Multi-Dimensional Encoding Image Based Time Series Feature Extraction for Improved Colorimetric Gas Concentration Estimation

Abstract

Industrial risk can be reduced through the early detection and intervention of gas leaks. Numerous early detection methods have been developed and applied in industrial areas, but they require expert analysis skills, a lengthy analysis time, or relatively expensive equipment. The development of novel, efficient, and reliable devices is continuously required as a strategy to reduce industrial gas risks.In the past decades, various types of gas sensors have been developed based on semiconductor, optical fiber, acoustic wave, and other technologies. A significant amount of work has been undertaken in recent years to make sensors smaller, more sensitive, more visually responsive, easier to handle, more personal, and more wearable. However, these functions require expensive equipment, complex skills, and large amounts of power. In this respect, colorimetric sensors are of great interest owing to their simplicity, easy fabrication, low cost, visual response, and ease of interpretation without power consumption. However, currently, colorimetric sensors have several limitations that need to be overcome, such as sensitivity, high cost, early detection, and accuracy for use in industrial safety systems.Existing colorimetric sensor studies use an algorithm that compares discoloration using the naked eye, estimates gas concentration with a simple statistical method, or uses a machine learning algorithm to estimate time series data by signal processing simple discoloration information. The algorithm learned the measurement image or the time series data extracted from the image. Signal processing was focused on noise removal, which was performed by scaling the time series data. These studies using machine learning and deep learning-based analysis for colorimetric sensor research have not sufficiently considered data coding. Machine learning of time series data is carried out using the time series decomposition method that separates time dependencies. Attempts are being made to perform time series data analysis through image encoding. To the best of our knowledge, the present study is the first to use image encoding for time series colorimetric sensor data.This study applied a time series data analysis technique using a convolutional neural network (CNN) with multi-dimensional image encoding to improve gas concentration estimation accuracy. The estimation accuracy obtained in this study was more than that obtained by applying existing statistical methods, such as auto-regressive integrated moving average (ARIMA) and exponential smoothing (ETS), and machine learning techniques such as multi-layer perceptron (MLP). We focused on gas concentration estimation using a time series-based colorimetric analysis method, which was developed based on colorimetric fabric detection data. The time series data obtained through the gas experiment were subjected to image encoding to classify the trend in the color value change of the gas concentration. Image encoding improves the efficiency with which the gas exposure section is specified, as well as the gas concentration estimation, within the entire section. The concentration estimation accuracy of the existing data classification method was 74.8%, while that of MLP was more than 80.5%. Furthermore, the accuracy was improved by applying a CNN that learned the multi-dimensional processing of sensor data through image encoding. From the results, an estimation accuracy of 90.1% was confirmed. The proposed technique can be applied to next-generation smart industrial safety systems. Figure 1