๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

๐Ÿ”ฌ Science/๐Ÿ“ป Signal

๋ ˆ์ด๋” ์‹ ํ˜ธ ์ธ์‹์„ ์œ„ํ•œ CNN ์„ค๊ณ„

๋ถ„์„ ๋…ผ๋ฌธ : Xuezhong Wang(2021), "Electronic radar signal recognition based on wavelet transform and convolution neural network", Alexandria Engineering Journal 61. 

 

CNN์„ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ๋ณธ์ ์ธ ๊ฐœ๋…

 

Layer

- 2๊ฐœ์˜ ์ธ์ ‘ ๋…ธ๋“œ๋Š” Convolution Layer์™€ Pooling Layer๋กœ ๋ฒˆ๊ฐˆ์•„ ๊ตฌ์„ฑ๋œ๋‹ค. ๋งˆ์ง€๋ง‰์€ Fully Connection Layer

 

Convolution Layer

- feature(ํŠน์ง•. ํ”ผ์ฒ˜)๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ณ„์ธต์œผ๋กœ, upper layer์˜ input์— ๋Œ€ํ•œ sliding window ํ•ฉ์„ฑ๊ณฑ์„ ํ•จ์œผ๋กœ์„œ,

๊ฐ convolution kernel์ด local feature์„ ์ถ”์ถœํ•˜๋„๋ก ๋งŒ๋“ ๋‹ค. 

- Convolution ๊ธฐ๋Šฅ์€ ๋‰ด๋Ÿฐ ๊ตฌ์กฐ์˜ ์ „๋‹ฌ ํ•จ์ˆ˜์™€ ์œ ์‚ฌํ•˜๊ณ , ๋‹ค๋ฅธ convolution window๊ฐ€ ๋‹ค๋ฅธ local feature ์ถ”์ถœ ๊ฐ€๋Šฅ

 

Pooling Layer

- feature์„ mappingํ•˜๋Š” ๋ ˆ์ด์–ด๋กœ, Convolution layer์—์„œ ์ถ”์ถœํ•œ ํŠน์ง•์˜ ํฌ๊ธฐ๋ฅผ ์ค„์ด๋Š” ๊ธฐ๋Šฅ์„ ํ•œ๋‹ค.

์œ„์—์„œ ์ถ”์ถœํ•œ data features์„ classifier๋กœ ๋ฐ”๋กœ ๋ณด๋‚ด ํ•™์Šต์‹œํ‚ค๋ฉด, feature์˜ dimension์ด ๋„ˆ๋ฌด ๋†’์•„์ง. ๋ณต์žกํ•ด์ง„๋‹ค๋Š” ๋œป

์ด ๊ฒฝ์šฐ ์ค‘๋ณต ์ •๋ณด๋„ ๋งŽ์ด ํฌํ•จ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด pooling ๊ณ„์ธต์ด ํ•„์š”.

 

- ์ข…๋ฅ˜๋Š” ๋‘ ๊ฐ€์ง€, Max & Average

Max pooling์€ maximum aggregation statistics์ด ์‚ฌ์šฉ๋˜๋ฉฐ, Average๋Š” mean aggregation statistics ์‚ฌ์šฉ.

Mean Aggregation Statistics์€ ํŠน์ง•์ ์˜ ํ‰๊ท ๊ฐ’์„ ๊ตฌํ•˜๋Š” ๊ฑด๋ฐ, ์ฃผ์š” ๊ฐ€์ค‘์น˜๋ฅผ ๊ฐ–๋Š” value์˜ ํŠน์„ฑ์ด ํฌ๋ฏธํ•ด์งˆ ์ˆ˜ ์žˆ๋Š” ๋ฌธ์ œ๋กœ ์ธํ•ด Max pooling์ด ๋งŽ์ด ์‚ฌ์šฉ๋จ. 

 

Fully Connection Layer

- ์•ž ๋„คํŠธ์›Œํฌ ๊ณ„์ธต์—์„œ ์–ป์€ ํŠน์ง•์„ ๋ถ„๋ฅ˜ํ•˜๊ณ , ์ธ์‹ํ•˜๋Š” ๊ธฐ๋Šฅ์„ ํ•œ๋‹ค.

 

 

Activation Function

- ๋น„์„ ํ˜• ๋ชจ๋ธ๋ง ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•˜๋Š” ํ•จ์ˆ˜. ๋‹จ์ˆœํžˆ ์œ„ layer๋“ค๋กœ ๊ตฌ์„ฑํ•œ CNN์€ ์„ ํ˜• ๋งคํ•‘๋งŒ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ,

์‹ค์ œ ํ™˜๊ฒฝ์—์„œ๋Š” ๋น„์„ ํ˜• ๋ถ„ํฌ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋ธ๋ง ํ•ด์•ผํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ํ›จ์”ฌ ๋งŽ์Œ.

(์ผ๋ฐ˜์ ์ธ ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋“ค : Sigmoid function, Tanh function, ReLu function )

 

ReLu function

- Sigmoid, Tanh ํ•จ์ˆ˜์™€ ๋‹ฌ๋ฆฌ, ์„ ํ˜• ๋ถˆํฌํ™”, ๋น ๋ฅธ ์ˆ˜๋ ด ์†๋„, ๊ฐ„๋‹จํ•œ ๊ตฌํ˜„ ๋ฐ ๋ณต์žกํ•œ ๊ณ„์‚ฐ์ด ํ•„์š” ์—†์Œ.

 

 

SoftMax

- SoftMax ํ•จ์ˆ˜๋Š” ์ตœ์ข… ouput layer์— ์‚ฌ์šฉ๋˜๋Š”๋ฐ, ์ด๋Š” ์ž…๋ ฅ๋ฐ›์€ ๊ฐ’์„ ๋ชจ๋‘ 0~1 ์‚ฌ์ด์˜ ๊ฐ’์œผ๋กœ ์ •๊ทœํ™”ํ•œ๋‹ค.

์ด๋ฅผ ํ†ตํ•ด ์ด๋ฏธ์ง€๊ฐ€ ๊ฐ label์— ์†ํ•  ํ™•๋ฅ ๊ฐ’์ด label ๋งˆ๋‹ค ๊ฐ๊ฐ ์ถœ๋ ฅ๋˜๊ณ , ์ด ์ค‘ ๊ฐ€์žฅ ๋†’์€ ํ™•๋ฅ ๊ฐ’์„ ๊ฐ€์ง€๋Š” label์ด ์ตœ์ข… ์˜ˆ์ธก์น˜๋กœ ์„ ์ •๋จ.

 

 

๊ตฌ์กฐ ์ˆœ์„œ Convolution Layer - ReLu Function (Activation) - Pooling Layer - Fully Connected Layer - SoftMax


๋ ˆ์ด๋” ์‹ ํ˜ธ ์ธ์‹์„ ์œ„ํ•œ CNN ์„ค๊ณ„

 

1. CNN scale ๊ฒฐ์ •

- Radar emitter ์‹ ํ˜ธ๋Š” ํ˜„์‹ค์—์„œ ๊ต‰์žฅํžˆ ์ž‘๊ธฐ ๋•Œ๋ฌธ์— network over fitting ํ˜„์ƒ์ด ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ์–ด, ๋ณต์žกํ•œ CNN์€ X

 

2. CNN ๊ณ„์ธต ์„ค๊ณ„ (3 convolution, 2 pooling, 2 fully connected layers)

- radar radiation source signal data size = 32*32

- convoluation neural network๋Š” ์‹ ํ˜ธ์˜ ๊ตฌ๋ณ„ ํŠน์ง•์„ ์ถ”์ถœํ•˜์—ฌ, ์ธ์‹๋ฅ ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ. 

 

๊ฐ ๊ณ„์ธต์˜ ์„ธํŒ…

Convolution Layer

Kernel Size๋Š” ํ™€์ˆ˜์—ฌ์•ผํ•จ. ์ดˆ๊ธฐ ํŠน์ง• ์ถ”์ถœ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋” ์ž์„ธํ•œ ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ณ , ์ปค๋„ ํฌ๊ธฐ 5*5๋กœ ์„ค์ •

(์ปค๋„ ์ˆ˜๊ฐ€ ๋„ˆ๋ฌด ์ž‘์œผ๋ฉด ์ผ๋ถ€ ํŠน์ง•์„ ํ•™์Šตํ•  ์ˆ˜ ์—†์–ด ๊ณผ์†Œ ์ ํ•ฉ์ด ๋ฐœ์ƒ. ๋„ˆ๋ฌด ๋งŽ์œผ๋ฉด ์ƒ˜ํ”Œ์ด ํฌ๋ฐ•ํ•ด์ ธ ๊ณผ๋Œ€ ์ ํ•ฉ ๋ฐœ์ƒ)

์ปจ๋ณผ๋ฅ˜์…˜ layer์ด ๋’ค๋กœ ๊ฐˆ ์ˆ˜๋ก ์ปค๋„ ์ˆ˜๋Š” ๋งŽ์•„์ง. So, 3๊ฐœ layer์˜ core ์ˆ˜๋Š” ๊ฐ๊ฐ 32, 64, 64๋กœ ์„ค์ •

 

Activation Function

ReLu ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉ. 

 

Pooling Layer

์ฝ”์–ด์˜ ํฌ๊ธฐ๊ฐ€ ๋„ˆ๋ฌด ํฌ๋ฉด ๋ฐ์ดํ„ฐ ์ •๋ณด๊ฐ€ ์†์‹ค๋˜์–ด ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋จ. ๋”ฐ๋ผ์„œ, ํ’€๋ง๋œ ์ฝ”์–ด์˜ ํฌ๊ธฐ 2*2๋กœ ์„ค์ •

 

Number of Network Parameter 

๊ฐ ๋ ˆ์ด๋” ์‹ ํ˜ธ ์ƒ˜ํ”Œ์˜ ํŠน์„ฑ์ด 32*32๋ณด๋‹ค ์ ๊ณ , cnn ๊ตฌ์กฐ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ๋„ˆ๋ฌด ๋งŽ์œผ๋ฉด ๊ณผ์ ํ•ฉ ๋ฌธ์ œ ๋ฐœ์ƒํ•จ.

๋„คํŠธ์›Œํฌ parameter ์ˆ˜๋Š” ์ „์ฒด ์—ฐ๊ฒฐ ๊ณ„์ธต์˜ ๋‰ด๋Ÿฐ ์ˆ˜์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง. 128๋กœ ์„ค์ •. 

 

3. CNN ํ›ˆ๋ จ ๊ณผ์ •

[1] ๋‚ฎ์€ level์—์„œ ๋†’์€ level๋กœ ์ƒ˜ํ”Œ ๋ฐ์ดํ„ฐ ์ „ํŒŒ (์ˆœ๋ฐฉํ–ฅ)

[2] ์˜ˆ์ƒ๊ฐ’ & ๊ฒฐ๊ณผ๊ฐ’ ์˜ค์ฐจ ๋†’์€ level to ๋‚ฎ์€ level๋กœ ์ „ํŒŒ (์—ญ์ „ํŒŒ)

 

a. ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์ดˆ๊ธฐํ™”๋จ. ๊ฐ€์ค‘์น˜ ๋ฌด์ž‘์œ„๋กœ ์ดˆ๊ธฐํ™”

b. ์ˆœ๋ฐฉํ–ฅ ์ „ํŒŒ : ์ž…๋ ฅ ์ƒ˜ํ”Œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๊ฐ neuron layer์˜ ์ถœ๋ ฅ๊ฐ’์ด ์ž…๋ ฅ layer to ์ถœ๋ ฅ layer๋กœ ๊ณ„์‚ฐ๋จ

c. ์—ญ์ „ํŒŒ : ๋„คํŠธ์›Œํฌ ์ถœ๋ ฅ ๊ฐ’๊ณผ ๋ชฉํ‘œ ๊ฐ’ ์‚ฌ์ด์˜ ์˜ค์ฐจ๋ฅผ ๊ณ„์‚ฐ.

์˜ค์ฐจ๊ฐ€ ์˜ˆ์ƒ์น˜๋ณด๋‹ค ํฌ๋ฉด, ์ฒซ๋ฒˆ์งธ hidden layer๋กœ ์žฌ์ „๋‹ฌ. ์ž‘์œผ๋ฉด ํ•™์Šต ์ข…๋ฃŒ

d. ๊ณ„์‚ฐ๋œ ์˜ค์ฐจ์— ๋”ฐ๋ผ ๋„คํŠธ์›Œํฌ ๊ฐ€์ค‘์น˜๋ฅผ ์—…๋Žƒํ•œ ํ›„ ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„๋กœ ๋Œ์•„๊ฐ. 

 

4. ์‹ ํ˜ธ ์ธ์‹์˜ ์‹คํ—˜์  ๋ถ„์„

์ƒ์„ฑ ๋ ˆ์ด๋” ์‹ ํ˜ธ์˜ ๋ณ€์กฐ ๋ชจ๋“œ CW, LFM, BPSK, bfsk4 ์œ ํ˜•

๋ฐ˜์†กํŒŒ ์ฃผํŒŒ์ˆ˜ ๋ฒ”์œ„๋Š” <= 1GHz์— ์ง‘์ค‘. 

 

์ธก์ •๋œ ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ - 10๊ฐ€์ง€ ์œ ํ˜•์˜ ๋ ˆ์ด๋” ๊ฐœ์ฒด๊ฐ€ ํฌํ•จ๋จ.

๊ฐ ์œ ํ˜•์˜ ํŽ„์Šค ์ˆ˜ {80, 63, 69, 59, 69, 72, 54, 45, 71, 83}์œผ๋กœ 665๊ฐœ์˜ ํŽ„์Šค ์กด์žฌ.

ํŽ„์Šค ์‹ ํ˜ธ๋Š” ๋‹จ์ผ ์ฃผํŒŒ์ˆ˜ ์‹ ํ˜ธ์ด๋ฉฐ, ํŽ„์Šค ํญ์€ ์•ฝ 400๊ฐœ์˜ ์ƒ˜ํ”Œ๋ง ํฌ์ธํŠธ.  SNR ์•ฝ 10db ~ 20db

ํŽ„์Šค ์‹ ํ˜ธ์—๋Š” ๋‹ค์ค‘ ๊ฒฝ๋กœ ๊ฐ„์„ญ์ด ์—†๊ณ , ํŽ„์Šค ์•ž๋’ค ๊ฐ€์žฅ์ž๋ฆฌ๊ฐ€ ๊ฐ€ํŒŒ๋ฅด๋ฉฐ ํŽ„์Šค ํญ์ด ๋™์ผ

 

์ „์ž ๋ ˆ์ด๋”์˜ ์ธก์ •๋œ ์‹ ํ˜ธ ๊ทธ๋ฃน์„ ์ถ”์ถœํ•˜์—ฌ ๋ถ„์„ ๋ฐ ์ธ์‹ ๊ณผ์ •์„ ์‹œ๊ฐํ™”ํ•œ๋‹ค. 

Wavelet Transform : ์ธก์ •๋œ ๋ฐ์ดํ„ฐ์˜ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ & ํŠน์ง• ์ถ”์ถœ์— ์‚ฌ์šฉ

Signal Envelope & ์ฃผํŒŒ์ˆ˜ ๋„๋ฉ”์ธ Features : ์‹œ๊ฐํ™”์— ์‚ฌ์šฉ

 

*Accurate carrier frequency estimation of emitter signal - ์ฃผํŒŒ์ˆ˜ domain์—์„œ ๊ฐœ๋ณ„ ํ”ผ์ฒ˜๋ฅผ ์ถ”์ถœํ•˜๋Š” ๋˜ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•

 

 

 

์ฐธ๊ณ  

https://www.sciencedirect.com/science/article/pii/S1110016821006049

 

๋ฐ˜์‘ํ˜•