Research On Counteraction Algorithms Of Long-term Drift In The Electronic Nose

In modern society, air pollution is one of the main problems related with humanlife in many scenes including indoor and in-car situations. Traditional air qualitydetection based on chemical gas analysis is not widely used because of its high cost,long time for detection and requirement for professional operations. On the contrary,electronic nose (EN) becomes a potential approach for traditional method for its fastdetection speed, convenience and so on. Current technology of ENs has made greatprogress, but still has a gap in practicability. Yet, the issues, namely long-term drift,background interference and individual differences, have not been solved well.According to the off-line analysis of long-term drift signals, long-term drift showsan uncertain and slow trend. Wilks criterion of principle component analysis (PCA)shows poor classification ability in principle component selection for visualization.And it is caused by neglecting the relationship between the samples at class plane andthe centre of gravity. To solve this problem, a new computation method has beenproposed for the centre of gravity in this research. Conventional orthogonal signalcorrection (OSC) cannot adjust algorithm model in on-line process. To adapt OSC toon-line work conditions, a modified OSC method with additional memory space andrecognition outcome process has also stated.Drift in EN can be categorized to two types: temporal and long-term drift and thelatter one is the focus in our research. In order to study long-term drift, an EN platformhas been established to collect the gas-sensor-array response involving long-term drifteffect. This platform is composed of a sampling system, an array of gas sensors and acomputer, these parts are combined together to perform different experiments. Thearray of gas sensors in the platform includes four gas sensors. In the work of gas sensorselection, the gas sensor responses under different heating voltages are examined.Considering the relationship between gas sensor characters and heating voltages,nonstandard heating voltages are used to heat some gas sensors to change theirselectivity and enhance the otherness of the responses of gas sensors.Long-term drift shows a slow and ruleless direction from off-line analysis. TheWilks rule of principle component analysis (PCA) ignores the effect of the samples onthe edge of classified planes to the center of gravity, thus the classification results arenot satisfied. In this research, a punishment mechanism is adopted to improve the computation of the gravity center. The experimental results show that the modifiedmethod avoids wrong selections of PCA components for classification.Orthogonal signal correction (OSC) cannot adjust algorithm model on-line.Memory space and recognition results process module are added to modified OSC towork on-line. The experimental results demonstrate that the modified OSC method candecompose long-term drift from on-line samples and minimize the effect of long-termdrift to gas sensor response compared with OSC.Conventional on-line anti-drift methods include multiple self-organizing maps(MSOM) and adaptive resonance theory (ART) network. These methods only adjusttheir neural weights for current sample in on-line retraining process, thus localcompensation problem may easily occur and make harm to the results of the driftcompensation. In this research, we have put forward a global compensation approachto overcome this drawback. During the retraining process, this method compensates allneural weights including the ones to current sample and the other ones. Theexperimental results show that the modified MSOM increases the accuracy of driftcompensation compared with MSOM under multiple kinds and discrete samplingcondition, while the modified ART network shows improved consistency ofrecognition outputs.The timeliness is an important index of on-line drift compensation. The timelinessof MSOM and ART can be optimized by getting rid of redundant retraining process.Thus two triggering methods of retraining have been proposed, one is based onadaptive mechanism and the other one is based on variable adaptive mechanism. Thefirst method adjusts the retraining interval by the error value changes of the currentsample. The results of experiments show that this method decreases the relationshipbetween initial parameter settings and timeliness and recognition rate. On the otherhand, the other method corrects the retraining interval according to the error valuechanges and the changing speed. Experiment results show that the method based onvariable adaptive mechanism keeps the advantages of the first one and also decreasesthe sensitivity to the setting values of parameters further.The algorithm would be invalid if the temporal state samples are used forretraining and it is called blind retraining problem. In order to solve this problem, weisolate the temporal samples from the retraining process using morphological andalgorithm descriptors. According to morphological descriptors method, somemorphological descriptors adapt to on-line conditions are discussed and real time slope is select as data selection criteria. Then, the algorithm obtains the slope from trainingset and determines the sample states by comparing the slopes value with slopethreshold. On the other hand, algorithm descriptors method uses target vector in earlytransition detection (ETD) model. This method obtains the thresholds according to thetendency of target vector and determines the sample states by comparing the targetvector value with the thresholds. The experimental results show that the two methodscan perform data choosing under fast sampling situation, and analysis points out thatETD descriptors method reveals better reliability under low sampling process.According to this idea, some morphological descriptors are checked. Aftercomparisons, the real-time slop of the response curve is selected as the criterion fordata selection. Experiment results show that this method can avoid the blind retrainingproblem and keep excellent drift compensation ability well. Additionally, the algorithmdescriptors in early transition detection method as the characters of neural network areused to estimate the sample state. The outcomes of the experiments demonstrate thatthis method can also avoid blind retraining problem.At the end of this study, the content above is summarized and it is pointed out thatthe current research is not satisfied in long-term drift orderliness, quantitativerecognition under drift condition and algorithm timeliness. The theory of long-termdrift and algorithm robustness should be studied further more in subsequent researches.