10 Prediction of Residual Life of Oil and Gas Pipeline Corrosion …
109
reconstructed by the hidden layer unit, and the new visible layer unit is mapped to
the hidden layer unit again to obtain a new hidden layer [6]. The unit, this step of
reversing and retreating is called Gibbs sampling. Since the difference in probability
distribution between the visual layer input and the hidden layer activation unit is the
main basis for weight update, the contrast divergence algorithm is used to pre-train
to obtain the weight of the generated model, and the training time is significantly
reduced, requiring only a small amount. It can converge in a few steps.
For the top two layers, the output of the lower layer provides relevant reference
clues to the top layer. The top layer links the output to its memory content based on
these clues, and finally completes the discriminant classification task. The performance of the network after this processing is better than that of the BP algorithm
alone, because the DBN BP algorithm only needs to perform a local search on the
weight parameter space, and the training and convergence time is less.
RBM is like a brick of architecture. It is easy to learn the weight of connection, which makes DBN have flexible expansion performance. CDBN is one of
the convolutional deep confidence networks [8]. CD-BN takes into account the
two-dimensional structure information of the image, and uses the spatial domain
relationship of the neighboring pixels to achieve the transformation invariance of
the generated model by convolving the RBM model, and can easily transform the
high-dimensional image.
10.3 Deep Life Prediction of Deep Pipeline Network
Pipeline
In this paper, the deep confidence network prediction method is used to realize the
remaining life prediction of pipelines. The feasibility is as follows:
(1) When using the acoustic emission detection method, the corroded gas pipeline
has a special elastic wave [9]. When the detection background is constant and the
detection position is fixed, the non-defective material has a uniform elastic wave
frequency, and the defective frequency is only partially grayscale. Differences,
these characteristics apply to the strong feature extraction capabilities of DBN.
(2) DBN combines elastic wave frequency processing and feature extraction into
one, which greatly shortens the defect detection time and can better meet the
online product detection. It also provides a theoretical basis for parallelization.
(3) DBN can quickly complete the weight training so that it can meet the production
needs when the product background, location and model change.
(4) DBN combined with traditional identification network (usually adding a recognition layer at the top layer), not only can achieve signal processing, feature
extraction [10], but also pattern recognition, providing data for further predicting
pipeline corrosion rate and further predicting pipeline remaining life Processing
route.
109
reconstructed by the hidden layer unit, and the new visible layer unit is mapped to
the hidden layer unit again to obtain a new hidden layer [6]. The unit, this step of
reversing and retreating is called Gibbs sampling. Since the difference in probability
distribution between the visual layer input and the hidden layer activation unit is the
main basis for weight update, the contrast divergence algorithm is used to pre-train
to obtain the weight of the generated model, and the training time is significantly
reduced, requiring only a small amount. It can converge in a few steps.
For the top two layers, the output of the lower layer provides relevant reference
clues to the top layer. The top layer links the output to its memory content based on
these clues, and finally completes the discriminant classification task. The performance of the network after this processing is better than that of the BP algorithm
alone, because the DBN BP algorithm only needs to perform a local search on the
weight parameter space, and the training and convergence time is less.
RBM is like a brick of architecture. It is easy to learn the weight of connection, which makes DBN have flexible expansion performance. CDBN is one of
the convolutional deep confidence networks [8]. CD-BN takes into account the
two-dimensional structure information of the image, and uses the spatial domain
relationship of the neighboring pixels to achieve the transformation invariance of
the generated model by convolving the RBM model, and can easily transform the
high-dimensional image.
10.3 Deep Life Prediction of Deep Pipeline Network
Pipeline
In this paper, the deep confidence network prediction method is used to realize the
remaining life prediction of pipelines. The feasibility is as follows:
(1) When using the acoustic emission detection method, the corroded gas pipeline
has a special elastic wave [9]. When the detection background is constant and the
detection position is fixed, the non-defective material has a uniform elastic wave
frequency, and the defective frequency is only partially grayscale. Differences,
these characteristics apply to the strong feature extraction capabilities of DBN.
(2) DBN combines elastic wave frequency processing and feature extraction into
one, which greatly shortens the defect detection time and can better meet the
online product detection. It also provides a theoretical basis for parallelization.
(3) DBN can quickly complete the weight training so that it can meet the production
needs when the product background, location and model change.
(4) DBN combined with traditional identification network (usually adding a recognition layer at the top layer), not only can achieve signal processing, feature
extraction [10], but also pattern recognition, providing data for further predicting
pipeline corrosion rate and further predicting pipeline remaining life Processing
route.