Multiple microphone sensor nodes are used to simultaneously collect time-series data of the leakage signals. â¢. Construct the RMS volume feature using four ...
DEEP-LEARNING-BASED PIPE LEAK DETECTION USING IMAGE-BASED LEAK FEATURES Ji-Hoon Bae1, Doyeob Yeo1, Doo-Byung Yoon2, Se Won OH1, Gwan Joong Kim1, Nae-Soo Kim1, and Cheol-Sig Pyo1, ETRI1, KAERI2, Rep. of Korea
INTRODUCTION Frequency domain H: 2.0 mm, P: 4 bar
t [sec]
Absolute magnitude
Amplitude level
H: 1.0 mm, P: 2 bar
Amplitude level
Time domain
t [sec]
H: 1.0 mm, P: 2 bar
Absolute magnitude
• Recently, in the field of plant industry, corrosion and leakage of pipes have been observed because of aging pipelines installed at the time of initial construction social problems, economic loss, and human injuries. • We consider acoustic-based leak detection technology (high leak-detection speed and ease of retrofitting). • Characteristics of leakage signals: 1) the magnitude of the measured leakage signal tends to increase with the increase in leakage scales (time-domain) and 2) the magnitude of the frequency spectrum also tends to increase as the leakage scales increase (frequency-domain). • Main objective: implement deep-learning-based pipe leak detection (PLD) using trajectorybased image features that reflect the aforementioned characteristics of the leakage signal.
H: 2.0 mm, P: 4 bar
f [kHz]
f [kHz]
PROPOSED IMAGE FEATURE EXTRACTION Generation of RMS-pattern image feature 1) For path-A, directly calculate the RMS level of the sampled input data x using a moving averaging window. 2) For path-B, calculated the RMS level of the preprocessed signal vector using the band-pass filter and amplitude weighing blocks. 3) Concatenate the two RMS-level feature vectors as . 4) Generate a 2D RMS-pattern image feature from 2D domain is discretized by a MN grid (M-levels in the time range and N-levels in the amplitude range). The amplitude values of the are quantized with the same Nlevels with respect to the corresponding M-intervals. Each quantized value is mapped to the corresponding lattice in the 2D MN domain. Small-scale Leak signal
Small-scale noise Large-scale Leak signal
Generation of frequency-pattern image feature 1) Perform a band-pass filtering for the sampled incoming data x. 2) Convert the filtered time-series acoustic data into frequency-domain data using 1D Fourier transform. 3) Generate a 2D frequency-pattern image feature from . 2D domain is discretized by a MN grid (M-levels in the frequency range and N-levels in the amplitude range) The amplitude values of the normalized absolute values of the are quantized with the same N-levels with respect to the corresponding M-intervals. Each quantized value is mapped to the corresponding lattice in the 2D MN domain Small-scale Leak signal
Large-scale Leak signal
Small-scale noise
Large-scale noise
Large-scale noise
DEEP-LEARNING-BASED PIPE LEAK DETECTION (PLD) Laboratory facility environment for PLD
Four 20KHz-microphone sensors. Sampling rate of 150 KHz for 0.5 [s]. 100 test positions of leak test region. 25 combinations of test conditions Hi and Pj (i = 1,2,…,5, and j = 1,2,…,5). Leak scale Values Leak hole 0.5, 1.0, 1.5, 2.0, 2.5 Diameter [mm] Escaping air 1, 2, 3, 4, 5 Pressure [bar]
• • •
Multiple microphone sensor nodes are used to simultaneously collect time-series data of the leakage signals. Construct the RMS volume feature using four RMS-pattern images extracted from the corresponding four sensor nodes. Construct the Frequency volume feature using four frequencypattern images extracted from the corresponding four sensor nodes.
Ensemble deep-learning structure for PLD • 8-layer residual network5 (ResNet) model is used as the base ensemble deep-learning (EDL) for PLD. • Training dataset ( H i Pj i=1with j=1, 3, and 5): 3,000 40404 volume features (1,500 for leakage signals and 1,500 for noises). 5 • Test dataset ( Hi Pj i=1with j=2 and 4): 2,000 40404 volume features (1,000 for leakage signals and 1,000 for noises). • Batch size of 100, training epochs of 33, and momentum optimizer.
Classification accuracy [%] for PLD Accuracy [%]
Method
Epochs
NN using 1160 RMS feature
77.805
88
NN using 1160 frequency feature
98.415
37
LSTM using 40 sequential length
V 3D 3D 8-layer ResNet using 40404 Vrms 8-layer ResNet using 40404
8-layer ResNet using 40404 V 3D 3D and EDL using Vrms V freq
3D freq
84.53
33
80.755
33
99.875
33
99.57
33
99.97
33