A multipoint optical fibre sensor system for use in process water systems based on artificial neural network pattern recognition techniques

A multipoint optical fibre sensor system for use in process water systems based on artificial neural network pattern recognition techniques

Sensors and Actuators A 115 (2004) 293–302 A multipoint optical fibre sensor system for use in process water systems based on artificial neural netwo...

398KB Sizes 0 Downloads 6 Views

Sensors and Actuators A 115 (2004) 293–302

A multipoint optical fibre sensor system for use in process water systems based on artificial neural network pattern recognition techniques D. King∗ , W.B. Lyons, C. Flanagan, E. Lewis Department of Electronic and Computer Engineering, University of Limerick, Castletroy, Limerick, Ireland Received 22 September 2003; received in revised form 12 February 2004; accepted 15 March 2004 Available online 20 May 2004

Abstract A dual-element multipoint optical fibre sensor system capable of detecting ethanol in water supplies is reported. The sensor system utilises a U-bend configuration for each sensor element in order to maximise the sensitivity of the system and is interrogated using a technique known as optical time domain reflectometry, as this method is capable of detecting attenuation over distance. Analysis of the data arising from the sensor system is performed using artificial neural network (ANN) pattern recognition techniques, coupled with Fourier transform-based signal processing. The signal processing techniques are applied to the obtained sensor system data, prior to the ANN analysis, with the aim of reducing the computational resources required by the implemented ANN in software. © 2004 Elsevier B.V. All rights reserved. Keywords: Optical fibre sensors; U-bend configuration; Optical time domain reflectometry; Measurement; Artificial neural networks

1. Introduction and background Over the past two decades, the use of optical fibre sensors for the purpose of environmental monitoring has expanded rapidly [1–4]. Optical fibre sensors possess a number of advantages over conventional electronic sensing techniques, which make them attractive for use in a wide range of application areas. These advantages include safety in chemically hostile environments, immunity to electromagnetic interference, electrically passive operation, sensitivity, weight and versatility. In this investigation two sensor elements are incorporated into a 1 km length of 62.5 ␮m core diameter polymer clad silica (PCS) optical fibre cable. In order to maximise the sensitivity of the optical fibre sensor system, a U-bend configuration is used for each of the sensor elements, where the buffer and cladding have been chemically removed and the fibre core exposed to the measurand under test. Compared to their networked and single point counterparts, multimeasurement, multipoint or distributed sensors have proven to be an attractive and less expensive option in the field of fibre optic sensing. In order to locate the position and changes in the sensor output on the fibre, multipoint sensors utilise optical time domain reflectometry (OTDR) ∗ Corresponding author. Tel.: +353-61-213558; fax: +353-61-202572. E-mail address: [email protected] (D. King).

0924-4247/$ – see front matter © 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.sna.2004.03.068

techniques. OTDR, since first being reported in 1976 [5], has become an established technique for attenuation monitoring in optical fibre networks within the telecommunications industry. OTDR has recently found many applications in the field of multipoint or distributed sensing on a single fibre loop [6] with various sensing and addressing methods possible when implementing a distributed sensor system [7]. Optical fibre sensor signals can often be complex and cross-coupling of signals from external parameters, e.g. temperature (the true measurand) and strain or microbending (interfering parameters in this case), adds to the difficulty interpreting data from such systems. It has been proposed that for many applications of optical fibre sensors, artificial neural network (ANN) pattern recognition techniques may be used to resolve the problems arising from cross-sensitivity to other parameters [8]. It is proposed in this investigation to apply ANN pattern recognition techniques to the obtained sensor system data with the aim of accurately classifying each of the sensor test conditions. Prior to the ANN analysis, novel Fourier transform-based signal processing techniques are applied to the obtained sensor data with the aim of reducing the required number of nodes in the input and hidden layers of the implemented ANN without affecting the accuracy of the ANN classifications. Previous work by King et al. has shown that it is possible to train a feed forward ANN based on the frequency domain response of a single sensor

294

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

element optical fibre sensor system [9,10]. In this investigation, the sensor system has been expanded to contain two U-bend sensor elements and the sensing elements are exposed to various combinations of test conditions: various combinations of air, 50% ethanol and distilled water at each of the sensing elements. For each set of test conditions, an ANN is implemented in software with the aim of accurately classifying each fluid under test at each of the sensing elements.

2. Experimental setup 2.1. U-bend sensor fabrication In this investigation the sensors were incorporated into a 1 km length of 62.5 ␮m core diameter polymer-clad silica (PCS) optical fibre. In order to maximise sensitivity, a U-bend configuration was used for each of the sensor elements where the cladding was removed and the core exposed directly to the measurand. In order to minimise optical power loss, the sensors were fabricated from the same fibre as that used to transmit light through the system (i.e. 62.5 ␮m PCS), removing the need to splice the sensors into the transmitting fibre and the resulting power losses. An evanescent wave absorption sensor was chosen as it offers high sensitivity and a linear dynamic range. High sensitivity is achieved by using a U-bend probe whose sensitivity increases with decreasing bend radius. A linear dynamic range is maintained by preventing surface reactivity of the silica core of the fibre with the solute molecules of the absorbing fluid by operating the sensors at a low pH level [11]. Applications of evanescent wave absorption sensors include the detection of contaminants (e.g. particles, inorganic or organic species) in water or other fluids and the detection of depositions coated on the sensor surface. The operation of the sensor is based on the modulation of the light intensity propagating in the fibre by the measurand as a result of the interaction with the evanescent field penetrating into the absorbing measurand. The evanescent field absorption, for a given length of the unclad fibre, depends on the number of ray reflections per unit length of the unclad fibre and the penetration depth of the evanescent field in the sensing region [12]. The penetration depth of the evanescent field for a liquid of refractive index n2 surrounding the sensing region is given by λ dp := (2.1) 2πn1 (cos2 θc − cos2 θ sin2 θφ )1/2 where λ is the wavelength of light launched into the fibre in free space, n1 the refractive index of the core, θ c (=sin−1 (n2 /n1 )) the critical angle of the sensing region with respect to the normal on the core cladding interface, θ the angle of the ray with the normal to the core cladding interface and θ φ is the angle of skew. As θ approaches θ c , the

penetration depth increases, which results in the increase in evanescent absorption and hence the sensitivity of the sensor [13]. Much experimental work has already been reported [14,15] for a single U-bend sensor detailing resulting sensitivity gains from evanescent wave increases from the curving of the sensing fibre. It has been shown by Gupta and coworkers that the sensitivity of the sensor increases with decreasing bend radius of the probe and also with the increase in refractive index of the fluid under test [12,14]. In order to produce the U-bend sensor elements in the fibre, the buffer and cladding were chemically removed from a 2 cm length section located at 665 and 756 m from the launch end of the fibre. In order to shape the fibre to the desired sensor configuration, the exposed fibre was cleaned using acetone and was then slowly bent into a U-shape using heat from a flame. The bending procedure was controlled using an in house developed fixture to improve the repeatability of the sensor manufacturing and hence improve the reproducibility characteristics between successive sensors, see Fig. 1. The final bend radii were measured to be 1 mm for each sensor using a conventional optical laboratory microscope. A schematic and a photo of a fabricated U-bend sensor are shown in Fig. 2. 2.2. Sensor interrogation—optical time domain reflectometry The sensor used in this investigation makes use of radarlike elastic Rayleigh backscattering to make continuous measurements on optical fibre cable. The basic method of optical time domain reflectometry (OTDR), devised by Barnoski and Jensen [5], represented the first distributed optical fibre sensor. OTDR was first reported in 1976 as a telecommunications application and became an established technique for attenuation monitoring and fault location in optical fibre networks. OTDR is capable of detecting attenuation as a function of distance along the fibre and therefore is able to locate position and changes in the sensor signals along the fibre. As a result of this, OTDR has found many applications in both single and multipoint sensors where the OTDR instrument is used to monitor the fluctuation in the optical fibre attenuation caused by an external parameter induced by a measurand [16,17]. Using an OTDR, the distance of any particular change in the backscattered light can be calculated by measuring the elapsed time of the return pulse. If the time required to propagate back and forth is τ, then change at the location L, along the fibre is given by L=

cτ 2n

(2.2)

where c is the velocity of light in a vacuum (3 × 108 m s−1 ) and n is the core refractive index of the optical fibre cable.

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

295

Fig. 1. U-bend sensor fabrication stages: (a) initial bend; (b) final bend.

Fig. 2. Fabricated U-bend optical fibre sensor.

In order to use the OTDR approach as a basis for developing a distributed or multipoint optical fibre sensor, modifications are made to the fibre at the sensing regions. As a result of these modifications, the local loss and/or backscattered characteristics at the specific locations can be altered by exposure to the required measurand. Any subsequent change in the measurand will also have an effect on the fibre characteristics at the modified sensing location. The OTDR method is shown in Fig. 3. A pulsed semiconductor laser is coupled into a section of fibre via a directional coupler, which serves also to couple the backscattered light fraction, captured and returned via the fibre under test, to the avalanche photodiode (APD) detector. In the example shown in Fig. 3, the OTDR pulses a signal and the returned back scattered signals are recovered from the same fibre end via an optical coupler. The OTDR is capable of detecting refractive index fluctuations (in this case Rayleigh scattering from fibre 1 and fibre 2) and any discontinuities along the fibres (in this case the launch end connection, the fibre end Fresnel reflection and the mid-fibre splice). When examining optical fibre systems using OTDR care must be taken with short wavelength sources (particularly <600 nm) to ensure that there is no significant fluorescence in the fibre [18]. In order to be guided by the fibre

back to the detector, the scattered light must couple into reverse propagating guide modes of the fibre. For a graded index multimode fibre, the relevant backscatter capture ratio, S, is defined as the ratio between the instantaneous coupled-reverse power and the total scattered power at the point of scattering, and is given by   NA 2 S = 0.25 (2.3) n1 where NA is the numerical aperture of the fibre and n1 is the core refractive index [19]. 2.3. Measurement system configuration The system configuration used in this investigation comprises of the U-bend optical fibre sensors, an OTDR test instrument and a Pentium MMX 200 MHz PC, see Fig. 4. The OTDR used in this investigation is an EXFO IQ7000 [20,21] with an 850 nm laser source, which offers high-resolution multimode operation with an event dead zone1 of only 1 Fresnel dead zone or minimum distance at which a 4% Fresnel reflection generated by a near end event can be detected.

296

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

Fig. 3. Basic OTDR concept [5]: (a) optical setup; (b) OTDR output signal.

Fig. 4. Measurement system configuration.

2 m. The EXFO IQ-7000 OTDR is capable acquiring up to 16,000 data points on a single trace and supports both local and remote control. In this investigation the OTDR was connected to the host PC via an IQ-7000 general purpose interface bus (GPIB) [22] port and therefore is capable of remote control, via a host PC. The host PC runs LabVIEW Virtual Instruments (VI) programs for data capture and preprocessing.

3. Results 3.1. Sensor test conditions investigated In order to train and test an ANN pattern recognition system, it is necessary to obtain a large number of patterns.

For this reason, numerous OTDR readings were taken for each of the sensing conditions under test, see Table 1. A total of 51 readings, each of 3 min duration, were taken for each of the sensors test conditions. In order to allow the sensor Table 1 Sensor test conditions investigated Sensor 1

Sensor 2

Expected output

No. of patterns

Air Air Air Ethanol Ethanol Ethanol Water Water Water

Air Ethanol Water Air Ethanol Water Air Ethanol Water

0 0 0 0 0 0 1 1 1

51 51 51 51 51 51 51 51 51

0 0 0 1 1 1 0 0 0

1 1 1 0 0 0 0 0 0

0 0 1 0 0 1 0 0 1

0 1 0 0 1 0 0 1 0

1 0 0 1 0 0 1 0 0

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

297

Fig. 5. OTDR output trace.

system to fully stabilise, the first OTDR reading taken for each test condition is ignored and the remaining 50 patterns are used in the training and testing of the ANN implemented in the SNNS [23] software package.

are shown in Fig. 6. Prior to the application of the Fourier transform-based signal processing, the extracted peaks are normalised between −1 and +1 using a standard scale 2D array LabVIEW VI.

3.2. OTDR sensor peak extraction and pre-processing 4. Signal processing analysis The sensing area of interest on the OTDR trace forms a relatively small part of the overall trace (approximately 256 points out of a total of 12,000, see Fig. 5) and therefore to maximise the efficiency of the computer algorithm, it was necessary to design an in-house LabVIEW VI that would locate the sensor peaks, select the required window width and save this data for analysis by the signal processing and the ANN software. Sample traces of the extracted OTDR sensor peaks for the test conditions listed in Table 1

Previous work by Lyons et al. [24] has shown it is possible to train a multilayer perceptron using the data obtained from a U-bend sensor interrogated by an OTDR. In this investigation it has been proposed to apply a digital signal processing technique to the obtained sensor data, prior to the training of the ANN, with the aim of reducing the computational resources of the implemented ANN, i.e. fewer nodes required in the input and hidden layers.

Fig. 6. Extracted OTDR sensor peaks.

298

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

It can be observed from the extracted OTDR peaks shown in Fig. 6, that it is relatively low frequency information that is of interest in this classification. Due to the low frequency nature of the information, a discrete Fourier transform (DFT) using an FFT algorithm can be directly applied to the OTDR peaks without having to apply any windowing transform. The low frequency nature of the information also reduces concerns over filtering and aliasing in the frequency domain [25,26]. The signal processing analysis in this investigation is performed using an in house developed MATLAB program where a 256-point DFT using an FFT algorithm is applied to the OTDR output data obtained. MATLAB was selected in preference to LabVIEW to perform the signal processing analysis as it offers a superior DSP toolbox to LabVIEW and is recognised as a DSP package, whereas LabVIEW is better suited to instrumentation applications. Once the extracted OTDR sensor peaks are inputted into MATLAB, a 256-point DFT of the peak is calculated and from the resulting Fourier transform the power spectral density (PSD) of the OTDR output is calculated. The resulting PSD plots are shown in Fig. 7. As anticipated, the main PSD area of interest is located in the low frequency region. As a result of the application of the discrete Fourier transform, the OTDR peak information is now more explicit and easier for the user to access in comparison to time domain-based results which require all of the extracted OTDR peak data points. An empirical decision was made to select the first 12 points of the PSD plot as the main area of interest on the trace and these 12 points form the input layer of the ANN implemented in SNNS. These 12 points adequately represent the area of interest of the PSD trace whilst minimising the required number of nodes in the input layer of the implemented ANN.

Fig. 7. Corresponding OTDR sensor peak PSD traces.

5. Artificial neural network pattern recognition 5.1. Artificial neural network architecture design The origins of artificial neural networks can be traced back to a publication by Mc Culloch and Pitts [27] of the first mathematical model of a biological neuron in 1945. During the mid-1980s there was a resurgence of interest in neural networks after a period of approximately 20 years when the research in the area effectively stopped. The resurgence was largely prompted by the publication of Rumelhart and Mc Clelland’s [28] book Parallel Distributed Processors. During the 1990s, ANNs have been generally accepted as a major tool in the developments of intelligent systems. There are many different types of neural networks in existence, ranging from relatively simple to highly complex. In this investigation a three-layer feed forward neural network was used, see Fig. 8. The multilayer feed forward network

Fig. 8. Three-layer feed forward artificial neural network.

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

defines one of the most widely used and successful neural network architectures. Numerous applications of using the multilayer feed forward network in conjunction with optical fibre sensors have been reported by Lyons and Lewis [8]. One useful application of the feed forward network is to extend the response of optical fibre sensors. Optical fibre sensors are often limited by their linear range of operation, good examples of this being some optical fibre pH sensors and humidity sensors [29]. However, Taib et al. have reported a technique for extending the response range of an optical fibre pH sensor using an ANN approach [30]. The implementation of the ANN with the optical pH sensor resulted in the extension of the useful working range of the sensor throughout the full calibration range used (pH 2.51–9.67) as compared with its limited linear response (pH 5.0–7.25). The ANN used was a three-layer feed forward ANN using a recursive prediction error (RPE) algorithm. Another example sensor which has its operational range extended is an optical fibre humidity sensor. Brook et al. reported a technique [31], using a similar ANN to that used by Taib et al., that was used to linearise the response of the sensor resulting in an extension of the linear range from 40–55 to 40–82% relative humidity (RH). Other applications of ANNs used in conjunction with optical fibre sensors include the automatic calibration and measurand reconstruction of measurement systems. Massicotte et al. [32] applied a three-layered feed forward ANN with a modified back propagation algorithm for static calibration of measuring systems and for measurand reconstruction for a high pressure system (1–100 MPa). Other applications areas where feed forward ANNs have been used in conjunction with optical fibre sensors include high-precision three-dimensional position measurement [33] and impact damage detection in smart structure applications [34]. The multilayer feed forward network consists of three sections: an input layer section, one or more hidden layers section and an output layer. The input layer is used to receive the inputs and acts as a distribution centre by fanning out the inputs to the first hidden layer. Within the hidden layer section, each hidden layer will first activate and transform the data before propagating them to the next layer. A hidden layer neuron has the following tasks: firstly it sums up all its inputs, and then transforms the sum by a suitable non-linear transfer or activation function. Output layer neurons are normally taken to be the same as for neurons in the hidden layer, but as a result limit the dynamic range of the output to between +1 and −1. The non-linear behaviour of the hidden and output layers generates classifier behaviour [35]. Using SNNS, a three-layer feed forward neural network was implemented, see Fig. 9, consisting of a 12 node input layer, a 5 node hidden layer and a 6 node output layer. The latter includes one node to represent each sensor test condition. To determine the optimal size of the hidden layer used, multiple trials using hidden layers consisting of one node up to eight nodes were performed. In this investigation,

299

Fig. 9. Feedforward artificial neural network implemented in SNNS software.

a hidden layer of five nodes was found to perform best, see Table 2. The use of a feed forward network is a robust way of building a non-parametric classifier. The feed forward network was selected in favour of other more simple methods, e.g. single layer perceptron, as the OTDR output data investigated in this study is not linearly separable and a single layer perceptron cannot classify data that is not linearly separable [35]. An attempt was made to train an ANN with no hidden layer nodes, however this ANN implementation failed to train successfully. The application of the discrete Fourier transform to the extracted output peaks from the OTDR has achieved it aim as it has significantly reduced the computational resources of the artificial neural network in comparison to time domain-based OTDR results [36], see Table 3. Table 2 ANN hidden layer resources trials No. of hidden layer nodes

Result

Required no. of epochs

Two nodes Three nodes Four nodes Five nodes Six nodes Seven nodes Eight nodes

Failed to train Failed to train Trained successfully Trained successfully Trained successfully Trained successfully Failed to train

900 900 700 500 700 800 900

Table 3 ANN resources required with and without DFT signal processing

No DFT signal processing DFT signal processing

Input layer

Hidden layer

Output layer

253 12

30 5

9 6

300

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

Table 4 ANN test condition outputs Sensor 1

Sensor 2

Expected output

Observed output

Air Air Air Ethanol Ethanol Ethanol Water Water Water

Air Ethanol Water Air Ethanol Water Air Ethanol Water

0 0 0 0 0 0 1 1 1

0.023 0.023 0.023 0.017 0.017 0.017 0.970 0.970 0.931

5.2. Artificial neural network training The objective of training the network is to determine neuron activation thresholds and weight of the connections between connecting neurons, such that the predicted output, is as close to the desired output as possible. The discrepancy between the two is called the prediction error or residuals. Taib et al. presented a mathematical treatment of the calculation of prediction error in 1996 [30]. A total of 360 patterns are used to train the ANN, 40 patterns for each test condition listed in Table 1. To train the implemented feed-forward ANN, 500 epochs were required using a backpropagation algorithm with a momentum term, see Table 2. The momentum term of the backpropagation algorithm introduces the old weight change as a parameter for the computation of the new weight change. The new weight change is calculated using the formula: ωij (t + 1) = ηδj Oi + α ωij (t)

(5.1)

α is a constant term specifying the influence of the momentum term. This helps to avoid common oscillation problems when the error surface has a very narrow minimum area. The main advantage of a back-propagation algorithm is that flat spots of the error surface are traversed relatively rapidly with a few big steps. In this investigation the learning function used a learning parameter η of 0.9, a momentum term µ of 0.1 and a flat spot elimination value c of 0.1. The network was initialised with randomised weights and trained with a topological order update function [23].

0 0 0 1 1 1 0 0 0

1 1 1 0 0 0 0 0 0

0 0 1 0 0 1 0 0 1

0 1 0 0 1 0 0 1 0

1 0 0 1 0 0 1 0 0

0.032 0.032 0.032 0.961 0.961 0.961 0.021 0.021 0.061

0.945 0.945 0.945 0.022 0.022 0.022 0.009 0.009 0.008

0.051 0.031 0.931 0.023 0.031 0.931 0.051 0.017 0.970

0.002 0.956 0.061 0.032 0.956 0.061 0.002 0.961 0.021

0.947 0.013 0.008 0.945 0.013 0.008 0.947 0.022 0.009

sifications. The ANN, using the frequency domain response of the sensor as its input layer, accurately classifies all of the sensor’s test conditions. Based on the results shown in Table 4, the example of the ANN shown in Fig. 9 shows sensor 1 detecting air and sensor 2 detecting 50% ethanol. Although other methods of ANN detection have been reported, e.g. self organising maps [37], adaptive resonance theory (ART) [38], the feed forward multilayer ANN was found to be successful in this application. It is proposed in future work to investigate other detection methods and hence form a performance comparison with the feed forward multilayer ANN.

6. Conclusion A reliable measurement system based on optical fibre sensors for the purpose of fluid monitoring has been presented. A dual-element optical fibre sensor system based on a 1 km length of 62.5 ␮m core PCS fibre has been investigated and proven to be capable of detecting varying combinations of air, distilled water and 50% ethanol at each of the sensor elements, using OTDR-based techniques. Although the length of fibre used in this investigation was 1 km, longer or shorter lengths may be used as required. Artificial neural network techniques have allowed the resulting OTDR signals to be accurately determined using pattern recognition. The ANN implemented in SNNS successfully classified each sensor test condition correctly based on the frequency domain response of the sensor. Due

5.3. Artificial neural network testing and classifications In order to test the trained ANN, an independent set of data to that which was used to train the ANN is used. This data consists of the remaining 10 patterns for each of the test conditions listed in Table 1. The resulting 90 patterns were applied to the trained ANN and all patterns were classified correctly, see Table 4. From the results obtained in Table 4, it can be seen that the application of the discrete Fourier transform has achieved its aim. The computational resources of the ANN have been significantly reduced by the application of the DFT, in comparison with previous work by Lyons et al. [36], without affecting the accuracy of the ANN’s clas-

Table 5 ANN classification accuracy Sensor

Test conditions

Expected output

Accuracy (%)

Air Air Air Ethanol Ethanol Ethanol Water Water Water

Air Ethanol Water Air Ethanol Water Air Ethanol Water

0 0 0 0 0 0 1 1 1

94.5, 94.5, 94.5, 96.1, 96.1, 96.1, 97.0, 97.0, 93.1,

0 0 0 1 1 1 0 0 0

1 1 1 0 0 0 0 0 0

0 0 1 0 0 1 0 0 1

0 1 0 0 1 0 0 1 0

1 0 0 1 0 0 1 0 0

94.7 95.6 93.1 94.5 95.6 93.1 94.7 96.1 97.0

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

to the application of the Fourier transform to the OTDR peaks, refinements have been made to the resources used by the feed forward ANN, see Table 3, without affecting the accuracy of the ANN’s classifications, see Table 5.

[19] [20] [21]

Acknowledgements The financial support of the European Commission Fifth Framework Growth as part of the MICROPRO project: Contract GRD2-2000-30203.

References [1] A.D. Kersey, A review of recent developments in fiber optic sensor technology, Opt. Fiber Technol. 2 (1996) 291–317. [2] E. Udd, An overview of fiber optic sensors, Rev. Sci. Instrum. 66 (1995) 4015–4030. [3] K.T.V. Grattan, T. Sun, Fiber optic sensor technology: an overview, Sens. Actuators A 82 (2000) 40–61. [4] C.M. Davis, Fibre optic sensors: an overview, Opt. Eng. 24 (2) (1985) 347–351. [5] M.K. Barnoski, S.N. Jensen, Fiber waveguides: a novel technique for investigating attenuation characteristics, Appl. Opt. 15 (1976) 2112–2115. [6] A.D. Kersey, A. Dandridge, Distributed and multiplexed fibre optic sensor systems, J. Inst. Electron. Radio Eng. (JIERE) 58 (5) (1988) 99–111. [7] A.D. Kersey, Multiplexed fiber optic sensors, SPIE Proc. Distrib. Multiplexed Fiber Opt. Sens. II 1797 (1992) 161–185. [8] W.B. Lyons, E. Lewis, Neural networks and pattern recognition techniques applied to optical fibre sensors, Trans. Inst. Meas. Contr. 22 (5) (2000) 385–404. [9] D. King, W.B. Lyons, C. Flanagan, E. Lewis, An optical fibre sensor system for water quality monitoring utilising Fourier transform techniques and artificial neural network pattern recognition, in: Proceedings of the Irish Signals and Systems Conference 2003 (ISSC 2003), Limerick, Ireland, 30 June–2 July 2003, pp. 224–230. [10] D. King, W.B. Lyons, C. Flanagan, E. Lewis, An optical fibre ethanol concentration sensor utilizing Fourier transform signal processing analysis and artificial neural network pattern recognition, IOP J. Opt. A: Pure Appl. Opt. 5 (4) (2003) S69–S75. [11] S.K. Khijwania, B.D. Gupta, Fiber optic evanescent field adsorption sensor with high sensitivity and linear dynamic range, Opt. Commun. 152 (1998) 259–262. [12] B.D. Gupta, H. Dodeja, A.K. Tomar, Fibre optic evanescent field absorption sensor based on a U-shaped probe, Opt. Quantum Electron. 28 (1996) 1629–1639. [13] A.W. Snyder, J.D. Love, Optical Waveguide Theory, Chapman & Hall, 1983. [14] S.K. Khijwania, B.D. Gupta, Maximum achievable sensitivity of the fibre optic evanescent field absorption sensor based on a U-shape probe, Opt. Commun. 175 (2000) 135–137. [15] S. Otsuki, K. Adachi, T. Taguchi, A novel fibre optic gas-sensing configuration using extremely curved optical fibres and an attempt for optical humidity detection, Sens. Actuators B 53 (1998) 91–96. [16] M. Tateda, T. Horiguchi, Advances in optical time domain reflectometry, IEEE J. Lightwave Technol. 7 (8) (1989) 1217–1224. [17] E.G. Neumann, Optical time domain reflectometry: comment, Appl. Opt. 17 (11) (1978) 1675. [18] J.P. Dakin, A.J. King, Limitations of a single optical fibre fluorimeter system due to background fluorescence, in: Proceedings of the First

[22] [23]

[24]

[25] [26]

[27] [28] [29] [30]

[31]

[32]

[33]

[34]

[35] [36]

[37] [38]

301

International Conference on Optical Fibre Sensors, London, UK, 1983, pp. 195–199. G. Neumann, Analysis of the backscattering method for testing optical fibre cables, Arch. Elektron Ubertragungs 34 (1980) 157–160. EXFO IQ-7000 Optical Time Domain Reflectometer Instruction Manual, P/N: MAN-136-I.3ACE, December 1999. EXFO IQ-200 Optical Test System Instruction Manual, P/N: MAN-056-I.6ACE, October 1998. EXFO IQ-200 GPIB and IQ Applications Development Guide, P/N: MAN-075-I.3AN, January 2000. Stuggart Neural Network Simulator (SNNS), User’s Manual, Version 4.1, Report No. 6/95, 1995. http://www-ra.informatik.unituebingen.de/SNNS/ W.B. Lyons, H. Ewald, C. Flanagan, S. Lochmann, E. Lewis, A neural networks based approach for determining fouling of multi-point optical fibre sensors in water systems, Meas. Sci. Technol. 12 (2001) 958–965. P.A. Lynn, W. Fuerst, Introductory Digital Signal Processing, Wiley, July 1994. W.H. Press, W.T. Vetterling, S.A. Teukolsky, B.P. Flannery, Numerical Recipes in C: The Art of Scientific Computing, Cambridge University Press, 1990. W.S. Mc Culloch, W. Pitts, A logical calculus of the ideas immanent in nervous activity, Bull. Math. Biophys. 5 (1945) 115–133. D.E. Rumelhart, J.L. Mc Clelland, (Eds.), Parallel Distributed Processing, MIT Press, Cambridge, MA, 1986. M.J.P. Leiner, P. Hartmann, Theory and practice in optical pH sensing, Sens. Actuators B 11 (1) (1993) 281–289. M.N. Taib, R. Andres, R. Narayanaswamy, Extending the response of an optical fibre pH sensor using artificial neural networks, Anal. Chim. Acta 330 (1996) 31–40. T.E. Brook, M.N. Taib, R. Narayanaswamy, Immobilization of ruthenium tris-bipyridyl complex for chlorine gas detection, Sens. Actuators B 38–39 (1997) 195–201. D. Massicotte, S. Legendre, A. Barweiz, Neural network based method of calibration and measurand reconstruction for a high-pressure measuring system, IEEE Trans. Instrum. Meas. 47 (2) (1998) 362–370. Q. Yang, C. Butler, Sensor signal processing using neural networks for a 3-D fibre optic position sensor, Sens. Actuators A 41–42 (1994) 103–109. R.T. Jones, J.T. Sirkis, E.J. Freiebele, A.D. Kersey, Location of magnitude of impact detection in composite plated material using neural networks, SPIE Proc. Smart Struct. Mater.: Smart Struct. Process. Instrum. 2444 (1995) 469–480. H. Simon, Neural Networks: A Comprehensive Foundation, Prentice-Hall, New Jersey, 1999. W.B. Lyons, D. King, C. Flanagan, E. Lewis, A three sensor multipoint optical fibre water sensor utilising artificial neural network pattern recognition, in: Proceedings of the 15th Optical Fiber Sensor Conference (OFS 2002), Technical Digest, Portland, OR, USA, 6–10 May 2002, pp. 463–466. T. Kohonen, Self-Organization and Associative Memory, 3rd ed., Springer-Verlag, New York, 1988. G.A. Carpenter, S. Grossberg, Adaptive resonance theory (ART), in: M.A. Arbib (Ed.), The Handbook of Brain Theory and Neural Networks, MIT Press, Cambridge, MA, 1995, pp. 79–82.

Biographies D. King was born in Kilkenny, Ireland in 1979. He received a B.Eng. degree in electronic engineering from the University of Limerick, Ireland in 2001. He is currently undertaking a PhD at the University of Limerick based on work in the area of applying artificial neural network pattern recognition techniques to multipoint optical fibre sensor systems. He is a member of the Optical Fibre Sensors Research Group at the University of Limerick and his research interests include water quality

302

D. King et al. / Sensors and Actuators A 115 (2004) 293–302

monitoring, environmental monitoring, optical fibre sensors and artificial neural network-based pattern recognition.

M.Eng. in computer engineering, and his PhD from the University of Limerick in 1986, 1988 and 1991, respectively.

W.B. Lyons was born in Dublin, Ireland in 1973. He received a NCEA diploma in electronics (product development) for Dundalk Institute of Technology in 1993. Following this he worked in the semiconductor industry for SVGL as an installation and qualification engineer. He received a B.Eng. degree in electronic engineering from University of Limerick in 1998. He was awarded a PhD from University of Limerick in 2002 for work in the area of multipoint optical fibre sensors utilising artificial neural network pattern recognition techniques to separate out cross-coupling effects. He is a junior lecturer in the Department of Electronic and Computer Engineering at the University of Limerick. He is a member of the Optical Fibre Sensors Research Group and his research interests are environmental sensing, food quality assessment, and optical fibre sensors.

E. Lewis was born in Holyhead, Wales in 1959. He received a B.Eng. degree in electrical and electronic engineering from University of Liverpool in 1981. He was awarded a PhD from University of Liverpool in 1987 for work on high-speed photography and spectroscopy of electric circuit breaker arcs during the current zero phase. Following this, he worked as development engineer with BICC Telecom. Cables, Prescott, Merseyside in conjunction with University of Liverpool developing chromatic modulation-based optical fibre sensors for a wide range of applications. In 1989 he joined Liverpool John Moores University where he initiated the research activity in optical fibre sensors. The group investigated sensors for environmental monitoring including water contamination and pH. In 1996 he joined University of Limerick. He is head of Department of Electronic and Computer Engineering and director of the Optical Fibre Sensors Research Group, which he founded in 1996. The group is primarily engaged in investigating sensors for environmental monitoring (e.g. water quality, vehicle exhaust emissions, UV light intensity), food quality assessment and parameters associated with high power microwave sources (e.g. electric field, electron beam proximity).

C. Flanagan is a senior lecturer in the Department of Electronic and Computer Engineering at the University of Limerick. His research interests include neural sensor systems, computer architecture and network processors. He received his B.Eng. degree in electronic engineering, his