.png){kind=icon}
.png){kind=icon}
Return
Gas recognition method based on the deep learning model of sensor array response map
.png){kind=icon}
PRE
AI
DOI：10.1016/j.snb.2020.129349.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
It is important to detect and recognize the unknown gases or VOCs (Volatile Organic Compounds) in industrial safety issues. Electronic nose is a novel and portable method to detect the VOCs with high accuracy combined with sensor array and artificial intelligence algorithm. The results indicated that the multidimensional dynamic response signals of the sensor array can be viewed as the image form. Thus, a new method coupled dynamic response map with deep learning model (DLM) was proposed to improve the accuracy of the sensor array. The error-correcting output codes (ECOC) model with support vector machine (SVM) learners was applied to discriminate different VOCs. The results showed that the model with the data from the sensor array classified the VOCs more accurately than that with just single sensor. Further, a simple DLM network was trained to classify the VOCs with the accuracy of 92 %. Then the transferred VGG-19 model was further adapted to improve the generalization property of DLM with the accuracy of 90 %. Moreover, all sensors' responses at certain time were normalized before building the model, which enhanced the prediction accuracy to 96 % for simple DLM and 94 % for transferred VGG-19. Finally, the concentrations of different substances were predicted with SVM and DLM. The results showed that the prediction error of SVM and DLM with multidimensional response map is lower that with the data from single sensor. Therefore, it is a feasible tool to detect VOCs with just one sensor module using the response map-DLM method proposed in this research.
Keywords:
Gas detection
Electronic nose
Artificial olfactory
Deep learning model
Dynamic response
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
7.7
Papers:
3.3W
Citations:
12.6W