Interpretable Feature Selection Method for Optical-Fiber Disturbance Signal Recognition

Min Sun; Nian Fang

doi:10.3788/AOS241101

Acta Optica Sinica, Volume. 44, Issue 21, 2106007(2024)

Interpretable Feature Selection Method for Optical-Fiber Disturbance Signal Recognition

Min Sun and Nian Fang^*

School of Communication and Information Engineering, Key Laboratory of Specialty Fiber Optics and Optical Access Networks, Joint International Research Laboratory of Specialty Fiber Optics and Advanced Communication, Shanghai University, Shanghai 200444, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Flow chart of signal recognition with explainable feature selection method

Download full size

Fig. 2. Structure of φ-OTDR system for dataset acquisition^[25]

Download full size

Fig. 3. Time-domain signals of six types of events. (a) Background noise; (b) digging; (c) knocking; (d) watering; (e) shaking; (f) walking

Download full size

Fig. 4. Confusion matrices of four models without feature selection. (a) SVM; (b) KNN; (c) DT; (d) RF

Download full size

Fig. 5. Influences of different features on model prediction results

Download full size

Fig. 6. Feature importance rank of four models based on SHAP. (a) SVM; (b) KNN; (c) DT; (d) RF

Download full size

Fig. 7. Confusion matrices of four models after feature selection. (a) SVM; (b) KNN; (c) DT; (d) RF

Download full size

Table 1. Details of dataset
View table
Table 1. Details of dataset
Event type Number of training samples Number of test samples Label
Background noise 2357 589 0
Digging 2010 502 1
Knocking 2024 506 2
Watering 1802 451 3
Shaking 2182 546 4
Walking 1960 490 5

Table 2. Definitions of time-domain features

View table

Table 2. Definitions of time-domain features

Feature	Definition	Feature	Definition
Peak value	$X_{P K} = m a x (x)$	Standard deviation	$X_{S t d} = \sqrt[]{X_{V a r}}$
Minimum	$X_{M i n} = m i n (x)$	Energy	$X_{E} = x_{i}^{2}$
Peak to peak value	$X_{P K ‑ P K} = X_{P K} - X_{M i n}$	Kurtosis	$X_{K u r} = \frac{1}{n} \frac{\sum_{i = 1}^{n} {(x_{i} - \bar{x})}^{4}}{X_{R M S}^{4}}$
Mean	$X_{M e a n} = \frac{1}{n} \sum_{i = 1}^{n} x_{i}$	Skewness	$X_{S k e w} = \frac{1}{n} {\sum_{i = 1}^{n} (\frac{x_{i} - \bar{x}}{X_{s t d}})}^{3}$
Rectified mean value	$X_{R M V} = \frac{1}{n} \sum_{i = 1}^{n} \| x_{i} \|$	Margin factor	$X_{M F} = \frac{X_{P K}}{X_{R S A}}$
Root mean square	$X_{R M S} = \sqrt[]{\frac{1}{n} \sum_{i = 1}^{n} x_{i}^{2}}$	Peak factor	$X_{P F} = \frac{X_{P K}}{X_{R M S}}$
Root square amplitude	$X_{R S A} = {(\frac{1}{n} \sqrt[]{\sum_{i = 1}^{n} \| x_{i} \|})}^{2}$	Impulse factor	$X_{I F} = \frac{X_{P K}}{X_{R M V}}$
Variance	$X_{V a r} = \frac{1}{n} \sum_{i = 1}^{n} {(x_{i} - \bar{x})}^{2}$	Shape factor	$X_{S F} = \frac{X_{R M S}}{X_{R M V}}$

Table 3. Definitions of frequency-domain features

View table

Table 3. Definitions of frequency-domain features

Feature	Definition	Feature	Definition
LLP energy	$X_{L L P E} = \sum_{k = 1}^{N} {\|D_{1} (k)\|}^{2}$	LHP energy	$X_{L H P E} = \sum_{k = 1}^{N} {\|D_{2} (k)\|}^{2}$
HLP energy	$X_{H L P E} = \sum_{k = 1}^{N} {\|D_{3} (k)\|}^{2}$	HHP energy	$X_{H H P E} = \sum_{k = 1}^{N} {\|D_{4} (k)\|}^{2}$
Wavelet entropy	$X_{W E} = - \sum_{i = 1}^{4} p (i) \cdot l o g_{2} p (i)$	Wavelet information quantum	$X_{W I Q} = \sum_{i = 1}^{4} p (i) \cdot E (i)$

Table 4. Recognition results of four models without feature selection

View table

Table 4. Recognition results of four models without feature selection

Model	Label	P /%	R /%	F₁ /%	A /%
SVM	1	70.38	76.69	73.40	81.68
	2	86.67	87.35	87.01
	4	86.34	93.77	89.90
KNN	1	87.72	91.04	89.35	94.29
	2	94.18	99.21	96.63
	4	97.79	97.07	97.43
DT	1	89.24	79.28	83.97	85.31
	2	91.55	92.09	91.82
	4	93.74	93.22	93.48
RF	1	96.16	89.84	92.89	96.11
	2	97.46	98.62	98.04
	4	97.61	97.44	97.53

Table 5. Recognition results of four models after feature selection

View table

Table 5. Recognition results of four models after feature selection

Model	Label	P /%	R /%	F₁ /%	A /%
SVM	1	72.28	82.07	76.87	83.07
	2	86.06	90.32	88.14
	4	86.82	94.14	90.33
KNN	1	88.34	92.03	90.15	95.04
	2	95.81	99.41	97.58
	4	98.17	98.35	98.26
DT	1	89.14	80.08	84.37	90.01
	2	92.46	92.09	92.28
	4	95.41	91.39	93.36
RF	1	96.61	90.84	93.63	96.5
	2	96.71	98.62	97.65
	4	98.35	97.99	98.17

Table 6. Comparison of recognition time

View table

Table 6. Comparison of recognition time

Model	Feature extraction time /s		Model recognition time /ms		Average time /ms
Model	Original	Proposed	Original	Proposed	Original	Proposed
SVM	252.35	191.46	6920.00	2670.00	84.07	62.95
KNN	252.35	182.17	5130.00	1092.00	83.00	59.69
DT	252.35	217.64	6.05	4.98	81.83	70.57
RF	252.35	203.58	42.11	37.90	81.82	66.01

Table 7. Comparison of average recognition accuracy A with different feature selection methods
View table
Table 7. Comparison of average recognition accuracy A with different feature selection methods
Model A /%
Proposed method Fisher score^［21］ Mutual information^［23］
SVM 83.07 82.74 81.82
KNN 95.04 94.63 94.38
DT 90.01 88.34 87.47
RF 96.50 96.26 96.15

Tools

Get Citation

Copy Citation Text

Min Sun, Nian Fang. Interpretable Feature Selection Method for Optical-Fiber Disturbance Signal Recognition[J]. Acta Optica Sinica, 2024, 44(21): 2106007

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites