Teaching Old Sensors New Tricks: Archetypes of Intelligence

Transcription

1 IEEE SENSORS JOURNAL 1 Teaching Old Sensors New Tricks: Archetypes of Intelligence Diosthenis Karatzas, Arsenia Chorti, Neil M. White, Christopher J. Harris Abstract In this paper a generic intelligent sensor software architecture is described which builds upon the basic requireents of related industry standards (IEEE 1451 and SEVA BS- 7986). It incorporates specific functionalities such as real-tie fault detection, drift copensation, adaptation to environental changes and autonoous reconfiguration. The odular based structure of the intelligent sensor architecture provides enhanced flexibility in regard to the choice of specific algorithic realizations. In this context, the particular aspects of fault detection and drift estiation are discussed. A ixed indicative/corrective fault detection approach is proposed while it is deonstrated that reversible/irreversible state dependent drift can be estiated using generic algoriths such as the EKF or on-line density estiators. Finally, a parsionious density estiator is presented and validated through siulated and real data for use in an operating regie dependent fault detection fraework. Index Ters intelligent sensor, software architecture, drift estiation, fault detection, data fusion, density estiation, adaptability, reliability, calibration I. INTRODUCTION THE TERM intelligent when used to qualify a sensor has at best a dubious eaning. Intelligent or sart sensors traditionally refer to sensors offering additional functionality provided by the integration of icroprocessors, icrocontrollers or application-specific integrated circuits (ASICs) within the sensing eleent itself. In any case, the exact functionality that would qualify a sensor as intelligent is not as yet strictly defined. As a result, there are any flavors of sensor intelligence available in the literature, ranging fro ipleentations including sensors coupled with application dependent electronics [1], [], [3], to Self-Validating (SEVA) sensors [4], [5] which place an ephasis on their ability to validate their output. In parallel, the issue of intelligent sensors has been approached fro a functional point of view in [6] and [7]. This work concentrates not so uch on how data is locally processed but on which type or results should be provided by an intelligent sensor in ters of available services. To address intelligent sensors, the IEEE introduced the IEEE 1451 faily of standards [8], [9], [1], [11]. IEEE 1451 provides a foral definition of the basic requireents of a sart transducer. These standards define the fundaental eleent of intelligence as the ability to self-identify, and copleent this with on board eory and a set of digital, analogue and ixed counication interfaces. IEEE 1451 has Manuscript received June 15, 6; D. Karatzas, A. Chorti, N.M. White and C.J. Harris are with the Electronic Systes Design Group, Departent of Electronics and Coputer Science, University of Southapton, SO17 1BJ, UK (e-ail: dk3@ecs.soton.ac.uk, ersi.chorti@gail.co, nw@ecs.soton.ac.uk, cjh@ecs.soton.ac.uk) been received well by industry, with a nuber of copliant products and ipleentations [1], [13]. While IEEE s intelligent sensor provides basic functionality, the SEVA sensors approach addresses the issue of intelligence at a higher level suggesting that an intelligent sensor should be able to validate its output and counicate sensor state inforation to a higher level. The SEVA approach recently becae a British Standard (BS-7986) [14] which has already been endorsed by parts of industry. BS-7986 defines a set of state values and couples these with rules for the propagation of the validated output and uncertainty values through any odules coprising the intelligent sensor. It is iportant to note two key facts, BS-7986 avoids the need to specify how the validated values and uncertainties are actually calculated or when specific sensor states should be entered into, since this functionality is deeed to be application-specific. In addition, it does not ake any assuptions regarding the counication interface nor it provides any standard way to identify the sensor. On the other hand, the IEEE 1451 standard defines this lower level functionality but avoids stepping into the real of higher level data processing. As such the two standards are essentially copleentary, but even so they still fail to propose a specific fraework for higher-level on-board data processing. This paper fills this gap by addressing a ultitude of coon situations encountered in sensory applications through a generic software architecture for intelligent sensors. It is argued that an intelligent sensor should be able to perfor on board signal processing within the sensor s software in order to produce the optial output signal in a variety of adverse circustances. The software architecture proposed provides a generic fraework to ipleent BS-7986, whilst building upon the requireents introduced by IEEE 1451 to incorporate the following functionality: Real-tie fault and drift detection, fault isolation and signal conditioning. Counication of sensor condition to the sensor anageent level. Adaptation to environental changes. Autonoous reconfiguration (where possible) to continue effective sensor operation despite sensor degradation. The key contribution of this paper is the presentation of the intelligent sensor software architecture. The architecture coprises odules which address coon issues of real world applications such as fault detection and drift estiation. Through a odular approach, the architecture can be easily adapted to fit specific applications, while retaining copatibility with industry standards. In addition to presenting the

2 IEEE SENSORS JOURNAL Intelligent Sensor Priary Measurand R Environent T H Sensor Interface 1 Sensor Interface Sensor Interface 3 Sensor Interface n Sensor Interface 1 Sensor Interface Sensor Interface n Sensor Interface 1 Sensor Interface Sensor Interface n Fault VV 1 Detection () Fault VV Detection () Fault VV 3 Detection () Fault VV n Detection () Environent Measurand #1 Fault VV 1 Detection () Fault VV Detection () Fault VV n Detection () Environent Measurand # Fault VV 1 Detection () Fault VV Detection () Fault VV n Detection () Coon Modality Internal Fusion Coon Modality Internal Fusion Coon Modality Internal Fusion VV IF Drift Evaluation and Copensation () Sensor Model Provider VV IF VV IF VV DEC Control Module IEEE 1451 TEDS Tiing IEEE Sensor Identification IEEE 1451 Counications Module Messaging Counications Interface (IEEE 1451) VV VVstatus VU VUstatus VDstatus Fig. 1. Intelligent sensor software architecture intelligent sensor software architecture, we discuss in detail possible algorithic approaches for the realization of the key coponents: fault detection, drift estiation and sensor odeling. The paper is organized as follows: an overview of the intelligent sensor software architecture is given in section II, along with a detailed description of the coprising software odules in subsections II-A to II-F. Section III offers a discussion on existing approaches to fault detection and proposes an algorith based on an Adaptive Kalan Filter (AKF) enhanced with outlier identification capabilities. In section IV, we analyze possible approaches for drift estiation and deonstrate the identification of reversible state dependent drift using the Extended Kalan Filter (EKF) algorith. We exaine the possibility of using on-line density estiators for irreversible state dependent drift estiation and present a theoretical analysis for ultiplicative drift in sensors that include internal oscillatory systes as a result of phase noise. Finally, in section V we deonstrate the use of a sparse density estiator, while section VI concludes the paper. II. INTELLIGENT SENSOR SOFTWARE ARCHITECTURE We consider an intelligent sensor to be a syste of priary sensing eleents and associated software odules acting as a single entity. The software architecture addresses the following issues: Specifies the functionality of the software odules. Defines how the odules are interconnected. Describes how data propagate between the odules. Explains how additional, specialized software odules can be incorporated to tackle application specific tasks. The software architecture supports input fro ultiple sensing eleents in different configurations, prooting redundancy of inforation through the use of sensor arrays for each of the quantities being easured. An overview of the odular structure is presented in Fig. 1 to deonstrate the key eleents of the architecture. These are explained in ore detail in subsequent subsections. In the configuration of Fig. 1, the quantity of interest (priary easurand) is easured by an array of sensors. Each sensor is interfaced with the software architecture through a Sensor Interface (SI) odule (subsection II-A), which is responsible for counicating with the hardware and perforing basic signal conditioning. The output of each SI odule is individually onitored by a Fault Detection (FD) odule (subsection II-C), which assesses each new easureent to produce an associated uncertainty value. The outputs of the FD odules in the array are subsequently cobined by an Internal Fusion (IF) odule (subsection II-D) and the single value and associated uncertainty is fed to the Drift Estiation and Copensation (DEC) odule (subsection II-E), which onitors the behavior of the signal over longer tiescale in order to identify and copensate for drift or bias.

3 KARATZAS ET AL.: TEACHING OLD SENSORS NEW TRICKS: ARCHETYPES OF INTELLIGENCE 3 Both the FD and the DEC odules generally use additional inforation about the sensor in ters of precalculated sensor odels. A bank of such odels is provided by the Sensor Model Provider (SMP) odule (subsection II-B). Depending on the algorith eployed by the FD or the DEC odule, the theoretical odel could refer to different types of sensor odels. Particular applications of a Kalan Filter (KF) based FD odule and an Extended Kalan Filter (EKF) based DEC odule will be described later on in sections III and IV respectively. In addition, it is possible that the selection of the theoretical odel is also dependent on the particular environental conditions (regie) within which the intelligent sensor operates. Generally, the environent can be onitored by one or ore arrays of sensors (two in the scenario of Fig. 1) easuring various environental attributes. Modules such as the SMP can use this inforation to facilitate the selection of the appropriate theoretical odel at any given tie. A parsionious density estiator is proposed in section V as a viable solution for the calculation of pdf odels. Each attribute onitored by the intelligent sensor is ipleented by a branch of odules. The Intelligent Sensor Control (ISC) odule (subsection II-F) is the final recipient of all inforation fro the odule branches and is responsible to counicate the final sensor output to higher level processes through an IEEE 1451 copliant counications interface. There are various advantages of having a odular software architecture. Each odule is self-contained; it is therefore trivial to exchange a specific odule with another of the sae functionality without affecting the overall design. The architecture specifies the functionality and the interconnections for each odule type, but does not dictate the use of a specific algorith to achieve the desired effect. The exact algorith used, is application-dependent and the end user will be able to either choose fro a library of alternatives, or produce an application specific ipleentation. Another advantage of the software architecture is that it allows odules to be cobined in a nuber of alternative ways. Two possible ipleentations of the architecture are shown in Fig. for illustration purposes. In ters of data propagation, the software architecture is fully copliant with standard BS All software odules in the data pipeline counicate inforation using a data structure coprising five pieces of inforation as required by BS-7986: i Validated value (V V ). This is the easureent possibly after certain corrections have been applied to it. ii Validated value status (V V status ). A byte (8-bit) value indicating how the associated V V data has been generated. iii Validated uncertainty (V U). Error band of the associated V V data value at 95% level of confidence represented in the sae units and with the sae precision as the V V. iv Validated uncertainty status (V U status ). A byte value indicating how the associated V U data has been generated. v Validated device status (V D status ). A byte describing the status of the odule reporting the above values (e.g. aintenance inforation and hardware probles). The intelligent sensor software architecture ipleents the transition decision rules for V U status and V D status, as well as T Environent T Fig.. Sensor Interface Sensor Interface 1 Sensor Interface Intelligent Sensor Priary Measurand Fault VV Detection () Environent Measurand Fault VV 1 Detection () Fault VV Detection () Sensor Interface n Fault VV n Detection () Environent T Drift Evaluation and VV DEC Copensation () Coon Modality Internal Fusion (a) Intelligent Sensor Priary Measurand Sensor Interface 1 (b) VV IF Control Module Control Module Counications Interface (IEEE 1451) Counications Interface (IEEE 1451) VV VVstatus VU VUstatus VDstatus Possible ipleentations of the intelligent sensor architecture VV VVstatus VU VUstatus VDstatus the rules for deriving V V status as specified in the standard. In general, the architecture assues V V and V U to be vectors of values. Software odules can be viewed as BS-7986 function blocks with the exception of the SMP odule, which provides inforation outside the ain data-pipeline. In the case of the SMP odule, the output refers to a theoretical sensor odel and the uncertainty value to the probability that this odel is valid for the current regie. BS-7986 function blocks are defined as software functional units. The function block output value (V V out ) is derived fro the function block input(s) by applying the algoriths ipleented in the block by its designer. So in the general case: V V out = f(v V 1, V V,..., V V N ) (1) where V V out is the function block output at the corresponding saple, f( ) is the function used to derive V V out fro the inputs to the function block and V V i, i = 1... N is the ith input to the function block. The function block output uncertainty value is then derived fro the inputs and their uncertainties by applying (): [ ] N U(V V out ) f = U(V V i ) () V V i i=1

4 4 IEEE SENSORS JOURNAL where U(V V out ) is the uncertainty associated with V V out at the corresponding saple tie, U(V V i ) is the uncertainty f associated with V V i and V V i is the sensitivity coefficient describing how V V out varies with changes in V V i. The intelligent sensor software architecture odules aintain copatibility with the above notion of function blocks, with the exception of the SMP as explained above. An additional feature of the architecture is support for counication between odules which is achieved through a essaging echanis. Although essaging is supported by all odules by default, its use is not necessary for ipleenting the basic functionality of the architecture. Nevertheless, it is iportant for such a echanis to exist, in order to ipleent ore advanced functionality, e.g. in the context of regie change detection. The types of software odules that coprise the intelligent sensor software architecture are described in detail in subsections II-A to II-F. The functionality of each odule is defined, along with the standard ipleentation and possible ipleentation alternatives where applicable. A. Sensor Interface Module (SI) The SI odule is responsible for interfacing the sensing eleent to the software architecture. As such, the possible ipleentations of this odule are as any and varied as the sensing eleents available. In order for a particular sensing eleent to be supported by the architecture the single requireent is that a SI ipleentation is provided for it. The basic functionality of the SI odule is to: Counicate with the sensing eleent hardware. Obtain easureents on deand. Perfor basic signal processing (linearization, A/D conversion, conversion to engineering units, etc.). The standard ipleentation of the SI odule reports the sensor s accuracy (known fro the technical characteristics) as the uncertainty value (V U). Subsequent odules (such as FDs) can ignore this inforation and produce their own. Nevertheless, in the absence of such subsequent odules, this is the best uncertainty that can be associated with the easureent at this level. B. Module (SMP) The SMP odule acts as a repository of pre-calculated sensor odels. The basic functionality of the SMP odule is to store and counicate pre-calculated odels for a specific sensor (or range of sensors), given a odel type. Although a theoretical sensor odel is sensor-specific inforation, it has been a design decision not to include this inforation in the SI odule for two reasons. First, it is a coon situation to have an array of sae sensors (which share the sae odels) in the design. Including sensor odel inforation in each instance of the SI would be superfluous. Instead a single SMP odule can serve any nuber of subsequent odules. Second, there are cases where the sensor odels refer to a virtual sensor 1 as opposed to a real one. An 1 The virtual sensor refers to the theoretical syste that has the sae acroscopic properties and characteristics with the array/set of the real sensors. exaple of such a situation will be discussed later on in the context of drift estiation and correction. In such situations an SMP odule can still act as a odel repository. The definition of a sensor odel is intentionally left open to ean any set of inforation that can describe the operation of the sensor. Three odel types are already defined: i Linear State-Space odels (intended to be used with KF based ipleentations of software odules). ii Non-linear State-Space odels (intended to be used with EKF based ipleentations of software odules). iii pdf odels (intended to be used with probabilistic hypothesis based ipleentations of software odules). Models are nevertheless not restricted to the above types, and users can enrich the list of odel types with their own. The selection of a specific odel in the standard ipleentation is based on a single attribute; the odel type. The standard ipleentation does not dictate a specific odel selection algorith in the case where ore than one odel is available. Derived ipleentations, of course, can ake use of additional inforation and elaborate algoriths to facilitate odel selection. Such extra inforation can be, for exaple, environental regie data as discussed in section V. C. Fault Detection Module (FD) The FD odule is responsible to assess and correct the incoing data and produce an uncertainty value. The basic functionality of the FD odule is defined as: Produce an uncertainty value for each input data value. Correct incoing data if possible. Copliant to the BS-7986 standard, the uncertainty value denotes the error band of the associated data value at 95% level of confidence and is represented in the sae units and with the sae precision as the data value. The uncertainty value is typically calculated by assessing the new data value in the context of recent data history. Specific algoriths for achieving that are discussed later on in section III. D. Internal Fusion Module (IF) The IF odule enables the configuration of sensors into arrays. It cobines an arbitrary nuber of input data values and associated uncertainties to produce a single sensory output and uncertainty value. Additionally, the IF odule is able to identify cases when an input value is inconsistent, and take action to report the proble. The functionality of the IF odule is suarized to the following: Generation of a single value and associated uncertainty fro ultiple inputs. Filtering of inconsistent values. The standard ipleentation of the IF odule is based on the Optial Weight Measureent Fusion (OWMF) algorith [15]. Let V V i be the ith input value of the IF odule and V U i its uncertainty value at 95% level of confidence. Assuing that the input values are drawn fro the Gaussian distributions N(V V i, σ i ), where σ i = V U i, a typical ethod to perfor

5 KARATZAS ET AL.: TEACHING OLD SENSORS NEW TRICKS: ARCHETYPES OF INTELLIGENCE 5 data fusion is by cobining the inputs based on a iniu ean square error criterion, where the cobined output V V out is coputed as follows: V V out = f(v V 1, V V,..., V V n ) = N i=1 σ i N i=1 σ i V V i The variance associated with the cobined values is given by: ( N ) 1 σout = (4) i=1 σ i (4) is the solution to () with f( ) given by (3). As such the OWMF algorith satisfies the BS-7986 requireents for data propagation through function blocks as detailed in the previous section. Although in ost practical cases, the original easureents are drawn fro noral distributions this assuption ay not always stand. In such cases an application specific ipleentation of the IF odule can be provided. In addition, alternative ipleentations ay perfor advanced cross-validation of input easureents and perfor fault detection at the sensor array level. Other alternatives for IF include an approach to consistency checking and data fusion between SEVA sensors proposed in [16]. E. Drift Estiation and Copensation Module (DEC) The DEC odule, ais to detect drift echaniss based on the data, and subsequently correct for the introduced easureent bias. This exercise ais to elongate the useful life span of the intelligent sensor and provide advanced aintenance inforation to higher level processes. The basic functionality of the DEC odule is therefore suarized as: Estiate the drift (or bias) fro historical data. Correct the input for drift and update the uncertainty value accordingly. Most of the algoriths for drift estiation and copensation tend to be application specific so the standard ipleentation acts as a placeholder and siply propagates the received data. Specific algoriths and ipleentations to address particular types of drift are discussed in section IV. The DEC odule, siilarly to the FD odule ay ake use of additional inforation about the sensor in the for of a theoretical sensor odel. F. Intelligent Sensor Control Module (ISC) The ISC odule is responsible for anaging the rest of the software odules in the architecture while it acts as the gateway to the world. As such, it is responsible for aintaining copatibility with industry standards. The basic functionality of the ISC odule is defined as: Counicate with higher level processes (IEEE 1451 interface). Manage the software odules (i.e. essaging, tiing). The standard ipleentation of the ISC odule ipleents a STIM (Sart Transducer Interface Module) as per IEEE [9] with a single channel for the priary easurand (3) data of type buffered data sequence sensor (IEEE 1451.), so that internal tiing is left with the Intelligent Sensor. Functional addresses open for industry read extensions are used to counicate the rest of the BS standard data to higher processes on deand [8], [9], [1], [11]. The ISC odule also takes care of odule essaging as a eans for odules to exchange inforation. This can be achieved either by one-to-one counication between odules, or by one-to-any counication, where odules announce inforation to all interested listeners. Messaging is iportant for the ipleentation of advanced functionality (e.g. the selection of a sensor odel based on environental data). In the rest of the paper we discuss algorithic approaches for the ipleentation of the fault detection, drift estiation and stochastic odeling of adverse sensory systes. For copleteness, an exaple realization of the intelligent sensor software architecture in the particular case of a piezoresistive pressure sensor can be found in the Appendix. III. FAULT DETECTION The objective of the FD odule is to assess incoing data, and produce an uncertainty value to characterize the output. A further desirable characteristic of fault detection, as listed in subsection II-C, is the ability to produce a corrected output if possible. Hereon the forer approach, naely the siple assessent of individual easureents, will be referred to as indicative fault detection. The latter approach, involving the correction of input easureents will be referred to as corrective fault detection. Correction here refers to the estiation of the ost probable value at the specific tie instance given a known history and a new easureent. There are advantages and disadvantages associated with each approach which ste fro their distinct response in particular scenarios. Indicative fault detection is able to identify the existence of outliers in the input data, but there is no echanis to ensure the continuation of the intelligent sensor s operation after such an event. Indicative fault detection is discussed in ore detail in subsection III-A. The actual perforance of corrective fault detection on the other hand is dependant on how uch confidence the algorith places on new easureents versus a-priori knowledge. Assuing higher confidence on a-priori knowledge, results in an algorith robust to the presence of invalid data, but slow to respond to changes in the easureent statistics. Corrective fault detection is discussed in ore detail in subsection III-B. Ultiately, a fault detection algorith should be reactive and robust to the presence of invalid data as well as able to provide a corrected output. A cobination of the above approaches is desirable, hence a ethod is presented in section III-C, based on the use of an AKF to provide corrected output values, whilst in parallel it assesses new easureents online in order to identify invalid data. A. Indicative Fault Detection In indicative fault detection each incoing easureent is assessed individually. Typically the assessent is achieved

6 6 IEEE SENSORS JOURNAL r iteration Fig. 3. Noralized KF innovations versus iteration index. A threshold equal to 4 correctly identifies the introduced outlier by coparing the incoing data to an estiate calculated based on a priori knowledge of the process and previous easureent history. Assuing a linear stochastic syste, residual-based indicative fault detection can be ipleented based on a KF. In this case the innovation E produced by the KF is used to assess new data. The agnitude of the inconsistency can be assessed in relation to the known a-priori observation R and state Q noise covariance atrices (for ore inforation on syste identification see [17]). The weighted innovation is given by: r = E σ, where σ = R + CQC T (5) where C is the observation apping atrix of the KF. In Fig. 3, the ratio r defined by (5) is plotted against the iteration index for siulated data including an outlier at the 1th iteration. Since r is given relative to σ, a threshold T = 4 would ensure that in noral operation % of the innovations will fall below the threshold. B. Corrective Fault Detection In corrective fault detection, the objective is to calculate the current best estiate given new observations as they arrive. In this section we will discuss the use of an Adaptive Kalan Filter (AKF) [18], [19] for corrective fault detection to illustrate the odus operandi of such algoriths. For fault detection, we are only interested in estiating the observation noise online, since we can typically assue that the process characteristics are invariant. The syste odel considered here for illustration is a linear, discrete, stochastic sequence given by: x k = Ax k 1 + w k (6) y k = Cx k + n k (7) where x is the state vector, A is the state transition atrix, y is the observation vector, C is the observation apping atrix, w is the state noise and n is the observation noise. The noise ters w and n are assued white Gaussian noise sequences with covariance atrices Q and R respectively. The state noise is assued zero-ean, while if systeatic errors occur the observation noise ean will be non-zero. The ethod to estiate the ean and covariance of the observation noise is based on a liited eory algorith proposed by Myers and Tapley []. An approxiation to the observation noise saple r k is: r k = y k C ˆx k (8) where ˆx k is the a-priory estiate of x k at tie k. An unbiased estiator for r is the saple ean estiated over N saples. It can then be shown [] that the unbiased estiate of R is given by: ˆR = 1 N 1 N i=1 [ (r i ˆr)(r i ˆr) T N 1 N C ˆP k CT ] (9) where ˆP k is the a priori state covariance estiate at tie k. The selection of the buffer size N iplies a trade-off between building a non-responsive filter and one prone to the presence of outliers in the observation sequence. The optial size of the buffer is proble specific (e.g. Mehra [19]). For illustration purposes, the response of a filter with A = 1, C = 1 and Q =.4 to siulated data (a step change of the real R) is shown in Fig. 4 for buffer sizes N = 5, N = 1 and N = points. An alternative approach to corrective fault detection is to base the estiate on the pdf of the process at the current iteration. This entails the ability to update the pdf on-line as new data becoe available. Existing ethods for calculating a pdf based on a data distribution are designed to work offline and their extension to the on-line doain is difficult. An exaple of such a ethod is discussed in section V. C. Mixed Fault Detection Let us consider again the use of the AKF as the basis of a fault detection algorith. A reasonable choice for the AKF buffer size, in the case of the exaple shown above, would be around N = 1. To exaine the response of such a fault detection algorith to the presence of outliers, an outlier was introduced in siulated data and the filter s response is shown in Fig. 5. This deonstrates that invalid data ay alter the statistics of the AKF buffer draatically. Moreover, since the data are weighted equally in the saple ean and (9), the effect of invalid data lasts until they exit the buffer. Instead of increasing the buffer size, resulting in a low-responsive fault detection odule, it is better to eploy an indicative fault detection approach to identify invalid data and prohibit their use in the estiators. A ixed fault detection approach is proposed here as an extension to the basic AKF algorith. We propose the use of the algorith described in subsection III-A for the identification of invalid data. Data points that are identified as invalid are subsequently suppressed in the AKF algorith, which calculates the estiates of the ean ˆr and Even a single point, depending on the size of the buffer used.

7 KARATZAS ET AL.: TEACHING OLD SENSORS NEW TRICKS: ARCHETYPES OF INTELLIGENCE 7.5 real R estiated R estiated P.5 real R estiated R estiated P variance variance variance variance iteration real R estiated R estiated P (a) iteration real R estiated R estiated P (b) iteration Fig. 4. AKF algorith with buffer sizes, fro top to botto, N = 5, N = 1, N = covariance ˆR of the observation noise based on the reaining data. Alternatively, if an indication of the event is desirable, the invalid data point ay be considered oentarily (resulting in a higher estiate of R), but reoved fro the buffer later on, avoiding an alteration of the AKF buffer statistics for the (c) iteration Fig. 5. AKF algorith response to the presence of an outlier. AKF buffer size N = length of the buffer. The algorith outline is suarized as: 1) At step k, calculate the a priori estiates for ˆx k and ˆP k based on the standard KF equations. If the easureent at tie k 1 was identified as invalid and if the estiates used in the KF calculations were derived including this inforation (see step 4), set ˆP k = ˆP k 1. ) Assess the validity of the new easureent y k based on its innovation using (5). The covariance R used for this step is the previous estiate ˆR k 1. 3) Estiate the new values for the ean ˆr k and the variance ˆR k based on the saple ean and (9) taking into account the new data point. 4) If the data point is identified as invalid: If it is desirable to indicate this event to higher level processes, use the estiated observation noise ean ˆr k and variance ˆR k for the rest for the calculations (resulting in a higher reported covariance). If it is desirable to suppress this event use the ean and variance values of iteration k 1, ˆr k 1, ˆRk 1. Set the observation noise saple r k to the ean of the valid r i in the buffer to suppress the use of this data point: r k = P k i=k N P airi k i=k N ai where a i = { 1 if point i is valid if point i is invalid (1) 5) Coplete this iteration of the KF and go back to step 1. The response of the fault detection odule for the sae data used to produce Fig. 5 is shown in Fig. 6. Fig. 6 (a) shows the ipleentation where the estiates ˆr k and ˆR k are used oentarily to produce an indication of the existence of invalid data and suppressed afterwards. Fig. 6 (b) shows the ipleentation where invalid data are copletely suppressed. The identification of invalid data enables the fault detection odule to eliinate undesirable effects in the AKF buffer statistics without necessitating the use of a larger buffer, retaining therefore the responsiveness of the fault detection

8 8 IEEE SENSORS JOURNAL.5 real R estiated R estiated P 1 5 Acceleroeter PSD as calculated fro easureents 1 deg Celcius, 11 Hz basic vibration 1/f 1.7 region variance PSD in db /f region variance Fig iteration real R estiated R estiated P (a) iteration (b) Mixed fault detection with suppression of invalid data algorith. The ixed fault detection algorith described in this section is the default ipleentation for the FD odule of the intelligent sensor architecture and has been eployed in the exaple realization of the intelligent sensor software architecture discussed in the Appendix. IV. DRIFT ESTIMATION ALGORITHMIC APPROACHES Various uncertain influences (e.g. ageing, environental conditions, poisoning etc.) ay generate drift in a sensor s response, introducing the need for self-calibration and selfvalidation. As previously argued, to avert these effects an integral part of the intelligent sensor is a drift copensation odule [16]. A unique generic and universal approach for drift estiation (either additive or ultiplicative) cannot, however, be proposed due to the essentially different underlying echaniss responsible for its generation. In the literature, specialized approaches have been proposed, e.g. van Putten et. al. [1] copare the chopping ethod, the sensitivity variation approach and the geoetrical van Putten ethod for drift estiation. Coonly, research have been focused on quantitative analyses (physical odels) of either the drift generation echaniss [], [3] or of the resulting drift itself [4]. We propose a unifying approach for Frequency in log(hz) Fig. 7. ADXL3 dual-axis acceleroeter power spectral density for a driving frequency of 11 Hz at a teperature of 1 o C drift classification inferred fro the dynaic and stochastic properties of the output. A central point we introduce in the drift estiation discussion regards the atter of whether drift stes fro processes stochastically related to the sensory operation. Furtherore, distinguishing between reversible and irreversible biases, we can classify drift into four ajor classes: A. Reversible state dependent drift. B. Reversible state independent drift. C. Irreversible state dependent drift. D. Irreversible state independent drift. In the following subsections each of these classes is discussed in detail, along with proposed approaches for their identification and copensation in specific cases; e.g. additive drift due to sensor nonlinearities and ultiplicative drift due to phase noise of internal oscillators in the sensory syste. We have used an Analog Devices ADXL3 dual-axis acceleroeter as a case study to test the possible ipleentations of drift copensation odules for these four classes of drift. The analysis coences with a siple high level odel of the acceleroeter. We have perfored spectral analyses of easureents and plot in Fig. 7 the Power Spectral Density (PSD) of the x- axis output when the acceleroeter was driven by an acoustic vibrator having a sinusoidal input of 11 Hz. Fro the sensor PSD we obtain a qualitative representation of the acceleroeter/vibrator syste as a weakly nonlinear syste of order four with the second, third and fourth haronics being clearly distinguishable in the graph. Furtherore, by siple inspection of the PSD at low frequencies we can identify a 1/f type power-law process. This kind of low-frequency processes are considered to arise due to intrinsic electronic noise sources in the acceleroeter circuit. Finally, the spectral broadening around the vibrating frequency of 11 Hz and its haronics due to phase noise can be odeled as a ultiplicative gain drift as stes fro theoretical analyses in [5]. A ore detailed analysis of this effect is presented in subsection IV-D.

9 KARATZAS ET AL.: TEACHING OLD SENSORS NEW TRICKS: ARCHETYPES OF INTELLIGENCE 9 TABLE I ACCELEROMETER WEAK NONLINEARITY COEFFICIENTS Polynoial coefficient (Pc) Pc ean value Pc variance a a a Based on the previous rearks, we can odel the acceleroeter following the input/output relation: 4 y(t) = d(t) + n(t) + b(t) a k x(t) k (11) k= where d(t) represents additive 1/f type noise sources, n(t) is a zero-ean white Gaussian process, b(t) is the ultiplicative gain drift due to phase noise, a k is the kth order coefficient in the Taylor (or Volterra) series expansion of the syste tie doain response. x(t) is the easurand process, coonly odeled as a Gaussian rando process. A. Reversible, state dependent drift Reversible, state dependent drift results fro non-idealities in the sensor response (the ost coon exaple being presented by nonlinearities). In the case where a weak-stationarity requireent is fulfilled, it is possible to develop approxiate state space odels of such sensors using the EKF [6]. In that aspect, a tangible estiation of the ean drift/offset in the sensor output as a function of the internal state oents is possible. Of greatest practical interest is when the sensor is a nonlinear device. A variety of sensors have been described by a nonlinear odel [7], and various approaches have been used for the circuvention of such natural phenoena [8]. In order to deonstrate the alternative use of the EKF in such sensory systes, we proceed by considering the specific case where the sensor output can be described by a low-order polynoial, representing the sensor as a weakly nonlinear syste. For the derivation of an approxiate nd order polynoial odel for the ADXL3 dual-axis acceleroeter during the design of the EKF, we have easured the sensor response for a driving frequency of 1 Hz and obtained N = 3 sets of n = 5 easureents. We have used of the sets to obtain a unique averaged nd order odel for the acceleroeter, while using these paraeters we algorithically estiated the drift in the reaining 1 sets. The estiated coefficients of the nd order polynoial are included in Table I. The algorithic estiation of the drift due to the second order nonlinearity is based on the subsequent equations: x k = x k 1 + w k (1) y k = a + a 1 x k + a x k + n k (13) σ x σ ˆx k = (k 1)σˆx k 1 + ˆx k k dc k = (k 1)dc k 1 + a σ ˆx k k Overall drift (14) (15) dc = dc k + a (16) Drift in Volts saple ean estiated drift Saple nuber Fig. 8. Estiation of state dependent drift of an acceleroeter. The solid lines represent estiations while the dot-lines are calculated as the set epirical offsets where variances of the white zero-ean Gaussian processes w and n are Q =.63 and R = 1 1 respectively. Assuing that x is a Gaussian rando process, the above expressions converge to the actual variance and ean respectively. (14) and (15) represent the ean values of the variance and dc offset respectively based on the EKF a-posteriori estiates ˆx k, while (16) accounts for the overall drift taking into account the polynoial coefficient a. Fig. 8 copares the estiated and epirical (calculated as the ean of the saple set) drift values for the 1 test sets of easureents. The ean estiated drift is 49. V, less than 1% different fro the ean epirical drift. Fig. 8 also shows the rapid convergence of the proposed algorith, with approxiately 1 saples being necessary for a convergence better than 5% error. The proposed algorith is eployed in the exaple realization of the intelligent sensor in the case of a piezoresistive pressure sensor included in the Appendix. The algorithic drift estiation discussed in the present section is not conceptually confined to the case of nonlinearities, but encopasses the whole class of state dependent drift. The non-ideal sensor tie-doain response h( ) along with H( ), the Jacobian atrix of its partial derivatives (with respect to x), have to be suitably odified. As an exaple, the additive drift in the acceleroeter x-axis due to a linear crosscorrelation to the y-axis can be evaluated through the following siple odifications; a coupled EKF should be constructed for the two acceleroeter axes based on relations of the type: h x (x, y) = a 1 x + λ xy y (17) H x (x, y) = a 1 (18) h y (x, y) = β 1 y + λ yx x (19) H x (x, y) = β 1 () In (17) λ xy represents the cross-correlation of the x-axis to the y-axis, while λ yx in (19) represents the cross-correlation of the y-axis to the x-axis and a 1, b 1 represent linear gains in the two axes.

10 1 IEEE SENSORS JOURNAL B. Reversible, state independent drift Reversible state independent drift can generally be evaluated. A representative exaple of this class of drift is the gravitational offset (g-offset) in an acceleroeter, generated fro its relative to the ground angle. A coonly-encountered approach for the acceleroeter calibration is through the use of a pair of acceleroeters at known relative positions. Alternatively, in the case where the dc coponent of the sensor output is of no interest, ac coupling or the use of differential circuits in the output is an effective low-cost ethod for its eliination. C. Irreversible, state dependent drift Drift due to irreversible state dependent phenoena is the result of sensor non-stationarity (dynaic behavior) in the tie interval of interest 3. Such drift can severely coproise the perforance of a sensory based syste, such as an electronic nose. It generally follows a specific trend [9] and various approaches for its copensation have been reported in literature [3] (ultiplicative step drift), [31] (additive dynaical drift). On-line density estiation algoriths specifically designed for systes that are governed by rapid dynaics are currently being researched. In [3] a novelty detection based approach is outlined using recursive dynaic principal coponent analysis for gas sensor arrays, while in [33] the use of independent coponent analysis is deonstrated for non-gaussian data. Conversely, a density estiation approach for drift copensation could be based on the on-line ipleentation of the density estiation algorith discussed in section V. D. Irreversible, state independent drift Irreversible state independent drift is due to the long-ter ean of intrinsic noise sources (such as 1/f noise in electronic sensors). In general it is the ost difficult part of drift to identify and copensate. In the following we will present a theoretical result for the ultiplicative drift estiation in the case of oscillatory-based systes such as acceleroeters, direct frequency output vibrating gyroscopes [34] or surfaceacoustic-wave and surface-transverse-wave based gas sensors [35] in the presence of phase noise in the internal oscillatory systes. In the case of the latter, close-to-the-carrier phase noise can be the ajor liiting factor to the sensor noise and its resolution. As discussed in detail in [5], real oscillators suffer fro phase noise that distorts the syste long-ter stability. The effect is anifested through a spectral broadening (indicating energy spreading) around the oscillation frequency and is odeled by eans of power-law phase noise processes with PSD of the type k α f α, where α {, 1,, 3, 4} and f is the frequency in Hz. In order to quantify the resulting drift effect, we begin with considering the coplex valued oscillation: ψ(t) = e j(ωosct+φ(t)) (1) 3 Any real syste exhibits non-stationary behavior if observed for a sufficiently long tie. In that context systes can be odeled as stationary as long as their dynaic behavior is negligible for application concerned purposes. where ψ(t) is an analytic version of a real oscillator at ω osc = πf osc assuing negligible aplitude noise. In [5] the correspondence in ters of ean and variance between a real valued oscillator and its coplex valued counterpart (analytical representation) is provided. In (1) we can isolate the effect of the phase noise coponent as a ultiplicative ter b(t), b(t) = e jφ(t) () with φ(t) representing the phase noise process. Modeling φ(t) as a zero-ean Gaussian rando process, we can estiate the expected value of the process b(t) as a function of the variance σφ (t) of the phase noise process φ(t): E[b(t)] = E[e jφ(t) ] = e σ φ (t) (3) where E[ ] denotes statistical expectation. (3) expresses the actual ultiplicative drift in the oscillator output. As a result, drift in the output of oscillatory systes corrupted by phase noise is expressed as a function of the overall variance of the phase noise process and can be tie dependent in the case of non-stationary phase noise 4. The above approach can be incorporated in the design process of sensors based on oscillators and can be particularly useful in early siulation stages. V. OFR WITH LOO TEST SCORE AND LOCAL REGULARIZATION SPARSE DENSITY ESTIMATION Fundaental to fault detection algoriths for coplex nonlinear sensors is the ability to generate an efficient pdf for the underlying process, so that a etric of its output easureent uncertainty can be evaluated (and propagated through a fusion process) as well as its utilization in a fault detection algorith. Nonparaetric kernel based approaches provide an efficient fraework for the estiation of pdfs of representative saple sets (e.g. of size at least N = 1). A nuber of density estiation ethods are available with the Parzen windows [36] being the ost widely used and universal approach. Recently, however, sparse density algoriths have becoe available, [37], [38], [39], in view of the need for decrease in CPU tie and eory size for the representation of a probabilistic syste odel. In applications related to intelligent sensing, the requireent for sparsity is acute. More iportantly though, a parsionious stochastic representation of a sensory syste is in better agreeent with intuition about the actual physical processes being odeled. In contrast to the Parzen window where the nuber of kernels eployed to represent the syste is equal to the size of the representative saple set (1s), sparse kernel density approaches only use a sall nuber of kernels (< 1). The physical syste is expected to have a ore concentrated behavior. In the present section, we eploy the use of the Orthogonal Forward Regression with Leave-One-Out test score and local regularization (OFR-LOO) algorith [37] for the construction of a sparse probabilistic odel for the sensory syste. The ain otivation behind using the above algorith is the 4 σφ can be a function of tie.

11 KARATZAS ET AL.: TEACHING OLD SENSORS NEW TRICKS: ARCHETYPES OF INTELLIGENCE 11 fact that the Leave-One-Out test score, a powerful crossvalidation technique in nonparaetric odeling, is convex in regard to the odel coplexity. In the case of kernel-based approaches, odel coplexity is siply represented by the nuber of kernels eployed. The use of the OFR-LOO as a construction 5 algorith allows the autoatic terination of the odel selection procedure. In the following we will deonstrate and validate the use of the OFR-LOO algorith through a siulated exaple and then ove to the odeling of an acceleroeter at different operation conditions. In this context, the proposed use of the OFR-LOO is for the provision of sparse kernel odels in operating conditions identified by the odule. In the siulated exaple we have considered the case of a piezoresistive pressure sensor. In [7] is argued that such sensors are sensitive to the operating teperature T, with their resistance R being described by an expression of the type: R(T ) = R()(1 + αt + βt ) (4) with α, β R, β < α. The dependance of the sensor resistance on the teperature as described in (4) introduces the following iplications: (i) there exists a teperature dependent dc offset in the sensor output (ii) the actual gain K of the sensory eleent is teperature dependent. As a result of the above, for the sae set of pressure easurands P, the probabilistic odel that describes the sensor output has an increasing ean and variance with increasing teperature. The above effects are deonstrated in the siulated piezoresistive pressure sensor, where the output voltage is odeled as: V = K(T ) P (5) K(T ) = K (1 + αt + βt ) (6) In our siulation we have chosen the values α =.1 and β =.1 that are of the sae order of agnitude with estiated values in [1]. In Fig. 9 we present the stochastic odeling of the siulated piezoresistive pressure sensor, using the OFR-LOO algorith and copare the results with a Parzen window of N = 1 6 saples that represents the actual pdf (the Parzen window converges to the actual pdf when the saple size is infinite). The pressure easurand is odeled as a zeroean Gaussian process of unit variance and the OFR-LOO pdf is estiated over N = 1 saples. The OFR-LOO window length was estiated through cross-validation as σ = 1.8 while the window length of the Parzen window is calculated as σ = 1. The OFR-LOO pdf atches well with the Parzen window pdf in all exained teperatures, thus validating the use of the OFR-LOO for pdf estiation, even when a sall nuber of saples (N = 1 in the specific exaple) are available. In the sae graph is also shown the slight drift in the ean value of the sensor output pdf, inferring the possible use of 5 We construct the odel starting fro a single kernel and oving to higher order representations following a tree based approach for the search of the optial overall sub-path at each stage of the construction. The process is terinated when a iniu in the LOO test score is achieved Probability Density Function Parzen o C OFR o C Parzen 5 o C OFR 5 o C Parzen 1 o C OFR 1 o C Parzen 15 o C OFR 15 o C Sesnor Output Fig. 9. Coparison of the pdf of a siulated piezoresistive pressure sensor, using the OFR-LOO algorith with window length σ = 1.8 and Parzen windows with window length σ = 1 TABLE II PDF ESTIMATION FOR THE SIMULATED PIEZORESISTIVE PRESSURE SENSOR Teperature Mean nuber of kernels Std. in kernel nuber o C o C o C o C the on-line version of the OFR-LOO for drift estiation. More iportantly, it is clearly shown that different stochastic odels (e.g. distinguishably different variance) should be eployed as a result of variations of the environental conditions. In Table II we present the results of 1 independent runs of the OFR-LOO algorith for the siulated piezoresistive pressure sensor and have calculated the ean nuber of kernels eployed in the proposed stochastic odeling along with the standard deviation in the nuber of kernels. It is worth noting that the stochastic process can be reliably represented by 1 or kernels, in contrast to traditional stochastic odels that ake use of the totality of the saple set (N = 1). The previous discussion deonstrates that density estiation algoriths can be eployed in the realization of novelty detection based fault detection operations. Moreover, on-line parsionious density estiators are suitable for the identification of irreversible state dependent drift as discussed in subsection IV-D. In the rest of the section we use the OFR- LOO algorith on real ADXL3 acceleroeter data. The OFR-LOO algorith has been used for the odeling of the acceleroeter in different operating teperature conditions, when driven by an acoustic vibrator at 1 and Hz and have found that the pdfs coincide for these two driving frequencies, inferring that the probabilistic acceleroeter odel is independent of the easurand haronics. Furtherore, the acceleroeter pdf was estiated for a range of operating

12 1 IEEE SENSORS JOURNAL pdf OFR acceleroeter 1D OFR o C OFR 5 o C OFR 1 o C OFR 15 o C x accelereoter axis Fig. 1. Estiation of pdf fro real acceleroeter data using the OFR-LOO algorith TABLE III PDF ESTIMATION FROM REAL ACCELEROMETER DATA Teperature Mean nuber of kernels Std. in kernel nuber o C o C o C o C teperatures, o C, 5 o C, 1 o C and 15 o C over N = 1 saples with Gaussian kernels of length estiated through cross-validation as σ =.85. The results are presented in Fig. 1. Using 1 sets of N = 1 easureents, we have repeated the above pdf calculation and have estiated the ean nuber of kernels and the standard deviation in the nuber of kernels required fro the OFR-LOO algorith to odel the acceleroeter in the previously entioned operation conditions. The results are presented in Table III. Two ain conclusions can be drawn fro the above. There appears to be a noticeable drift in the acceleroeter output as shown by the ean of the pdf in different teperatures. Although we believe that this is an artefact of the experiental set-up due to acceleroeter rotating during the easureents, it is clear that on-line density estiation can provide inforation about drift. Secondly, the shape of the pdf at 15 o C is distinguishably different fro the pdf of the rest of the operating teperatures. The noinal teperature operating range is defined as o C-11 o C and as a result the stochastic odeling eploying the OFR-LOO algorith can separate the noral sensor output fro the corrupted due to operating outside the range of noinal conditions. The benefit of the approach is that it utilizes sensor data alone, no a-priory physical odels are necessary and conventional statistical hypothesis fault condition algoriths can be used directly with the derived pdfs. Such abstract sensor representations capture the underlying physical processes avoiding the developent of detailed quantitative odels. Also, the approach is particularly aenable to sensors with nonlinearities dependent on the operating regie. VI. CONCLUSIONS In this paper we have proposed a generic, odular software architecture as an advantageous intelligent sensor ipleentation. Such an approach enables on-board signal processing to produce the optial signal output. The proposed architecture bridges the gap between existing industry standards (IEEE 1451 and BS-7986) by exposing higher level signal processing (BS-7986) while adhering to low-level prerequisites set by IEEE The software architecture ais at addressing the totality of the intelligent sensor goals, naely optial data fusion, real-tie fault detection, calibration and autonoous reconfiguration. We have exained a variety of possible algorithic ipleentations of the above operations and have proposed a ixed indicative/corrective approach for fault detection. In ters of drift estiation, we have shown that the EKF can be eployed in reversible state dependent drift estiation, while on-line density estiation can be utilized when the drift is irreversible state dependent. Furtherore, a state of the art sparse density estiation algorith was used in a parsionious odel selection context. It was deonstrated that it can be used with a statistical hypothesis algorith to generate fault condition, even when no physical sensor odel is available over a range of operating regies. A snapshot of the developed deonstrator of the intelligent sensor software architecture is included in the Appendix, using synthetic data. APPENDIX The exaple ipleentation illustrated in Fig. 11 features a piezoresistive pressure sensor as the priary sensing eleent while teperature sensors are used to onitor the environent and choose the right sensor odel for drift copensation during the operation. The current ipleentation of the intelligent sensor software architecture is developed in software using an object oriented language to facilitate the specialization of odules as necessary. However, as long as the architecture s definition is adhered, all or parts of the odules of the architecture can be ipleented in hardware where this offers an advantage. A drift estiation and copensation odule is used for calibration of the pressure sensor output V. The dependance of V on the easurand P and the teperature T is described below: V = K(T )(P +.1P ) (7) K(T ) = T T (8) As a result, the pressure sensor suffers both fro additive drift because of the second order nonlinearity in (7) and fro ultiplicative gain drift as expressed in (8). The additive drift is evaluated using the EKF algorith as outlined in subsection IV-A. For the identification of the ultiplicative gain drift we use the teperature easureents for a straightforward evaluation based on (8).

13 KARATZAS ET AL.: TEACHING OLD SENSORS NEW TRICKS: ARCHETYPES OF INTELLIGENCE 13 Fig. 11. Deonstrator of the intelligent sensor software architecture featuring a piezoresistive pressure sensor as the priary sensor and teperature sensors as environental conditions onitoring sensors As far as the teperature sensors are concerned (two in the specific ipleentation), a ixed indicative/corrective AKF based approach is used, as discussed in subsection III-C, for fault detection before the internal fusion odule. In the snapshot, two outliers are identified in the output of the first teperature sensor and are reoved before the teperature easureents are fused. REFERENCES [1] M. A. Clapp and R. Etienne-Cuings, A dual pixel-type array for iaging and otion centroid localization, IEEE Sensors Journal, vol., no. 6, pp , Dec.. [] L. Kish, J. Solis, W. Marlow, R. V. ad C. Granquist, V. Lantto, J. Sulko, and G. Scera, Detecting harful gases using fluctuationenhanced sensing with Taguchi sensors, IEEE Sensors Journal, vol. 5, no. 4, pp , Aug. 5. [3] N. A. Riza, M. A. Arain, and F. Perez, Harsh environents inially invasive optical sensor using free-space targeted single-crystal silicon carbide, IEEE Sensors Journal, vol. 6, no. 3, pp , June 6. [4] J. Yang and D. Clarke, A self-validating thero-couple, IEEE Transactions on Control Systes Technology, vol. 5, no., pp , Mar [5] M. Henry, Recent developents in self-validating (SEVA) sensors, Sensor Review, vol. 1, no. 1, pp. 16, 1. [6] J. P. Cassar, M. Bayart, and M. Staroswiecki, Hierarchical data validation in control systes using sart actuators and sensors, in IFAC Syp. Intelligent Coponents and Instruents for Control Applications (SICICA 9), 199. [7] M. Staroswiecki, Intelligent sensors: A functional view, IEEE Trans. on Industrial Inforatics, vol. 1, no. 4, pp , Nov. 5. [8] IEEE, IEEE Std Standard for Sart Transducer Interface for Sensors and Actuators - Network Capable Application Processor, J. C. Edison, Ed. IEEE Standard Association,. [9], IEEE Std Standard for Sart Transducer Interface - Transducer to Microprocessor Counication Protocols and Transducer Electronic Data Sheets (TEDS) Forats, E. V. El-Kareh, Ed. IEEE Standard Association, [1], IEEE Std Standard for Sart Transducer Interface - Digital Counication and Transducer Electronic Data Sheet (TEDS) Forats for Distributed Multidrop Systes, L. Eccles, Ed. IEEE Standard Association, 4. [11], IEEE Std Standard for Sart Transducer Interface for Sensor and Actuators - Mixed Mode Counication Protocols and Transducer Electronic Data Sheet (TEDS) Forats, T. Licht, Ed. IEEE Standard Association, 4. [1] A. de Castro, A syste-on-chip for sart sensors, in Proc. of the IEEE International Syposiu on Industrial Electronics (ISIE ),. [13] J. Schalzel, F. Figueroa, J. Morris, S. Mandaya, and R. Polikar, An architecture for intelligent systes based on sart sensors, IEEE Transactions on Instruentation and Measureent, vol. 54, no. 4, pp , Aug. 5. [14] B. S. B. 7986:5, Data Quality Metrics for Industrial Measureent and Control Systes - Specification. BSi, Apr. 5. [15] C. Harris, X. Hong, and Q. Gan, Adaptive Modelling, Estiation and Fusion fro Data. Springer,. [16] M. Duta and M. Henry, The fusion of redundant SEVA easureents, IEEE Transactions on Control Systes Technology, vol. 13, no., pp , Mar. 5.