Balanced Distribution Adaptation for Metal Oxide Semiconductor Gas Sensor Array Drift Compensation

Abstract

Drift compensation is an important issue for metal oxide semiconductor (MOS) gas sensor arrays. General machine learning methods require constant calibration and a large amount of label gas data. At the same time, recalibration will cause a lot of costs, and label gas is difficult to obtain in practice. In this paper, a novel drift compensation method based on balanced distribution adaptation (BDA) is proposed. First, the BDA drift compensation method can adjust the conditional distribution and marginal distribution between the two domains through the weight balance factor, thereby more effectively reducing the mismatch between the two domains. When the BDA method performs classification tasks through machine learning, no labeled data is required in the target domain. Then, the particle swarm optimization algorithm is used to improve the accuracy of drift compensation. Individuals in the population are initialized randomly, and their fitness values are calculated. Iterative optimization of the population individuals is conducted until the optimal weight balance factor parameters are calculated. Finally, the BDA method is experimentally verified on the public gas sensor drift data set. Experimental results showed that the BDA method was significantly better than the existing joint distribution adaptation (JDA) method and other standard drift compensation methods such as K-Nearest Neighbor (KNN). In the two setting groups, the recognition accuracy was 4.54% and 1.62% ahead of the JDA method, and 12.23% and 15.83% ahead of the KNN method.

Review–Modern Data Analysis in Gas Sensors

“…Liu et al 75 extracted features using Fisher linear discriminant approach for drift compensation which decreases the inter-concentration discrepancy of the feature distribution, resulting in concentration independent features. Balanced distribution adaptation algorithm showed better performance over joint distribution adaptation according to Jiang et al 76 where the marginal and conditional distributions are considered. An earlier article 77 combined simple classifiers with these features, resulting in an improved accuracy.…”

“…KNN provided promising results (when integrated with enhance feature selection algorithms) with an increased accuracy up to 97.5% in a drifted E-nose dataset. 76 It was used alongside balanced distribution adaptation (BDA) optimization of features. In another study, KNN showed promising result in determining the perishable quality of shrimp when used with softmax regressor.…”

A Study of Drift Effect in a Popular Metal Oxide Sensor and Gas Recognition Using Public Gas Datasets

“…However, since there is not enough prior information, these methods usually cannot effectively handle samples that differ significantly from the initial distribution. Adaptive correction methods were applied to drift revision through classifier integration by [25] for the first time in 2012. According to the experiment, an ensemble method based on support vector machine was shown to deal well with sensor drifts and was certificated to fulfill the task better than baseline competing method.…”

“…Domain adaptation and transfer learning approaches [26][27][28] focus on removing the influence of cross-domain sensitivity to make the prediction model more robust. The balanced distribution adaptor drift compensation method improved the accuracy of drift compensation by reducing the discrepancy between the two domains, as it can adjust the conditional and marginal distributions between the two domains through a weight balancing factor [25]. Additionally, to correct for drift and device variation, a transfer learning-based drift correction autoencoder model for adaptive domain adaptation was proposed [29].…”

Sensor drift compensation for gas mixture classification in batch experiments

“…Furthermore, component‐correction methods are used to identify and eliminate drift‐sensitive components before building the model. However, owing to a lack of prior information, such methods are generally incapable of processing drifted samples that deviate significantly from the initial distribution 15,21 …”