Scholarly article on topic 'Wireless Sensor Networks for Distributed Chemical Sensing: Addressing Power Consumption Limits With On-Board Intelligence'

Wireless Sensor Networks for Distributed Chemical Sensing: Addressing Power Consumption Limits With On-Board Intelligence Academic research paper on "Electrical engineering, electronic engineering, information engineering"

0
0
Share paper
Keywords
{""}

Academic research paper on topic "Wireless Sensor Networks for Distributed Chemical Sensing: Addressing Power Consumption Limits With On-Board Intelligence"

Wireless Sensor Networks for Distributed Chemical Sensing: Addressing Power Consumption Limits With On-Board Intelligence

Saverio De Vito, Paola Di Palma, Carmine Ambrosino, Ettore Massera, Gianbattista Burrasca,

Maria L. Miglietta, and Girolamo Di Francia

Abstract—Chemicals detection and quantification is extremely important for ensuring safety and security in multiple application domains like smart environments, building automation, etc. Characteristics of chemical signal propagation make single point of measure approach mostly inefficient. Distributed chemical sensing with wireless platforms may be the key for reconstructing chemical images of sensed environment but its development is currently hampered by technological limits on solid-state sensors power management. We present the implementation of power saving sensor censoring strategies on a novel wireless electronic nose platform specifically designed for cooperative chemical sensing and based on TinyOS. An on-board sensor fusion component complements its software architecture with the capability of locally estimate air quality and chemicals concentrations. Each node is hence capable to decide the informative content of sampled data extending the operative lifespan of the entire network. Actual power savings are modeled and estimated with a measurement approach in experimental scenarios.

Index Terms—Electronic noses, power saving, sensor censoring, wireless chemical sensing.

I. Introduction

POLLUTANT and toxic gases affect human health in several scenarios. Their effects can vary from localized, low intensity to systemic with potential life threats, depending mainly on exposure time and concentrations. Hazardous gas, like explosives or flammable ones are also source of increasing concern for security reasons due to their possible use in terrorists attacks to military installations. Some of them are currently in use or are foreseen to be used as energy carrier for automotive transports and so their diffusion is expected to

Manuscript received September 03,2010; accepted September 08,2010. Date of publication September 27, 2010; date of current version February 09, 2011. This work was supported in part by P.O.R. Campania 2002/2006 mis. 3.17, under Grant 4997 "Sistema Integrato Multisensoriale per il Monitoraggio Ambientale" (SIMMA-ICT) and by Project FP7-ENCOMB. The associate editor coordinating the review of this paper and approving it for publication was Prof. Ralph Etienne-Cummings.

S. De Vito is with the Italian National Agency for New Technologies, Energy and Sustainable Development (ENEA), Portici Research Center, 80055 Portici (NA), Italy, and also with the University of Cassino, DAEIMI Department, 03043, Cassino, Italy (e-mail: saverio.devito@enea.it).

E. Massera, G. Burrasca, M. L. Miglietta, and G. Di Francia are with the Italian National Agency for New Technologies, Energy and Sustainable Development (ENEA), Portici Research Center, 80055 Portici (NA), Italy.

P. Di Palma and C. Ambrosino are with the National Research Council, CNR-ISAFOM, 80056 Ercolano (NA), Italy.

Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/JSEN.2010.2077277

grow significantly, for example, hydrogen powered car refilling stations could become very common in the near future [1]. The capability to detect and monitor the presence of such substances in the work or living environments, being indoor or outdoor, is thus of paramount importance in multiple applicative scenarios and should be considered as a potential life saving assets. Unfortunately, the characteristics of chemical signal propagation make the use of single measurement point unfeasible and mostly inefficient. Actually, propagation of chemical plumes in real environments is primary related to complex fluid dynamics effects (e.g., turbulence) with diffusion being negligible in many circumstances [2]. Actually, diffusion velocities of gas molecules are too slow to transport chemical signals under many conditions. As such, gas concentrations in real environments can easily become very difficult to predict both in space and time domain.

Just as an example, the plume generated by an H2 spill in a hydrogen-based car refilling station could move in rather unpredictable paths and the probability of a single solid-state chemical sensor to be hit by it with a significant concentration in a timely way could be considered negligible.

Recently, researchers began to tackle these scenarios with novel approaches that refer to two main classes. A first approach is based on the use of a moving detector. Together with appropriate modeling information these detectors can follow random paths, exploring a particular environment before being hit by a chemical plume [3], [4]. After that, by using chemical spills search algorithms, often biomimetic, they try to detect the source of contamination (source declaration problem). The other approach, basically rely on the use of multiple, low cost and autonomous distributed fixed detectors, that try to cooperate in reconstructing a chemical image of the sensed environment [1], [5]-[7]. Both approaches could maximize probabilities to detect and quantify appropriately the presence of hazardous or toxic gas in environment with different specificities.

Advantages of distributed approach are identifiable in flexibility, scalability, enhanced signal-to-noise ratio, robustness, and self healing. Several sensors node can be placed in different locations, each one with its own characteristics in terms of environmental conditions (air flow, temperature, humidity, different gas concentration, etc.) contributing to describe more thoroughly the environment in which they are embedded. Each smart chemical sensor, composing the distributed architecture has its own communication capabilities and its information is available for more than one client. The network can adapt itself to a variable number of chemical sensor improving reliability.

1530-437X/$26.00 © 2010 IEEE

If a sensor fails, the network can estimate its response on the basis of the previous behavior and of the response of the closest sensors while being able to self-heal the network structure by reconstructing routing trees [8].

Sensor calibration could be performed by temporarily adding other reference sensors or, thanks to data fusion techniques, using self recalibration strategies. Practically, the development of application specific algorithms could allow self calibration by exploiting networked cooperation in a totally unmanned fashion [9]. A distributed architecture means also that every sensing node has computing capabilities; it can provide diagnostic services and it can elaborate strategies, for instance, to improve the stability of its metrological characteristics. In particular this can be significant for implementing drift correction, accommodating a paramount problem in solid-state chemical sensing.

The characteristics of an optimal sensor for distributed chemical sensing should include low power operations capability, low cost, long-term reliability and stability; additionally, it should also be easy to integrate with simple signal conditioning schemes. Depending on the application, the sensor should possibly express good specificity properties and very low detection limits. By far, this depiction applies more to an ideal device than a real one, in fact, no current chemical sensor technology seems near to obtain such results simultaneously [10], [11].

In particular, the well known metal oxide chemical sensor technology (MOX) requires high working temperatures and, in fact, they are mostly coupled with micro-hotplates used to increase the temperature of the sensitive layer [12]. This represents a serious issue for their use in wireless chemical sensing motes because their average power consumption is in the range of 200-800 mW with continuous operation. Instead, polymer-based chemiresistor, resonators and mass sensors (QMBs, SAWs) are usually operated at room temperature [13]-[16]. Although they are not as common as their MOX counterparts, their low-power operation capability can be recognized as a huge advantage with respect to the other technologies especially when VOC (Volatile Organic Compunds) detection is concerned [17]. Unfortunately, they may not be as efficient at very low concentrations and most of them are significantly affected by humidity.

Since a decade, polymer/nanocomposites reactivity to chemicals have been also applied to the development of passive resonant sensors that are capable of wireless remote operations in a very simple way [18]. LED/Polymer-based optical sensing can be an extremely interesting solution for all applicatiosn where high limits of detection are not an issue because of very low cost, reliability, and very low-power demand [7], [1]. However, in order to obtain suitable sensitivity in most applications laser source should be used and this may increase significantly both costs and power needs. Practically, all solid-state sensors are widely affected by being not specific, this issue has been tackled with the widespread electronic nose design often turning it into an advantage for mixture classification applications [10].

Current e-nose are not designed to tackle distributed chemical sensing problem, however, during last years a number of novel approaches have been proposed and experimented. This field is, hence, rapidly evolving exploiting the plethora of results obtained by researchers in the wireless technology field. Bicelli et al. have investigated the use of commercial low-power

MOX sensors in a WSN network for indoor gas detection applications. Their primary focus was the analysis and reduction of power consumption in order to meet DCS criteria. They suggested a particular heating procedure in an attempt to achieve a significant increase in the wireless sensor battery life (about one year) with a sample period in the range of two min. Unfortunately, this resulted in increasing the actual response time [19].

Pan et al. realized a single w-nose for online monitoring of livestock farm odors integrating meteorological information and wind vector [5]; detection performances and power consumption have not been reported so it was not possible to estimate autonomous life expectation. Warmer et al. recently presented a four MOX sensors-based WCSN network for flammable detection in military docks, but power needs restrict the application to the availability of power mains [20].

During the last years, we have developed a wireless electronics nose platform, based on TelosB mote using the TinyOS operating system. The platform, makes use of four room temperature operating low-power polymeric chemical sensors that extends the entire sensory network lifespan.

In this work, on-board intelligence is implemented to provide the basis of sensor censoring strategy that allow for transmission of only informative data packets. We propose the use of such a strategy for achieving a dramatic reduction of the power needed for distributed chemical sensing networks further extending its lifespan. The target scenario for the application of the proposed methodology is the distributed monitoring of indoor air quality and, specifically, the detection of toxic/dangerous VOC spills.

In the following, a description of the architectural features of the proposed platform is provided focusing on the design and test of on-board intelligence for power management. An instrumental investigation on the performance of the overall intelligent power management methodology has been carried out.

II. Wireless Sensor Network Architecture

The developed platform is a wireless network consisting of spatially distributed autonomous devices that use sensors to monitor and detect the presence of organic vapors in air.

The modules were designed to operate in a mesh shaped network in order to pursue the final goal of building an olfactive image of a sensed environment by means of a distributed platform connected to a PC for data management purposes (Fig. 1).

A. Modules Sensors Array

The sensing board is based on four polymer-carbon black nanocomposites sensors (see Table I), a class of chemical detecting devices that have been already selected for electronic noses applications [13]-[18]. Their selectivity can be adapted via the use of the wide range of polymeric materials and conductive fillers available [14].

In our opinion, this class of sensing devices is particularly suitable for the use in distributed chemical sensing application for their interesting low power and rapid switch on characteristics. Their mechanism of response is, at its simplest level, based on swelling. When the isolating polymer film, within which conductive particles have been dispersed, is exposed to a particular

Fig. 1. UML Deployment diagram of the software architecture for the proposed platform.

TABLE I

Sensor Array Composition

Polymer Chemical Structure

Sensl Poly-(methylmethacrylate) [CH2C(CH3)(C02CH3)]n

(PMMA)

Sens2 Poly-(2hydroxy- H(NHCH2CH2)nNH2

ethylmethacrylate)

(PHEMA)

Sens3 Poly-(styrene) (PS) (C6H1003)n

Sens4 Poly-(ethylenimine) [CH2CH(C6H5)]n

linear(PEI)

vapor, it swells while absorbing a varying amount of organic vapors depending of polymer's type.

The swelling disrupts conductive filler pathways in the film by pushing particles apart and the electric resistance of the composite increases [16]. As a result of this room temperature operating mechanism, they exhibit very low-power consumption and, if properly conditioned, a nice stability. However, they suffer for poor selectivity and sensitivity.

In order to obtain a suitable voltage signal, i.e., showing proportionality to sensor resistance variation, a simple resistance to voltage converter signal conditioning system has been designed. Suitable choice of circuit parameters allow the proper operation of the board within a wide range of base resistance. In Fig. 2, the functional scheme of the board is shown.

The overall sensing subsystem is then directly connected to four A/D input ports provided by the core mote of the proposed platform, a commercial Crossbow TelosB mote.

B. Core Mote

The commercial Crossbow TelosB platform is a research oriented mote platform that has shown its operative potential over time [21]. Each TelosB node is powered by 2 AA batteries for standalone operations, while it can be powered by a standard USB interface for its usage as a datasink related module. The chosen platform is based on a Texas Instruments MSP430F1611 microcontroller which is a 16-bit RISC processor featuring 10 kB of RAM, 48 kB of flash, and 128B of

Fig. 2. Current absorbed by TelosB in its active operating phase, in particular, during the radio wakeup, sensing, and transmission phases in a complete cycle.

information storage. It operates at 1.8 V and allows for very low-power consumption states from which it can recover in few ms. It also has eight built-in (ADC) channels providing also DAC capabilities. The basic node is equipped with a set of commercial sensors among which the SHT11 humidity and temperature sensors that can be very useful in chemical sensing applications both as a primary information source and for multivariate sensor calibration. The embedded low-power radio, Chipcom© CC2420, represent the kernel for 802.15.4 protocols stack support.

Basically, the choice of this module as the core mote is based on very low-power sleep state of both MCU and radio sections, fast recovery, and the availability of community support with a wide number of educational-based users. Data transmission, in fact, is one of the primary power draining activity in WSN applications and the choice of a low-power transmission with low-power states and fast recovery is crucial.

C. DCSN Software Architecture

The heterogeneity of the practical applications and the limited energy availability required a careful design of the software (sw) structure and protocols for the management of the wireless e-nose platform. The possibility to focus primarily on domain specific optimization, led to the choice of the TinyOS architecture as a runtime support and operating system.

TinyOS is an open source operating system for WSN applications developed by the University of California at Berkeley [22]. Essentially, TinyOS's component library includes network protocols, distributed services, basic sensor drivers, data acquisition tools-all of which can be further refined for custom applications. Specific software modules realize basic functionalities control of the sensor node, like access microcontroller registers, writing and reading from EPROM, the interface radio management for transmission, and reception data. In particular, the-power management model of TinyOS allow for the automatic management of module subsystems switching among active, idle, and sleep phases. Often, provided that the custom sw design meets the

TinyOS requirements, the programmer can focus on specific optimization letting the runtime operate for general power saving options.

A C derived language, NesC, is the reference programming language for TinyOS-based programming. NesC relies on a component-based programming module. Users can expand basic TinyOS functionalities building their own components relying on NesC itself. The executable code generated by the NesC compiler includes only the TinyOS modules actually needed by the custom applications, minimizing memory footprint.

Fig. 1 shows an UML deployment diagram of the overall platform software architecture from which its three-layers design is clearly apparent. The first-layer encompasses all the embedded sw components that, being developed in NesC, can make use of TinyOS directives to provide on-board w-nose control and networking functionalities. In particular, the embedded component is responsible for pump ON/OFF operations control (optional, its presence depends on applications), sensors signal conditioning board control (ON/OFF switching), data acquisition and processing, and eventually data transmission and routing. At the second-layer, a PC-based component coded in Java capture data packets from a sensorless node that act as a datasink and is connected, via a USB connector, to the datasink PC. This component provides rebroadcasting features for distributing the data provided by the networked nodes over TCP/IP networks. The component has been developed by using as a core component the SerialForwarder class provided by the TinyOS standard installation. At the third level, multiple GUIs, also coded in Java, can provide visualization and recording features while remotely controlling relevant parameters for embedded sw operations (e.g., duty cycle parameters). Separate tabs provide visualization of raw data or pollutant concentration local estimations provided by each node. Actually, the overall architecture has been designed to host two possible pattern recognition and sensor fusion subsystems. The first interface defines methods to connect a sensors raw data processing component that will allow for local situation awareness. Actually, it is connected to the local sensor fusion component that allows for the local estimation of pollutants concentration by using a trained neural network algorithm. The second interface, should provide second-layer sensor fusion services, allowing for integrating estimation coming from the several deployed wireless e-noses. In this way it will be possible to cooperatively reconstruct an olfactive image of the environment in which they are deployed. The second-layer sensor fusion algorithm and its implementing component is currently under design.

D. Duty Cycle and Power Consumption

The proposed platform has been designed for continuous realtime monitoring of VOC with a programmable duty cycle including sensors data acquisition, processing, and transmission toward data sink. Duty cycle parameters and, in particular, the length of each phase and the sampling frequency are fully programmable by the application designer, sample frequency can be set dynamically at runtime by GUI. For this work, it has been set to 1 s.

In order to fully characterize the power consumption, we can easily separate the duty cycle in four separate phases each one having a different power need.

1) Sleep phase. In this phase, each node is put asleep. MCU and radio are turned to standby mode.

2) Sensing phase. Sensors driving, data acquisition, and ADC conversion is carried out.

3) Computing phase. Data processing is performed in order to prepare data to be transmitted towards datasink.

4) Transmission/Reception phase. The actual data transfer.

The whole embedded software has been designed and implemented to optimize the battery life by controlling the duration of each phase and the activation of sensors driving electronics.

At any instant, a single module power consumption can be computed as a function of its microcontroller power state and whether the radio, pump, sensor driving electronic are on, and what operations the active MCU sub-units are performing (analog to digital conversion). By using appropriate programming models, e.g., relying on split-phase operations, it is possible to best utilize the features provided by TinyOS to keep the power consumption to a minimum. In particular, as regards as radio stack management, we have chosen to rely on the LPL—Low-Power Listening algorithm. As such, node radio can be programmed to switch on periodically just long enough to detect a carrier on the channel. If a carrier has been detected, then the radio remains on long enough to detect a packet to be routed to the data sink for mesh shaped networking. After a timeout, the radio can be switched off.

III. Results

In this work, our goal is to present a sensor censoring (see [23]-[25]) power management strategy based on on-board sensor fusion component for long-term operation, real-time distributed chemical sensing with wireless electronic noses.

Our approach, is based on a simple sensor fusion component designed to locally detect chemical species presence, simultaneously estimating their concentration. In this way, the mote is able to assess if the data are informative and apply sensor censoring, i.e., to decide whether to transmit it or not. Alternative approaches, are based on data buffering (store and forward), i.e., sending data samples collection in burst, extending latency (see [24]). Unfortunately, in real-time chemical sensing this approach is not feasible for obvious reasons: sampled data, if deemed significant, should be transmitted as soon as possible. We wanted to assess the feasibility and potential benefits of using such an approach in distributed chemical sensing by measuring the actual power consumption of a sensor node in a star shaped networking configuration to assess the module operative lifespan.

The measurement setup is based on a Tektronix TDS 3032 digital oscilloscope set to measure Vshunt on 10 Ohm shunt resistance and to let us derive current flows (Fig. 2).

In its simplest configuration with sample frequency set at 1 Hz, the sensor node, as mentioned above, remains in sleeping mode for most of the time (Trs time interval). In this phase, a current Irs, measured as 12 /xA, is drawn. During the Twr wake up period, the sensor node turns on the chip radio awakening from the sleep mode. In this stage, we measured a 5 mA

Operation phase

Radio Sleep (RS) Radio Wakeup (RW) Acquisition (A) Computing (C) Transmission (T)

Time [ms]

Trs = 888 Trw — 7 Ta = 37 Tc = 25 Tt = 68

Current request [mA]

Irs=0,012 Irw—5

Ia=19 Ic=2,5 IT=18

TABLE III

Platform Signal Conditioning Board Consumption

Operation phase Operating condition

Time [ms] T0N= 30

Current request [mA]

Fig. 3. Acetic acid concentration estimation (red) performed by the FFNN component plotted against true concentration (blue). X axis depict time (samples), while y axis depicts real and estimated concentrations values.

TABLE II

Core Mote Consumption in Different Phases of the Duty Cycle

of such strategy is only made possible by the peculiar sensor response mechanism of the nonconductive polymer sensors. In fact, swelling is only negligibly affected by actual sensor power up so that sensors do not need warming up before their resistance could be sampled. Conversely, MOX sensors should be carefully driven to achieve their best response characteristics [19].

Usually, module lifespan can be estimated using the mean current consumption, namely Icc,mean, of the wireless sensor, considering batteries capacity, conversion efficiency, and power supply output voltage gain.

In their related work, Bicelli et al. suggest the use of a simple reference formula in order to compute Icc,mean by averaging current consumption over the multiple phases of a module duty cycle [19]. In order to compare results with that specific work, we set the sampling period T to one second, furthermore, initially, no sensor fusion take place and data are sent to datasink at the end each sampling cycle. The mean current value, /cc,mean, obtained under the TelosB operating conditions described above, can be computed according to (1) by using the current measured in the proposed setup

Icc,mean

— jRS ^RW —/A +

mean current draw. After that, the system sets up itself for data capture and conversion with a mean current of 19 mA. Even in this stage, provided that data are sent every acquisition cycle, radio activity is the main source of power consumption. In fact power consumption, in the active phase, is dominated by radio activity until switch off timeout inset; after ADC converter timeout, it can be measured as 18 mA. Eventually, the RF circuit is turned off again and the state change from "Transmission" mode to sleeping mode.

Fig. 3 shows the detailed evolution of the TelosB drawn current during acquisition, processing, and data transmission of sensors data, while Table II shows the TelosB consumption in terms of measured currents during each operating phase. Although in the operating conditions polymeric sensors power consumption is negligible, the overall electronic conditioning board, requires significant current draw. Actually, we have measured, under normal operating conditions, that the current needed for sensor array driving and signal conditioning board is equal to 38 mA (see Table III). The adopted experimental solution provides the implementation of solution for the control of power supply of the signal conditioning board. A digital signal obtained from a TelosB digital output drives a switch that lets the electronics power supply to switch on only during the time slice in which data capture take place. The implementation

Table II summarizes the results of performed measurements. According to it, Irs is the current drawn from the entire platform during sleeping phase. Circuitry elements that must be turned on during the acquisition phase (e.g., microcontroller, interface circuits, ADC, etc.) require a total Ia = 19 mA contribution; however, the radio, that must be switched on for communication purposes, account for the most of Ia current contribution during active phase.

In fact, assessment of the duration of Ta phase has been conducted in a previous step by avoiding transmission phase. The actual current drain in Ta phase is due to the simultaneous consumption of data capture and radio equipment. We found that the reported duration of Ta is mainly due to ADC timeout length. Furthermore, for our goal there is no need of assessing precisely when the data transmission occur.

Neglecting conversion efficiency, battery life (BL) can be computed as a function of the battery capacity C and total current draw. For our w-nose, C is equal to 3500 mAh, while total current is equal to the sum of mean current absorbed by conditioning board /6?mean (30/1000 * 38 = 1, 14 mA) and the ^cc,mean (1.97 mA), related to the core mote, computed using (1). Expressing BL as the ratio C/(/cc?mean + h,mean), we can estimate that the overall e-nose battery life, with sampling and transmission frequency of 1 Hz, is roughly 47 days. In order to comment on comparison of our results with those reported by literature, we should take into account that this platform is equipped by a small array of sensors, while typical results are reported for single sensors wireless nodes. Anyway, extending the sample period to 10 s, hence loosing real-time characteristics, this basic setting will account for a battery life in excess of one year.

A. Power Saving Using Sensor Censoring

As previously mentioned, a local sensor fusion component can classify a sampled sensor response pattern to assess the

significance of the sampled data. The further power consumption needed by these computational activities should be taken into account to evaluate a tradeoff that depends basically on the rate of significant events occurrence. Actually, in most chemical sensing scenarios, the probability p of a significant event to occur (e.g., chemical spills) is expected to be very low, while the timely transmission of relevant data is needed for rapid situational awareness in security application or for multiple mote data fusion at data sink. These constraints do not permit the relaxation of the latency parameter. It is also important here to recall that sensor fusion at mote level is needed for the intrinsic nonspecificity of solid-state chemical sensors that will make threshold-based approach almost useless both for chemical identification and quantification.

In order to perform an experimental check of the exposed concepts in distributed gas sensing scenarios, we have designed and implemented an ad hoc lab scale experiment. We have designed and developed a sensor fusion component that have been trained for a distributed pollutant detection application with wireless e-noses. Deployment network topology was star shaped. Hence, we have evaluated the power savings obtained by using sensor censoring in this common application scenario. Actually, we assumed a general chemical sensing problem characterized by the presence of two pollutants whose toxic/dangerous concentration limits are different. To save battery energy, the single mote should be able to decide whether to transmit or not the sampled data on the basis of the concentration of the two gas estimated by sensor responses. The e-noses sensor arrays have been exposed to four concentrations levels of Acetic Acid and Ethanol in a controlled environment setup and their responses, sampled by the motes, have been recorded to build a suitable data set. Concentration levels have been chosen to be [225, 450, 900, 1800] ppm for Acetic Acid and [200, 500, 1000, 2000] for Ethanol, while relative humidity was fixed to 30%. Five complete exposure/purge cycles have been conducted and instantaneous response to base response ratio of each sensor was used as descriptive features. We have designed and implemented a three-layer feedforward neural network (FFNN) component using NesC. FFNN architecture have been chosen on the basis of considerations about its peculiar characteristics. In fact, FFNN shows a high degree of flexibility and a high capacity in pattern recognition sense. Their computational footprint, at least in the operational phase, is relatively small; furthermore, their highly compact internal knowledge representation paradigm allow for a very limited memory footprint. The component has been linked to the platform embedded software architecture so to run on-board. The neural network component design foresee four input neurons, ten hidden layer neurons, and two output neurons whose outputs indicate concentrations of the simulated pollutants. The network component has been then loaded with networks weights obtained by training an identical model in MATLAB with the recorded dataset, to reach a reasonable point-to-point real-time estimation of concentrations of different analytes, thus simultaneously solving a classification and regression problem.

Practically, normalized steady-state response have been used as training set, while the remaining samples have been used for validation purposes. By choosing this particular training dataset

Fig. 4. Ethanol concentration estimation (red) performed by the FFNN component plotted against true concentration (blue). X axis depict time (samples), while y axis depicts real and estimated concentrations values.

TABLE IV

Memory Footprint Increase in the Embedded Component Resulting From the Linking of Sensor Fusion Neural Component

split, we set apart a significant number of test samples; on the other hand, we can expect significant errors to occur in the estimation of transient samples. Since estimation error is averaged on all samples, transient cause a significant hit on the chosen synthetic performance estimators. In this laboratory-based setup, the proposed architecture performance was evaluated by dividing validation set Mean Absolute Error (MAE) by analyte concentration range span, obtaining, despite the above mentioned setup, a 6% (standard deviation: 10%) value for Acetic Acid concentration estimation problem and 11% (standard deviation: 11%) for the Ethanol problem (see Fig. 4). Preliminary performance operative tests with a simpler architecture were also conducted and results were reported in [26], anyway significant performance boosting, in particular, as regarding dynamics, could be achieved by using a two-stage detector/quantifier approach with tapped delay architectures as reported in [27].

In this paper, however, we focus on sensor censorship strategy transmission versus processing tradeoff that would be hampered by sensor fusion component complexity. For this reason, assessment of the memory footprint increase, resulting from the implementation of the local sensor fusion approach, has been conducted and results has been reported in Table iV. Approach, has been conducted and results has been reported in Table iV.

As regards to the computational footprint impact, we can expect several hyperbolic tangent function calls as well as multiplications and sums, the number of which depending only on the particular network architecture. For the current four hidden layer neurons architecture these figures are reported in Table V.

Following a measurement approach, the execution of the NN sensor fusion component account for additional 2.5 mA con-

TABLE V

Computational Complexity in Terms of Functional Calls

sumption over a total time span of 25 ms (function call overhead included).

In order to reassess total power consumption, we write (1) again by taking into account computing phase

TRW + TA + TT + Tc\

1cc,mean

+ + (2)

This equation provides the mean current supplied by the batteries in case of significant event occurrence: the mote perform signals conditioning and data sampling, data fusion by means of NN and data transmission every T seconds.

Using a Bernoulli random variable X ~ B(l,p) which would model the result of the computation performed by the NN to verify whether to transmit the acquired data on the radio channel, (2) can be rewritten as

Icc,mean

where , are, respectively

/tV _ rri rri r~fl rri rri

1rs — 1 ~ ÍRw - -La — J-t - J-c n* = T-TA-TC

By exploiting equation (3), we can finally discuss the advantages of NN-based sensor censoring for the implementation of power saving strategies in the proposed architecture.

Worst case is obtained when p = 1, i.e., when all samples refers to significant events, then the mean current computed using (3) is obviously greater than the one calculated by (1), i.e., i^Jnean = 2.013 mA. However, by equating (1) and (3) and selecting p as independent variable, we obtain

p = 0.97.

The obtained value represents the percentage threshold of significant event under which the NN-based sensor censoring become more efficient as regards as power management. This computed threshold level make the use of the proposed approach feasible for most of the analyzed distributed chemical sensing scenarios.

In industrial chemical spills monitoring, even considering false positive generation, it is reasonable to expect values of p that reflect a number of only a few (<100) significant samples a day. This is true even in the case of performing only rough local concentration estimations and being very conservative in

the censoring criteria. Censoring criteria, i.e., spiking threshold, should be chosen exploring the tradeoff among sensitivity and false positive rate considering the danger/toxicity level of the target gas. With this regard, danger/toxicity levels should, of course, be taken into account also during training selecting an appropriate interval that match working conditions. The nature of second-level sensor fusion subsystems that collect data from the different networked nodes also has significant implications in the threshold choice. Just as an example, a second-level sensor fusion subsystem may decide to initially ignore time isolated detection reported from a single node waiting for a number of received spike to accumulate before declaring an emergency or tagging the node as malfunctioning.

For an experimental evaluation, using the above mentioned setup and allowing a small slack to the network estimation of ethanol (detection threshold on NN response = 10 ppm w.r.t. a 2000 ppm max exposure level) we obtain a false positive rate of 5% for Ethanol and less than 1% for Acetic Acid. In case of no positive events during node battery lifetime, a reasonable expectation in safety oriented leakage detection scenario, a proposed node can be expected to experiment a maximum operative lifespan that is very near to the intrinsic limit now dominated by the power needed by the signal conditioning board. In particular, considering a p = 0.01, the expected lifetime computed by taking into account (3) reach 113 days (110 days for a p = 0.05), that is a rather interesting value for a four sensor wireless electronic nose with real-time operating characteristics.

IV. Conclusion

We have developed a distributed chemical sensor architecture based on intelligent electronic nose motes that relies on room temperature operating solid-state polymer sensors. The proposed w-noses have proven capable, if properly trained in lab conditions, to locally infer chemicals presence and their specific concentrations. This capability can be used for the implementation of sensor censoring strategies, i.e., to decide on sample data significance, so to improve power management by simultaneously avoiding unneeded transmissions and retaining low latency characteristics.

Influence on mean current drawn by the on-board sensor fusion component have been modeled and discussed in different scenarios characterized by different event frequency.

By using a measurement approach we have demonstrated that this capability can allow for remarkable extension of network operating lifespan in real-time distributed chemical sensing applications that are currently strongly limited by both sensor arrays and data transmission power needs. However, sensors driving and signal conditioning stage are currently in a prototype stage so we expect a significant reduction in power needs in the following releases that could allow the proposed platform to reach the one year target lifespan.

References

[1] C. Grimes, K. G. On, O. K. Varghese, X. Yang, G. Mor, M. Paulose, E. C. Dickey, and C. Ruan, "A sentinel sensor network for hydrogen sensing," Sensors, vol. 3, pp. 69-82, 2003.

[2] T. Nakamoto and H. Ishida, "Chemical sensing in spatial/temporal domains," Chem. Rev., vol. 108, no. 2, pp. 680-704, 2008.

[3] A. J. Lilienthal, A. Loutfi, and T. Duckett, "Airborne chemical sensing with mobile robots," Sensors, vol. 6, pp. 1616-1678, 2006.

[4] A. J. Lilienthal and T. Duckett, "Building gas concentration gridmaps with a mobile robot," Robot. Autom. Syst., vol. 48, no. 1,1, pp. 3-16,

[5] L. Pan, L. Rui, P. Shanghong, Y. Chai, and S. X. Yang, "A wireless electronic nose network for odors around livestock farms," in Proc. 14th Int. Conf. Mech. Mach. Vision in Practice, M2VIP 2007, Xiamen, 2007, pp. 211-216.

[6] S. De Vito, E. Massera, G. Burrasca, A. Di Girolamo, M. Miglietta, G. Di Francia, and D. Della Sala, "TinyNose: Developing a wireless e-nose platform for distributed air quality monitoring applications," in Proc. IEEE Sensors Conf., Lecce, Italy, Oct. 26-29,2008, pp. 701-704.

[7] R. Shepherd, S. Beirne, K. T. Lau, B. Corcoran, and D. Diamond, "Monitoring chemical plumes in an environmental sensing chamber with a wireless chemical sensor network," Sens. Actuators B Chem., vol. 121, no. 1, pp. 142-14, 2007.

[8] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, "Wireless sensor networks: A survey," Comput. Networks, vol. 38, pp. 393-422, 2002.

[9] W. Tsujita, A. Yoshino, H. Ishida, and T. Moriizumi, "Gas sensor net-workfor air-pollution monitoring," Sens. ActuatorsB: Chem., vol. 110, pp. 304-311, 2005.

[10] K. Arshak, E. Moore, G. M. Lyons, J. Harris, andS. Clifford, "Areview of gas sensors employed in electronic nose applications," Sensor Rev., vol. 24, no. 2, pp. 181-198, 2004.

[11] D. Diamond, S. Coyle, S. Scampagnani, and J. Hayes, "Wireless sensor networks and chemo-/biosensing," Chem. Rev., vol. 108, no. 2, pp. 652-679, 2008.

[12] N. Barsan and U. Weimar, "Conduction model of metal oxide gas sensors," J. Electroceram., vol. 7, no. 3, pp. 143-167, 2001.

[13] L. Quercia, F. Loffredo, M. Bombace, I. Nasti, and G. Di Francia, "Effects of thin film polymer composite preparation on vapor sensing devices characteristic," Sens. Actuators B, vol. 111-112, pp. 166-170,

[14] L. Quercia, F. Loffredo, B. Alfano, V. La Ferrara, and G. Di Francia, "Fabrication and characterization of carbon nanoparticles for polymer based vapor sensors," Sens. Actuators B, vol. 100, pp. 22-27, 2004.

[15] T. J. Plum, V. Saxena, and J. R. Jessing, "Design of a MEMS capacitive chemical sensorbased on polymer swelling," in Proc. IEEE, Workshop of Microelectronics Ad Micro Devices, Boise, ID, Apr. 14, 2006, pp. 49-50.

[16] E. J. Severin, B. J. Doleman, and N. S. Lewis, "An investigation of the concentration dependence and response to analyte mixtures of carbon black/insulating organic polymer composite vapor detectors," Anal. Chem., vol. 72, pp. 658-668, 2000.

[17] S. C. Ha, Y. Yang, Y. S. Kim, S. H. Kim, Y. J. Kim, and S. M. Cho, "Environmental temperature independent gas sensor array based on polymer composite," Sens. Actuators B, vol. 108, pp. 258-264, 2005.

[18] H. Yoon, J. Xie, J. K. Abraham, V. K. Baradan, and P. V. Ruffin, "Passive wireless sensors using electrical transition of carbon nanotube junctions in polymer matrix," Smart Material and Structures, vol. 15, no. 1, pp. 14-20, 2006.

[19] S. Bicelli, A. Depari, G. Faglia, A. Flammini, A. Fort, M. Mugnaini, A. Ponzoni, V. Vignoli, and S. Rocchi, "Model and experimental characterization of the dynamic behavior of low-power carbon monoxide MOX sensors operated with pulsed temperature profiles," IEEE Trans. Instrum. Meas., vol. 58, no. 5, pp. 1324-1332, May 2009.

[20] C. Becher, P. Kaul, J. Mitrovics, and J. Warmer, "The detection of evaporating hazardous material released from moving sources using a gas sensor network," Sens. Actuators, B: Chem., vol. 146, no. 2, pp. 513-520, Apr. 29, 2010.

[21] J. Polastre and R. Szewczyk, "Telos: Enabling ultra-low power wireless research," in Proc. Int. Symp. Inform. Process. Sensor Networks, 2005, pp. 364-369.

[22] TinyOS. [Online]. Available: http://www.tinyos.net

[23] E. Y. Chang and A. Jain, "Adaptive sampling for sensor networks," in Proc. 1st Int. Workshop on Data Management for Sensor Networks, DMSN 2004, Toronto, Canada, Aug. 30, 2004, vol. 72, pp. 10-16.

[24] C. Rago, P. Willett, and Y. Bar-Shalom, "Censoring sensors: A low-communication rate scheme for distributed detection," IEEE Trans. Aerosp. Electron. Syst., vol. 32, no. 2, pp. 554-568, Apr. 1996.

[25] S. Appadwedula, V. V. Veeravalli, and D. L. Jones, "Decentralized detection with censoring sensors," IEEE Trans. Signal Process., vol. 56, no. 4, pp. 1362-1373.

[26] S. De Vito, G. Burrasca, E. Massera, M. Miglietta, and G. Di Francia, "Power savvy wireless e-nose network using in-network intelligence," in Proc. 13th ISOEN, AIP Conf., vol. 1137, pp. 211-214, ISBN: 978-07354-0674-2.

[27] S. De Vito, A. Castaldo, F. Loffredo, E. Massera, T. Polichetti, I. Nasti, P. Vacca, L. Quercia, and G. Di Francia, "Gas concentration estimation in ternary mixtures with room temperature operating sensor array using tapped delay architectures," Sens. Actuators B: Chem., vol. 124, no. 2, pp. 309-316, 2007.

Saverio De Vito received the M.S. degree in informatics engineering from the University of Naples, "Federico II," Naples, Italy, in 1998.

During 1998 and 1999, he was a Research Fellow at the Artificial Vision and Intelligent Systems Laboratory, University of Naples "Federico II," working on breast cancer computer-aided diagnosis. From 1999 to 2004, he worked in ICT Engineering as a R&D Project Manager in the framework of satellite-based telemedicine, earth observation, and distance learning. In June 2004, he joined ENEA, as a Researcher. He is currently based in Portici Research Center. His research interests include statistical pattern recognition, artificial olfaction, wireless sensor networks, and computer-aided diagnosis. Since 2005, he is a Contract Professor of Applied Informatics at the University of Cassino.

Paola Di Palma received the M.S. degree in electrical engineering from the University of Cassino, Cassino, Italy, in 2009.

She was a Research Fellow at CNR-ISAFOM and ENEA pursuing research on artificial olfaction and wireless sensor networks with applications to distributed olfactive nuisance detection, indoor air quality, and power management in wireless sensor networks. She is now a system designer for an engineering firm specialized in photovoltaic systems.

Carmine Ambrosino received the M.S. degree in informatics engineering from University of Sannio, Benevento, Italy, in 2007.

During 2007 and 2008, he was a Fellow of GRACE Laboratory, University of Sannio, working on control automation systems for automotive applications. During 2009, he was Fellow at CNR-ISAFOM and ENEA pursuing research on artificial olfaction and wireless sensor networks with applications to distributed olfactive nuisance detection, indoor air quality, and power management in wireless sensor networks. He is now with CNR-ITIA developing software components for researches in control automation for manufacturing systems.

Ettore Massera received the Degree in physics from the University of Naples, "Federico II," Naples, Italy, in May 1997.

He has been working at the ENEA Research Center, Portici (NA), Italy, since June 2003. At present, he is in charge of research activity on gas sensor devices based on nanostructured materials. Previously, he worked on the study of thermal and optical properties of porous silicon at the Department of Physics, University of Naples.

Gianbattista Burrasca graduated with a Degree in electronics engineering from the University of Ancona, Ancona, Italy, in 2000. His thesis was on electromagnetic diffraction problems.

In March 2001, he began activities in photovoltaic systems design and electronic circuits design for photovoltaic applications at a private company. Since October 2004, he has been a Researcher at ENEA. From 2004 to 2009, he was focusing its activities on Ink-Jet techniques for functional material's deposition and embossing techniques. Furthermore, he also worked on electronic interfacing for sensors finalized to electronic nose development. Since 2009, he has been involved in electronic systems development for testing of heat exchange components in nuclear fusion reactors.

Girolamo Di Francia received the Degree in physics from the University of Naples "Federico II," Naples, Italy.

In 1985, he started his research activity in the field of fabrication and characterization of semiconductor solar cells (c-Si, GaAs), formerly in the Ansaldo Company, Genova, and then in the ENEA Research Center of Rome, where he was appointed Researcher in 1988. In 1991, he joined the ENEA Research Center of Naples where, starting from 1992, he investigated porous silicon-based devices. In 1996, he established there the Gas Sensor Laboratory mainly devoted to the fabrication and characterization of devices based on nanomaterials and on polymers nanocomposites.

Mara Miglietta received the B.S. degree in chemistry from the Universita degli Studi di Bari, Bari, Italy, in 1997.

In 1999, she joined ENEA as a Researcher working on environmental monitoring and on liquid and solid-state NMR (HR-MAS, CP-MAS) of foodstuffs and polymers for food packaging. Since 2007, she has been working in the fields of nanotechnologies, specifically on polymer nanocom-posites and conductive polymers for VOC sensing applications and physico-chemical characterization

of nanomaterials.