Weak Life Signal Recognition Based on Deep Neural Network

Yan Li; Liang Li; Chenyu Zhao; Yulu Zhang; Yun He; Pei Liang

doi:10.3788/AOS240906

Acta Optica Sinica, Volume. 44, Issue 21, 2117001(2024)

Weak Life Signal Recognition Based on Deep Neural Network

Yan Li¹, Liang Li¹, Chenyu Zhao¹, Yulu Zhang^2、*, Yun He^3、**, and Pei Liang^1、***

Author Affiliations

¹College of Optical and Electronic Technology, China Jiliang University, Hangzhou 310000, Zhejiang , China

²School of Electronics & Information Engineering, Nanjing University of Information Science & Technology, Nanjing 210044, Jiangsu , China

³School of Information Engineering, Wuhan University of Technology, Wuhan 430070, Hebei , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Millimeter wave radar data feature extraction. (a) Original heat map of positive data; (b) amplitude-focus map of positive data; (c) static filter map of positive data; (d) normalized curve plot of positive data; (e) original heat map of negative data; (f) amplitude-focus map of negative data; (g) static filter map of negative data; (h) normalized curve plot of negative data

Download full size

Fig. 2. Structure of Trans-shrink-Net neural network

Download full size

Fig. 3. Accuracy, loss and F1-score curves for the training and validation process of the Trans-shrink-Net

Download full size

Fig. 4. Comparison of ROC and AUC of each modelunder standard data set and comparison of model validation accuracy under different input data conditions. (a) Comparison of ROC and AUC of Trans-shrink-Net, Transformer_Net, LSTM, InceptionNet and ResNet-30; (b) comparison of model validation accuracy under different data conditions

Download full size

Table 1. Millimeter wave radar parameter table
View table
Table 1. Millimeter wave radar parameter table
Parameter Value
Output power / dBm 12
Center carrier frequency /GHz 77
Bandwidth /GHz 4
Modulation type Sawtooth modulation
Bmatrix stop 21
Smatrix stop 13

Table 2. Parameter setting of network training, testing and verification process

View table

Table 2. Parameter setting of network training, testing and verification process

Parameter	Value	Parameter	Value
Batchsize	512	Layer normalization	ε=10^-6
Batch	33	Regularization	L4 population
Epochs	35	Training layer by layer	No
Learning rate	0.00015	Strategy of curing model selection	Best epoch
Optimizer	Adam （weight decay is 10^-4）	Number of layers	3
Training/validation frequency （number of batches）	After epoch	Hidden size	18
Input data shape	137×18×6	Number of attention heads	6

Table 3. Comparison of accuracy and AUC accuracy of the model validation set

View table

Table 3. Comparison of accuracy and AUC accuracy of the model validation set

Model	Accuracy /%			AUC
Model	Original data	CA-CFAR processed data	Mixed data	AUC
Trans-shrink-Net	88.19	96.00	90.09	0.96
Transformer_Net	82.09	91.03	89.91	0.89
LSTM	50.98	69.24	67.20	0.64
InceptionNet	68.09	76.02	75.20	0.77
ResNet-30	70.29	84.69	81.55	0.82

Table 4. Comparison of accuracy of the described model on different AS validation sets

View table

Table 4. Comparison of accuracy of the described model on different AS validation sets

Model	Body wash	Facial cleanser	Laundry detergent	Colonia	Beverage
Trans-shrink-Net	92.23	92.09	97.33	95.75	91.17
Transformer_Net	90.79	90.03	93.55	90.89	90.89
LSTM	66.98	62.24	67.28	60.64	66.76
InceptionNet	77.09	74.72	72.22	70.77	78.45
ResNet-30	88.29	87.69	81.55	77.82	89.12

Table 5. Trans-shrink-Net key component ablation validation

View table

Table 5. Trans-shrink-Net key component ablation validation

Model	Val_acc	Test_acc	Difference value
Remove custom convolution block	91.05	80.09	10.96
Remove the shrink threshold residuals block	87.49	80.91	6.58
Two Transformer blocks	92.44	83.62	8.82
Four Transformer blocks	90.35	84.92	5.43
Learned positional encoding	86.11	76.10	10.01

Tools

Get Citation

Copy Citation Text

Yan Li, Liang Li, Chenyu Zhao, Yulu Zhang, Yun He, Pei Liang. Weak Life Signal Recognition Based on Deep Neural Network[J]. Acta Optica Sinica, 2024, 44(21): 2117001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Medical optics and biotechnology

Received: Apr. 24, 2024

Accepted: Jun. 11, 2024

Published Online: Nov. 19, 2024

The Author Email: Zhang Yulu (zhangyulu@nuist.edu.cn), He Yun (heyun@whut.edu.cn), Liang Pei (plianghust@126.com)

DOI:10.3788/AOS240906

CSTR:32393.14.AOS240906

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Millimeter wave radar parameter table

Table 1. Millimeter wave radar parameter table

Table 2. Parameter setting of network training, testing and verification process

Table 2. Parameter setting of network training, testing and verification process

Table 3. Comparison of accuracy and AUC accuracy of the model validation set

Table 3. Comparison of accuracy and AUC accuracy of the model validation set

Table 4. Comparison of accuracy of the described model on different AS validation sets

Table 4. Comparison of accuracy of the described model on different AS validation sets

Table 5. Trans-shrink-Net key component ablation validation

Table 5. Trans-shrink-Net key component ablation validation