Early-Stage Gas Identification Using Convolutional Long Short-Term Neural Network with Sensor Array Time Series Data

Area of Science:

• Computational intelligence
• Pattern recognition
• Chemical sensing

Background:

• Gas identification commonly relies on equilibrium sensor responses or full time-series data.
• Utilizing diverse gas sensing kinetics during transitional phases offers potential for improved identification.

Purpose of the Study:

• To develop a computational intelligence-based meta-model for automatic feature extraction and gas identification from time-series data.
• To incorporate temporal dependencies within time-series data for enhanced gas identification performance and reliability.

Main Methods:

• A convolutional long short-term memory (CLSTM) neural network was developed to analyze time-series data from gas sensor arrays.
• The CLSTM approach was compared against baseline models like multilayer perceptron (MLP) and support vector machine (SVM).

Main Results:

• The CLSTM model achieved a classification accuracy as high as 96%, demonstrating enhanced accuracy and robustness.
• CLSTM showed excellent gas identification performance even at early stages of gas exposure, indicating practical significance.

Conclusions:

• The proposed CLSTM approach effectively leverages temporal characteristics in time-series data for superior gas identification.
• This method holds significant promise for real-time gas sensing applications, validated through comprehensive testing.