Performance Improvement of Multisensor Systems Using Autocompensation Strategy-Based LSTM

Abstract

Sensor drift has long been an issue that has perplexed researchers attempting to implement electronic noses (e-noses). However, for cases in which the target domain is inaccessible and the models must be updated frequently to adapt to data variations, routine subspace learning-based drift compensation approaches appear to be powerless. In this paper, we present a novel recurrent neural network consisting of layer-normalized long short-term memory cells with an autocompensation strategy, aiming to automatically generate compensation values based on external time drift factors without target domain data for drift compensation. Specifically, various strategies can be utilized to generate compensation values, and we present plans based on a quadratic function and a multilayer perceptron. In addition, this model can be updated based on previously trained models by feeding it new data without retraining the whole model. These advantages imply that the model is more flexible and more in line with actual application scenarios involving target domain inaccessibility. In the comparison experiments on a real benchmark sensor drift dataset, the proposed model obtained 89.64%, 94.15% and 77.40% classification accuracies for three different experimental settings, which are much greater than those of the comparison methods. These experimental results demonstrate that the proposed models can realize long-term prediction with higher accuracy than do some mainstream drift compensation approaches.

Topics & Concepts