TL;DR
This paper introduces a novel electronic nose (eNose) algorithm that can identify chemicals independently of their concentration and estimate that concentration, even for previously unseen levels. It also incorporates a "reject option" to identify samples from unknown chemical types, offering a more robust and versatile solution for chemical sensing.
What Problem Does This Paper Address?
Electronic noses (eNoses) are analytical devices that use an array of broadly tuned chemical sensors to detect and identify airborne chemicals. A major challenge in eNose applications is the dependency of sensor responses on the concentration of the detected chemical. Traditional classification algorithms, trained on specific concentrations, often struggle to accurately identify chemicals when their concentration deviates from the training data. This limitation hinders their practical use in environments where concentration control is difficult or where precise concentration information is crucial. Furthermore, existing systems are typically unable to identify chemicals that were not part of their initial training set. This paper addresses these limitations by proposing an algorithm that can identify chemicals irrespective of their concentration and estimate that concentration, while also being able to flag unknown substances.
Methods
The proposed algorithm is inspired by Hopfield's work on olfactory processing in the brain. It leverages the multiplicity of sensors in an eNose system. The core idea is to treat each sensor as a "concentration sensor" by establishing an invertible relationship between a sensor's response and the concentration of a specific chemical.
- Training Phase: For each known chemical (candidate), the eNose is exposed to it at several different concentrations. The response of each sensor is recorded. These data points are used to construct a function, $r_{ij} = f_{ij}(c_j)$, where $r_{ij}$ is the response of sensor $i$ to candidate $j$ at concentration $c_j$. Crucially, these functions must be monotonic to allow for inversion. Piecewise cubic spline interpolation is used to model these functions.
- Classification and Concentration Estimation: When a new sample is presented, the algorithm uses the inverse function, $c_{ij} = f_{ij}^{-1}(R_i)$, to estimate the concentration of candidate $j$ based on the response $R_i$ of sensor $i$. This is done for all sensors and all candidates.
- Identity Determination: For each candidate $j$, an average estimated concentration $\hat{c}_j$ is calculated across all sensors. The variance $\sigma_j$ of these estimates is then computed. If $\sigma_j$ is below a predetermined threshold ($\sigma_0$), the sample is identified as candidate $j$, and its concentration is estimated as $\hat{c}_j$.
- Reject Option: If the variance $\sigma_j$ for all candidates exceeds the threshold, the sample is rejected as an unknown chemical type.
- Robustness Enhancements: The algorithm also includes checks for responses exceeding maximum expected values ($r_{max}$) or falling below minimum expected values ($r_{min}$), flagging such estimates as "INVALID" to improve reliability. A count of "INVALID" estimates for each candidate is also used to further refine the rejection process.
Datasets
The algorithm was tested using two datasets:
- Training Dataset: This dataset consisted of 10 candidate chemicals, each measured at six different concentrations. The concentrations were expressed as molar fractions in a PEG-400 solution. A total of 269 measurements were collected.
- Validation Dataset: This dataset was designed to rigorously test the algorithm's performance on unseen data and comprised three groups:
- Group (a): Six candidates from the training set, measured at concentrations that were also present during training.
- Group (b): Seven candidates from the training set, measured at concentrations not present during training.
- Group (c): Four non-candidate chemicals (unknowns), measured at seven different concentrations. This validation dataset contained a total of 40 measurements.
Key Insights and Ideas
- Sensor as Concentration Meter: The core innovation is treating each sensor's response not just as a qualitative indicator but as a quantitative measure of concentration for a specific chemical, provided the response is monotonic.
- Variance as Identity Indicator: The algorithm cleverly uses the variance of concentration estimates across multiple sensors to distinguish between a true match (low variance) and a false alarm (high variance). This is a powerful way to leverage the collective information from the sensor array.
- Concentration Independence: The ability to identify chemicals regardless of their concentration is a significant leap forward, addressing a fundamental limitation of many existing eNose systems.
- Implicit Reject Option: The variance-based decision threshold naturally provides a mechanism to reject samples that do not fit any of the trained chemical profiles, enhancing the system's reliability and safety.
Key Results
- Training Data Performance: The algorithm achieved 100% correct classification and an average relative error of 1.4% for concentration prediction on the training dataset.
- Validation Data Performance:
- Known chemicals at known concentrations (Group a): 100% correct classification with a 7.8% average relative error.
- Known chemicals at unknown concentrations (Group b): 100% correct classification with an 11.8% average relative error. This is highlighted as a particularly exciting result.
- Unknown chemicals (Group c): An 83.3% correct classification rate (treating rejection as correct classification), with only two misclassifications.
Limitations and Future Work
- Monotonicity Requirement: The algorithm relies on the sensor responses being monotonic with respect to concentration. If this is not the case for a particular sensor-analyte pair, the inversion step becomes problematic.
- Number of Sensors: While 16 sensors were sufficient for good results, the authors suggest that increasing the number of sensors is expected to further improve performance and robustness, similar to how more samples improve statistical inference.
- Single Free Parameter ($\sigma_0$): The algorithm currently uses a single global threshold $\sigma_0$. The authors propose that introducing candidate-specific thresholds ($\sigma'_j$) or even candidate-specific, concentration-dependent thresholds could further enhance performance by accounting for the "natural" variance of different chemicals.
- Generalizability: The authors believe the algorithm's applicability extends beyond eNoses to any multisensor system with monotonic responses, suggesting broad potential.
Summary updated