Upload
timothy-west
View
217
Download
0
Tags:
Embed Size (px)
Citation preview
Real-Time Odor Classification Through Sequential Bayesian Filtering
Javier G. Monroy Javier Gonzalez-Jimenez
http://mapir.isa.uma.es
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
Toxic or dangerous chemicals
Localization of multiple gas sources
Gas distribution maps
In real-time
Absence of steady state
Desirable: uncertainty about class prediction?
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
Training: 4 classesTest: Only 1 class
Only instantaneous e-nose readings are used as features.
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
SEGMENTATION
SLIDING WINDOW
• Complex.• Influences the classification
performance.• Not real time.
• Optimization for selecting the appropriate window size.
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
BAYES FILTER: Probabilistic framework to efficiently integrate information from previous e-nose observations.
Bayesian network of the HMM for the real-time odor classification problem.
Bel(Ct) = P(Ct|Z1:t)Bel(Ct) = P(Ct|Z1:t)
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
Bel(Ct) = P(Ct|Z1:t)
Normalization constant
Number of class labelsObservation likelihood
Transition probability
Bayes rules + HMM conditional indep. assumptions
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
ps is the probability that at two consecutive instants of time the e-nose is exposed to the same odor ( that is, the same class).
Transition probability
Observation likelihood
We do not propose any new classifier.
To apply SBF to the posterior provided by any probabilistic classifier working on the most recent e-nose observation:
P(Ct|Zt)
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
Removing terms not dependent of P(Ct) + Marginal class probability
P(Ct) is time-independent
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
EXPERIMENTAL RESULTS
Real-time classification of volatiles in uncontrolled, real environments.
Naive Bayes classifier working on the instantaneous response of the e-nose.
DS-UMA: 4-classes, array of 5 MOX gas sensors.
DATASETS
DS-UCI: subset corresponding to the parameters L4, Vh=5, fan=100, down-sampled to 1Hz, and restricted to only 4 gas-classes.
No Mixture of gases, just Gas transitions!
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
EXPERIMENTAL RESULTS
For a 4-class problem, ps=0.25 indicates uniform class-transition probability.
ps=0.99 indicates that the most likely class at time «t», is that at time «t-1»
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
EXPERIMENTAL RESULTS
SBF is specially effective when the posterior probabilities havesimilar values, but not so much when one class has a much
higher probability that the others (as is desired).
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
EXPERIMENTAL RESULTSWITH ARTIFICIALLY SIMULATED CHEMICAL TRANSITIONS
Real-Time Odor Classification Through Sequential Bayesian Filtering http://mapir.isa.uma.es
Sequential Bayesian filtering (SBF) approach to the real-time classification of volatile substances.
Integration of information from previous e-nose observations, without relying on segmentation or sliding window approaches.
Evaluation with two public olfaction datasets (chemical transitions).
Results show that SBF is particularly suitable for high dynamic environments where spurious class transitions produced by the sensor dynamics are effectively removed.
CONCLUSIONS
Test alternative classifiers (SVM).
Comparison with sliding window techniques.
Adaptive approaches to SBF to automatically adjust the parameters of the model (e.g. class transition probabilities).
FUTURE RESEARCH
Real-Time Odor Classification Through Sequential Bayesian Filtering
THANK YOU!
http://mapir.isa.uma.es