100
B. P. Duong et al.
for leak detection in different leak sizes, the pipeline has five types of recognition
states 0.3 mm, 0.5 mm, 1 mm, 2 mm, and normal. As a result, we need to conduct
five SVMs to respect to the five kinds of states. For the i
th SVM, the recognition
output value of i
th pipeline state should be (+1) while the output values of other
states should be all (−1). The classification principle of SVM for distinguishing
multi-classes is based on the kernel method and the structural risk minimization
principle.
f (z) = sgn
n
i=1
λ i y i ψ(z i , z) + b
(4)
where z i and z describe training and measured feature samples, respectively;
y = {−1, +1} is the target label. λ represents the Lagrange coefficient, which
satisfies the condition that 0 ≤ λ i ≤ C,
n
i=1 λ i y i = 0 with i = 1, 2, ..., n. The
parameter C is the penalty factor, b represents the threshold value for classification. ψ(z i , z) is the kernel function. Depending on whether specific engineering
problems of leakage features in the pipeline, the Gaussian radial basis kernel (RBF)
ψ(z i , z) = exp
−γ z i − z
2
is considered as the kernel function. To solve the issue
of parameter selection, the hyperparameters (C, γ ) of the SVMs, which can directly
affect classification accuracy, can be obtained by using the grid search method to
obtain the maximum accuracy.
9.4 Pipeline Fault Experimental Analyzed Results
Applying the proposed method to the collected dataset from the experimental testbed,
which is introduced in Sect. 9.2. Considering different leak sizes, including 0.3 mm
orifice, 0.5 mm orifice, 1 mm orifice, 2 mm orifice, each type of fault signal is put
into the burst detector to detect and segment to the list of burst segment signals.
Hence, for each burst segment, the features are extracted and input to the OAAMCSVMs. The classification result for five classes, four sizes of the leak and the
normal state, is illustrated in Fig. 9.4a with the confusion matrix describes the result
of the OAA-MCSVMs classifier of the ECFAR. As depicted in Fig. 9.4a, the average
accuracy of the MCSVMs with the proposed method is equal to 93%. For illustration, the research also compares the result of the ECFAR burst detection with
the wavelet-based threshold burst detection. The wavelet-based threshold is another
methodology used for burst detection as described in [8]. Observing that wavelet
spectrums of noise signals and leak signals have different performance characteristics
in each scale, signals are decomposed into both wavelet coefficients and the scaling
values through the wavelet decomposition technique. The complete decomposition
hierarchy is provided based on this technique. Consequently, because of uniform
frequency secondary groups, the wavelet decomposition is extremely adaptable. The
first some decomposition levels, extremes of the details are both due to noise and
B. P. Duong et al.
for leak detection in different leak sizes, the pipeline has five types of recognition
states 0.3 mm, 0.5 mm, 1 mm, 2 mm, and normal. As a result, we need to conduct
five SVMs to respect to the five kinds of states. For the i
th SVM, the recognition
output value of i
th pipeline state should be (+1) while the output values of other
states should be all (−1). The classification principle of SVM for distinguishing
multi-classes is based on the kernel method and the structural risk minimization
principle.
f (z) = sgn
n
i=1
λ i y i ψ(z i , z) + b
(4)
where z i and z describe training and measured feature samples, respectively;
y = {−1, +1} is the target label. λ represents the Lagrange coefficient, which
satisfies the condition that 0 ≤ λ i ≤ C,
n
i=1 λ i y i = 0 with i = 1, 2, ..., n. The
parameter C is the penalty factor, b represents the threshold value for classification. ψ(z i , z) is the kernel function. Depending on whether specific engineering
problems of leakage features in the pipeline, the Gaussian radial basis kernel (RBF)
ψ(z i , z) = exp
−γ z i − z
2
is considered as the kernel function. To solve the issue
of parameter selection, the hyperparameters (C, γ ) of the SVMs, which can directly
affect classification accuracy, can be obtained by using the grid search method to
obtain the maximum accuracy.
9.4 Pipeline Fault Experimental Analyzed Results
Applying the proposed method to the collected dataset from the experimental testbed,
which is introduced in Sect. 9.2. Considering different leak sizes, including 0.3 mm
orifice, 0.5 mm orifice, 1 mm orifice, 2 mm orifice, each type of fault signal is put
into the burst detector to detect and segment to the list of burst segment signals.
Hence, for each burst segment, the features are extracted and input to the OAAMCSVMs. The classification result for five classes, four sizes of the leak and the
normal state, is illustrated in Fig. 9.4a with the confusion matrix describes the result
of the OAA-MCSVMs classifier of the ECFAR. As depicted in Fig. 9.4a, the average
accuracy of the MCSVMs with the proposed method is equal to 93%. For illustration, the research also compares the result of the ECFAR burst detection with
the wavelet-based threshold burst detection. The wavelet-based threshold is another
methodology used for burst detection as described in [8]. Observing that wavelet
spectrums of noise signals and leak signals have different performance characteristics
in each scale, signals are decomposed into both wavelet coefficients and the scaling
values through the wavelet decomposition technique. The complete decomposition
hierarchy is provided based on this technique. Consequently, because of uniform
frequency secondary groups, the wavelet decomposition is extremely adaptable. The
first some decomposition levels, extremes of the details are both due to noise and