Travel & Places

Evaluation of a radial basis function neural network for the determination of wheat quality from electronic nose data

Evaluation of a radial basis function neural network for the determination of wheat quality from electronic nose data
of 11
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  Ž . Sensors and Actuators B 69 2000 348– r locate r sensorb Evaluation of a radial basis function neural network for thedetermination of wheat quality from electronic nose data Phillip Evans a , Krishna C. Persaud a, ) , Alexander S. McNeish b , Robert W. Sneath c ,Norris Hobson c , Naresh Magan d a ( ) ( ) Department of Instrumentation and Analytical Science DIAS , Un Õ ersity of Manchester Institute of Science and Technology UMIST , Chemistry Tower,Faraday Building, Sack  Õ ille Street, Manchester, M60 1QD, UK  b Osmetech plc, Electra House, Electra Way, Crewe, CW1 6WZ, UK  c Silsoe Research Institute, Wrest Park, Silsoe, Bedford, MK45 4HS, UK  d Cranfield Biotechnology Centre, Cranfield Uni Õ ersity, Cranfield, Bedfordshire, MK43 0AL, UK  Received 14 October 1999; accepted 8 February 2000 Abstract Odorous contaminants in wheat have been detected using a conducting polymer array. A radial basis function artificial neural network  Ž . RBFann was used to correlate sensor array responses with human grading of off-taints in wheat. Wheat samples moulded by artificialmeans in the laboratory were used to evaluate the network, operating in quantitative mode, and also to develop strategies for evaluatingreal samples. Commercial wheat samples were then evaluated using the RBFann as a classifier network with great success, achieving a Ž . predictive success of 92.3% with no bad samples misclassified as good in a 40-sample population 24 good, 17 bad using a training set Ž . of 92 samples 72 good, 20 bad .  q 2000 Elsevier Science S.A. All rights reserved. Ž . Keywords:  Radial basis function; Artificial neural network; Electronic nose; Wheat quality; Sensors; Conducting polymer s 1. Introduction Cereal grains, wheat in particular, represent one of themost important crops in global terms. There is a significantrequirement to ensure the organoleptic quality of suchcrops to ensure good commercial returns and ensure safety w x of the product 1,2 .Often, the initial screening procedure is on the basis of an inspector’s or buyer’s olfactory perception. However,this is a subjective measurement and is unsuitable inmodern agricultural economies where large financial gainsand loses can be made as a result of changing the gradingof a crop.The European ISO605 standard and American United Ž . States Department of Agriculture USDA grain gradingprocedures in place relate to the odour determination of grains but these are vague and do not define any standard ) Corresponding author. Tel.:  q 44-161-200-4912; fax:  q 44-161-200-4879. Ž .  E-mail address: K.C. Persaud . or surrogate odours against which malodorous samplesmay be compared or against which inspectors and buyers w x can be trained 3–5 . Attempts have been made to qualifysome of the odour descriptors used in the industry todescribe malodorous samples but a consensus has not yet w x been reached, even for the most typical malodours 6 .Consequently, there are strong commercial reasons fordeveloping an odour sensing system for the determinationof grain quality at points of transfer and purchase.Allied to and perhaps more important than commercialconsiderations are the health and safety aspects involved in‘sniffing’ grain. Concurrent with the well-documented po-tential mycotoxin dangers are the chronic respiratory dan-gers associated with exposure to small particulates and w x possibly toxic volatile organic compounds 2 .Fungal contamination is the major problem leading todowngrading or rejection of grain. This is usually as aresult of poor or inadequate storage after harvesting. Grainaffected in this way is often described as being musty orsour smelling. Early detection of fungally infected graincan in some cases be remediated and the crop saved. 0925-4005 r 00 r $ - see front matter q 2000 Elsevier Science S.A. All rights reserved. Ž . PII: S0925-4005 00 00485-8  ( )P. E  Õ ans et al. r Sensors and Actuators B 69 2000 348–358   349 Numerous studies have been carried out that have inves-tigated the association between fungal infections, storageconditions and the volatiles evolved by the various species w x of fungi 7–9 .A number of workers have already reported attempts atmeasuring cereal grain quality, specifically wheat, using w x volatile chemical sensor array based technologies 10–14 .Our approach has been to evaluate wheat moulded undercontrolled conditions using a variety of sampling ap-proaches and then to develop suitable sensors and proto-cols for in-line r at-line monitoring of grain at the point of transfer. For the task at hand, there is a need to map thesensor array response to human organoleptic parametersthat determine whether a particular wheat sample is to be w x accepted or rejected 6 . In addition, there is a requirementfor any system developed to provide a rapid and reliableanswer to the operator r inspector. The system must berobust and reliable with a minimum of maintenance anddata interpretation and also operate in real time.A variety of pattern recognition techniques includingneural networks may be applied to the classification of different odours and quantitative prediction and recogni-tion of unknown gases and odours. Backpropagation, amodel of multilayer perceptron networks, is probably themost widely used neural network paradigm. One disadvan-tage of this model is the difficulty in classifying a previ-ously unknown pattern that is not classified to any of theprototypes in the training set. This paper focuses on theapplication of a rapid data interpretation system using a Ž . radial basis function artificial neural network RBFann to w x map grain odour to human organoleptic perception 15–17 .RBF networks train rapidly, usually orders of magni-tude faster than backpropagation, while exhibiting none of backpropagation’s training pathologies such as paralysis or w x local minima problems. A RBF network 15 is a two-layernetwork where the output units form a linear combination Ž . of the basis functions computed by hidden units Fig. 1 .The basis functions in the hidden layer produce a localised Fig. 1. Schematic network to represent the basic RBFann architecture,where  x  ,  x  ...  x  are input neurones with the non-linear function 1 2 m Ž embedded in the hidden layer, and  S   is the linear combiner  l  – l  are 1 c . the hidden layer outputs,  l  is the weighting factor . 0 response to the input and typically, uses hidden layerneurones with Gaussian response functions: y  2 F y   s exp  y  1 Ž . Ž . 2 ž / b  where  b   is a real constant. The outputs of the hidden unitlie between 0 and 1; the closer the input to the centre of the Gaussian, the larger the response of the node.The activation level of an output unit: O  s Ý W O  2 Ž .  j ji i where  W   is the weight from hidden unit  i  to output unit  ji  j , forms a linear combination of the non-linear basis func-tions.Finding the centres, widths, and the weights connectinghidden nodes to the output nodes does the training in aRBF network. The performance of radial basis functionclassifiers is highly dependent on the choice of centres andwidth. This has been the focus of our attention in order to w x optimise RBF networks for odour classification 16,17 .For a minimum number of nodes, the selected centresshould be closely representative of the training data foracceptable classification.The subtleties and complexities in the optimisation of  w x the RBFann are dealt with elsewhere 16,17 . This paperfocuses on how the network may be applied to rapid anduseful classification of real data from electronic nose datato human psychophysical parameters. 2. Method 2.1. Moulding of wheat samples Ž Field-harvested wheat Silsoe Research Institute, Beds., . UK was modified to different water contents using amoisture sorption isotherm. Known amounts of water wereadded to grain samples in 500-ml flasks and stored at 4 8 Cfor 24 h with regular shaking to obtain an even moisturecontent. The grain was then incubated at 25 8 C and re-moved when visible moulding was noticed in the wettestsamples. This enabled a range of samples from unmouldedto completely moulded to be obtained. When not in use,the wheat was stored in a refrigerator at 5 8 C to arrest thegrowth of the fungi present. Wheat was incubated at 18%, Ž . 20% and 25% moisture contents mc .After moulding, the wheat samples were divided intotwo batches, one supplied ‘as moulded’ and the otherair-dried to a nominally constant moisture content of 15%. Ž . Colony counts colony forming units, CFU were mea-sured using standard techniques as follows. Ž . Sub-samples 1 or 5 g were weighed and placed in 9 or95 ml of 0.1% water agar diluent. A serial dilution serieswas carried out and 0.1 ml spread plated onto 2% malt Ž extract agar, and 2% malt, 10% salt low water availabil-  ( )P. E  Õ ans et al. r Sensors and Actuators B 69 2000 348–358  350 . ity using a sterile bent spreader. The plates were incu-bated at 25 8 C for up to 7 days and the fungal coloniescounted where they appeared at between 10 and 100 perplate. The major dominant species were identified. Ž . Commercial wheat samples UK grown were obtainedfrom a commercial grain trader via Silsoe Research Insti-tute. The samples were provided with some analytical datasuch as moisture content, protein content, density, Hagburgfalling number and screening measurements and were usedas received. Reject grains typically only contained datapertaining to their moisture content and density. 2.2. Odour measurement basis The basis of the odour sensing system has previously w x been described elsewhere in Ref. 18 . The odour emittedby the wheat samples was actively sampled by passingpre-conditioned air through the sample and then over a Ž . sensor array see later . The array consisted of 32 conduct-ing polymer elements. Each element possessing a broadspecificity with overlap between the responses of all 32elements. The sensor array response was recorded eachsecond and data transmitted to an IBM compatible PC. Ž Raw data was processed using Osmetech software v3.1, . Osmetech, Crewe, UK to produce normalised patterns forinput to the RBFann. 2.3. Radial basis function artificial neural network  We have developed a RBFann adapted for rapid odourclassification. This was evaluated for its ability to discrimi-nate between mouldy and good wheat samples. It wasinitially trained using data from wheat samples producedby moulding under controlled conditions. Later, sampleswere obtained from a commercial supplier to provide amore representative population of wheat encountered un-der normal conditions and the correlation between theinspectors evaluation and the sensor array output evalu-ated. 2.4. Sensing apparatus2.4.1. Manual sampler for e Õ aluating laboratory moulded wheat  Ž . An Osmetech formerly Aromascan 32-sensor arraymounted inside an A8S sample station for extra tempera- Ž . ture control was used Osmetech, Crewe, UK . Wheat Ž . samples 50 g were weighed into screw neck Pyrex tubes Ž . Fisher Scientific, Loughborough, UK fitted with phenolicplastic lids when not in use. A sparging system was usedto transfer wheat odour form the sample tube to the sensorarray. The sensor array was set to sample reference air for1 min, then to sample wheat odour for a period of 2 min.This was followed by a wash cycle of 100% RH air at30 8 C for 1 min and finally, a return to sampling reference Ž  y 1 . air for 1 min all at 160 ml min . The reference air wasmaintained at a humidity of 30% RH at 30 8 C and wasswitched to the sample tube during the wheat odour mea-surement period. The A8S was held at 30 8 C while thearray itself was held at 35 8 C throughout. In total, includingthe 5-min sample equilibration time at 30 8 C, each vial took 10 min to be processed. From the raw data, odour patterndatabases were constructed using averaged time slices overthe last 20 s of the odour exposure or using three 5-s slicestaken at the end of the exposure to the sample, as appropri- Ž . ate. Principal components analysis PCA maps were cal-culated for comparison with the RBFann output. 2.4.2. Autosampler arrangement for e Õ aluating commer  - cial wheat  An Osmetech 32-sensor array coupled with autosamplerfacility was used to obtain the data presented for the Ž . commercial wheat samples Osmetech, Crewe, UK . Sam-ples were 10 g aliquots weighed into standard sample vials Ž . 22 ml, Osmetech, Crewe, UK and crimp sealed with Fig. 2. Sammon map of wheat moulded under controlled conditions when mapped along with standards.  ( )P. E  Õ ans et al. r Sensors and Actuators B 69 2000 348–358   351Table 1Colony counts for mould on the surface of the wheat moulded undercontrolled conditionsCFUs Ž . Moisture content % Log counts 10 First batch 12.5 6.81915 5.49518 6.37520 7.39725 6.765 Second batch 6.8195.49514.2 6.86915.7 7.43718.2 7.284 aluminium hole caps lined with polytetrafluroethylene Ž . PTFE  r silicone septa. Prior to each set of samples, two Ž . distilled water vials 5 ml each were run to ensure thatcontamination from any previous samples used were notcarried over during the measurement. A distilled water Ž . sample 5 ml was also run at the end of each batch. Thesensor array was set to sample for a period of 3 min, at aflow rate of 60 ml min y 1 , followed by a wash cycle of distilled water for 4 min at 160 ml min y 1 . The referenceair was maintained at a humidity of 30% RH at 30 8 C whilethe wash was again 100% RH air at 30 8 C. The platen wasmaintained at 30 8 C, the sample loop at 40 8 C and thetransfer line was at 50 8 C throughout the experiment. Thesensor array and data processing arrangements were asdescribed in Section 2.4.1. 2.4.3. E  Õ aluation of RBFann using measurements of artifi - cially moulded wheat  The RBFann can be used in quantitative or qualitativemode. The network was used in quantitative mode andtrained to predict values from the sensor outputs for three Ž . possible classification scenarios. These were: i angle of  Ž . clusters from the 0,0 srcin of a Sammon map see Fig. 2 ; Ž . Ž . ii log of CFU on the wheat samples; and iii the 10 moisture content of the wheat when moulded.The data used for the trial of the RBFann was basedupon 20 wheat sample runs using the manual samplingA8S system. The samples were four aliquots each of 12.5%, 18%, 20% and 25% mc laboratory moulded wheat.The data were taken as three 5-s slices at the end of sampling period rather than averaged over the period as forthe autosampled wheat data, thereby creating three sets of data per run and a total of 60 data sets for the evaluation intotal. The data order was randomised and the first 15samples taken as the training set in the first instance andthe remaining 45 samples used as the unknowns. Afternetwork evaluation, the data was re-sorted into order to aidclarity when presenting the results. 3. Results 3.1. Artificially moulded wheat  At 12% and 15% mc the fungi were almost entirelyfield fungi with a very small population of   Penicillium spp. present, whereas at 18% and 20% mc the predominantfungi present were  Penicillium  spp. At 25% mc, there was Ž Fig. 3. Distribution of fungal species in wheat moulded under controlled conditions NB no accurate identifications were carried out at 12.5% but these . were almost entirely field fungi with only a trace of   Penicillium  spp. present .  ( )P. E  Õ ans et al. r Sensors and Actuators B 69 2000 348–358  352Fig. 4. Comparison of RBFann training errors for each of the four classes and three classification scenarios used for wheat moulded under controlled Ž . conditions open bars are for 15 data points in the training set and the hatched bars are for 45 data points in the training set . Ž . a broader distribution of fungi see Table 1 and Fig. 3with  Acremonium  sp. dominating. 3.2. E  Õ aluation of RBFann using measurements of artifi - cially moulded wheat  Artificially moulded wheat was used to evaluate theperformance of the network operating in quantitative mode.The data presented in Fig. 4 is for the error in prediction of the learning set vs. the actual values for the three dataclasses evaluated, while the data presented in Fig. 5 is thecorresponding error output for the unknowns presented tothe network after the initial training phase. Data is pre-sented for two training and unknown scenarios. The openbars represent a training set of 15 data points and anunknown set of 45 data points whereas the grey barsrepresent the reversed case of 45 training data points and15 unknown data points. Note that the data was evaluatedin a random order and resorted after processing for clarity.The data in Fig. 6a–b represents the success of theRBFann in predicting the differences between wheat sam-ples moulded under controlled conditions using naturallyoccurring fungi. The dotted lines represent the error inprediction associated with each class while the grey barsare the predicted values based upon the sensor valuesproduced on exposure. The data presented in Fig. 6a are Fig. 5. Comparison of RBFann predictive errors for each of the four classes and three classification scenarios used for wheat moulded under controlled Ž . conditions open bars are for 15 data points in the training set and the hatched bars are for 45 data points in the training set .
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks