Active Intracavity Mixed Gas Inversion Algorithm Based on Multi-Task Learning (Invited)

Kun Liu; Hui Yin; Junfeng Jiang; Tiegen Liu; Chengwei Zhao

doi:10.3788/LOP231913

Laser & Optoelectronics Progress, Volume. 61, Issue 3, 0306002(2024)

Active Intracavity Mixed Gas Inversion Algorithm Based on Multi-Task Learning (Invited)

Kun Liu^1,2,3、*, Hui Yin^1,2,3, Junfeng Jiang^1,2,3, Tiegen Liu^1,2,3, and Chengwei Zhao^1,2,3

Author Affiliations

¹School of Precision Instruments and Opto-Electronics Engineering, Tianjin University, Tianjin 300072, China

²Key Laboratory of Opto-Electronics Information Technology, Ministry of Education, Tianjin University, Tianjin 300072, China

³Institute of Optical Fiber Sensing of Tianjin University, Tianjin 300072, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Schematic diagram of gas sensing system based on two-stage amplified thulium-doped fiber ring cavity laser

Download full size

Fig. 2. Physical diagram of gas sensing system based on two-stage amplified thulium-doped fiber ring cavity laser

Download full size

Fig. 3. Output spectra of two-stage amplified thulium-doped fiber ring cavity laser

Download full size

Fig. 4. Absorption spectra of mixed gas. (a) Absorption spectrum after baseline removal in wavelength range1; (b) absorption spectrum after Lorentz fit in wavelength range1; (c) absorption spectrum after simulation using HITRAN data in wavelength range1; (d) absorption spectrum after baseline removal in wavelength range2; (e) absorption spectrum after Lorentz fit in wavelength range2; (f) absorption spectrum after simulation using HITRAN data in wavelength range2

Download full size

Fig. 5. One-dimensional convolution and one-dimensional maximum pooling operations. (a) Convolution; (b) maximum pooling

Download full size

Fig. 6. LSTM cell structure diagram

Download full size

Fig. 7. Structure of the MTL-1DCNN-LSTM model

Download full size

Fig. 8. MTL-1DCNN-LSTM model learning results. (a) Loss function plots for training and validation processes; (b) confusion matrix for gas classification; (c) comparison of NH₃ predicted values with true values during the testing process; (d) comparison of CO₂ predicted values with true values during the testing process

Download full size

Fig. 9. Single-task learning training and validation process loss function curves. (a) Classification task; (b) regression task

Download full size

Table 1. Data distribution of mixed gas
View table
Table 1. Data distribution of mixed gas
Gas NH₃ volume fraction /% CO₂ volume fraction /% Number
NH₃+CO₂ 0.2 4.5 60
0.4 4.0 60
0.6 3.5 60
0.8 3.0 60
1.0 2.5 60
1.2 2.0 60
1.4 1.5 60
1.6 1.0 60
1.8 0.5 60

Table 2. Multi-task learning
View table
Table 2. Multi-task learning
Hardware name Detailed information
Computer operating system Windows 10 64
CPU Intel（R）Core（TM）i5-7200U CPU @ 2.50GHz
GPU NVIDIA GeForce 940MX
Random access memory（RAM） 4 GB

Table 3. Model hyperparameter settings
View table
Table 3. Model hyperparameter settings
Hyperparameter name Value
Learning rate 0.001
Batch-size 8
Epochs 150
Optimizers Adam
$ω_{1}$ 1
$ω_{2}$ 10

Table 4. Comparison of multi-task learning, single-task learning, BPNN, and SVM gas inversion performance

View table

Table 4. Comparison of multi-task learning, single-task learning, BPNN, and SVM gas inversion performance

Algorithm	A_cc /%	R²（NH₃）/%	R²（CO₂）/%	R_MSE（NH₃）/%	R_MSE（CO₂）/%	Inference time /s
MTL-1DCNN-LSTM	100	99.84	99.62	0.025	0.089	0.922
STLC-1DCNN-LSTM	100	‒	‒	‒	‒	0.578
STLR-1DCNN-LSTM	‒	99.53	98.62	0.040	0.118	0.875
SVM	95.00	‒	‒	‒	‒	‒
SVR	‒	98.37	98.13	0.111	0.297	‒
BPNN	98.89	‒	‒	‒	‒	‒
BPNNR	‒	93.57	85.92	0.148	0.594	‒

Table 5. Comparison of LSTM usage methods

View table

Table 5. Comparison of LSTM usage methods

Algorithm	A_cc /%	R²（NH₃）/%	R²（CO₂）/%	R_MSE（NH₃）/%	R_MSE（CO₂）/%	Inference time /s
MTL-1DCNN-LSTM	100	99.84	99.62	0.025	0.089	0.922
MTL-1DCNN	98.89	99.36	97.88	0.380	0.179	0.838
MTL-1DCNN-DLSTM	100	99.87	99.33	0.023	0.132	1.109

Tools

Get Citation

Copy Citation Text

Kun Liu, Hui Yin, Junfeng Jiang, Tiegen Liu, Chengwei Zhao. Active Intracavity Mixed Gas Inversion Algorithm Based on Multi-Task Learning (Invited)[J]. Laser & Optoelectronics Progress, 2024, 61(3): 0306002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites