and Malcolm Slaney.Fixed-point fast fourier
transform.http://www.jjj.de/fft/int