Home
|
About Journal
|
Editorial Board
|
Instruction
|
Subscriptions
|
Contacts Us
|
中文
Office Online
Submission Online
Peer Review
Editor Work
Editor-in-chief
Office Work
Journal Online
Accepted
Current Issue
Advanced Search
Archive
Read Articles
Download Articles
Email Alert
Quick Search
Adv Search
2021 Vol. 42, No. 3
Published: 2021-06-28
221
Machine Learning based prediction methods on shale gas recovery
Shale gas is a kind of unconventional natural gas which exists in mud shale with adsorption, free and fluid phase. The proved reserves in China are abundant and widely distributed, and the buried depth is below 3000 meters. The key technologies of shale gas production are horizontal well and hydraulic fracturing, while the more difficult and challenge for efficient production is to predict recovery. If the recovery can be predicted, we can evaluate the degree of reservoir reconstruction, and know the direct effect of current construction parameters on gas production. In recent years, with the rise of deep learning, it has become a popular method to solve engineering problems. In this paper, three machine learning methods, DNN, SVM and Xgboost, are used to establish the prediction models from reservoir and construction parameters to recovery factor by analyzing data of horizontal wells in shale gas production in the Fuling District area. The advantages and disadvantages of various models are analyzed, and the importance of relevant parameters is also discussed. Under the condition of small data in shale gas construction, the reasonable recovery prediction models are set up, which has the prospect of engineering application.
2021 Vol. 42 (3): 221-232 [
Abstract
] (
367
)
HTML
(1 KB)
PDF
(0 KB) (
157
)
233
Data-driven elastoplastic constitutive law of gradient structure materials
Gradient structure materials are widely used in engineering applications due to their excellent mechanical properties. However, because its composition or microstructure varies gradually with spatial position and the related mechanical properties, simulating engineering structures containing such materials often faces great challenge and further limits the applications. In order to simulate this kind of materials, the present work proposes a new approach to construct elastoplastic constitutive law of gradient structure materials by integrating the plastic theory and machine learning technology. The proposed approach first builds different representative volume elements (RVE) based on the microstructures at different positions to generate homogenized stress-strain data numerically and then train an artificial neural network (ANN) through the generated data. Then the trained ANN can be used as the homogenized elastic-plastic material model for the gradient structure materials, replacing the yield function in conventional plasticity. The obtained homogenized material model is implemented through UMAT interface of ABAQUS for ease of use. Illustrative example shows that the constitutive model constructed by the present approach can be used to solve the boundary value problems involving elastoplastic materials with gradient structure under complex loading paths (e.g., cyclic/reverse loading) effectively. Compared with the results of direct numerical simulation on the gradient microstructure, the accuracy of the proposed approach in computing the macroscopic mechanical response is verified. But the degree of freedom of the engineering structure and the computational cost are greatly reduced. This approach provides a new way for computing the elastoplastic macroscopic mechanical response of structure involving complex gradient structure materials. At present, the approach can only be used for materials with isotropic hardening and meeting the associated flow law. Further research is needed to overcome these issues.
2021 Vol. 42 (3): 233-240 [
Abstract
] (
397
)
HTML
(1 KB)
PDF
(0 KB) (
160
)
241
An efficient hierarchical data searching scheme for data-driven computational mechanics
The most shining advantage of the data-driven computational mechanics is directly embedding the material data into mechanical simulations. Its basic idea is to assign to each integration point the optimal stress-strain data from the database. However, the original data searching process during the data-driven computing is time-consuming in cases with high-dimensional and high-density databases. This work aims to propose a new hierarchical data searching scheme to improve the data-driven computing efficiency. To this end, the original database is pre-processed to be a multi-layer tree data structure firstly, in which the amount of data decreases layer by layer. Then the tree search algorithm is adopted to narrow down the data search and to reduce the time cost for data searching. Within this data searching scheme, a benchmark test is considered to discuss the influences of the number of database layers and different data allocation settings on the computing efficiency. It is found that the computation time decreases significantly with the increase of database layers and allocating data uniformly for each sub-database allows to further improve the computing efficiency.
2021 Vol. 42 (3): 241-248 [
Abstract
] (
270
)
HTML
(1 KB)
PDF
(0 KB) (
157
)
249
The application of deep collocation method and deep energy method with a two-step optimizer in the bending analysis of Kirchhoff thin plate
With the advancement of the computing power and machine learning algorithms, deep learning method has been widely applied in a wide range of fields. In this manuscript, we have developed the deep collocation and energy method fitted to engineering computation and further applied to solve Kirchhoff thin plate bending problems. The deep collocation method, which adopted the physics-informed neural networks, incorporating the strong form governing equations into the loss function, which reduces the solving of thin plate problem into an optimization problem. On the other hand, the deep energy method adopted an energy driven neural networks based on the Principle of Minimum Potential Energy, indicating that of all displacements satisfying given boundary and Equilibrium conditions, the actual displacement is the one that minimizes the total potential energy at stable Equilibrium. Thus, we can build a loss function from the total potential energy. Boundary conditions are penalized to the loss form, which reduced to an unconstrained optimization problem. The physics-informed and energy driven neural networks are based on the universal approximation theorem. Due to the introduction of physical and energy information, the neural networks become difficult to train, an improved two-step optimization algorithm is presented to train the neural network. From numerical results, it is clearly the both models are suitable to solve thin plate bending problems and contributes to they are easy in implementation and is truly “Meshfree”.
2021 Vol. 42 (3): 249-266 [
Abstract
] (
303
)
HTML
(1 KB)
PDF
(0 KB) (
148
)
267
A new method for fault diagnosis of rolling bearing under cross working conditions based on domain adaptive
When the mechanical equipment is running, the monitoring data of rolling bearing have large distribution differences and vibration signals with labels are difficult to be obtained under different conditions, which leads to the low accuracy of existing models. In order to solve the problem of cross working condition diagnosis of rolling bearing, this paper proposes a fault diagnosis method based on domain adaptive. Firstly, the fault sensitive features of vibration signal are extracted and embedded into Grassmann manifold space to avoid feature distortion in the migration process. Secondly, dynamic distribution alignment is used to adaptively adjust the difference of distribution between source domain and target domain. Moreover, by increasing the regularization term of within-class scatter, the distance within the class is minimized and the separability of the class is increased. Finally, according to the principle of structural risk minimization, the domain invariant classification model is established. Compared with the existing methods, the proposed method can effectively improve the classification accuracy.
2021 Vol. 42 (3): 267-276 [
Abstract
] (
267
)
HTML
(1 KB)
PDF
(0 KB) (
152
)
277
An Intelligent Digital-Twin-Based Strategy for the Inspection and Maintenance of Aircraft Skin Cracks
Fatigue growth of cracks in aircraft structures is affected by many aleatory and epistemic uncertainties. In order to effectively track and control such uncertainties by using the prediction results of physical knowledge-based models and observation results of manual inspections, this paper proposes an intelligent digital-twin-based strategy for the inspection and maintenance of aircraft skin cracks. Taking the single edge crack near a rivet hole as the object of study, this strategy takes advantages of the reduced order fracture mechanics simulation, the fatigue crack growth model, the data of crack length inspections, and the framework of the dynamic Bayesian network, comprehensively considering the uncertainties of the initial crack size, material parameters, flight load, measurement error etc., in order to dynamically adjust the inspection and maintenance time according to the probabilistic damage diagnosis results. Simulation results show that the method can effectively track the uncertain crack propagation, and it can provide a referencing basis for the intelligent inspection and maintenance of aircraft skin cracks.
2021 Vol. 42 (3): 277-286 [
Abstract
] (
431
)
HTML
(1 KB)
PDF
(0 KB) (
150
)
287
Back analysis of the TBM collapse section based on convolutional neural networks
This paper explores the possibility of training and predicting the collapse section of TBM excavation using the indexes TPI and FPI by the convolution neural network and time series prediction method based on the existing research work. On the basis of the principle that the input power is equivalent to the rock breaking efficiency, the authors put forward a technical approach by taking the field penetration index FPI and the torsional shear tunneling index TPI as the training and prediction elements. There are 199 columns of parameters and 17 collapse events recorded in the TBM database of Jilin pine diversion project. The total amount of data is large and the quality is high, so it has high scientific research value. Taking the above indexes as the training objects of machine learning, the normal driving section and large-scale collapse section 66 + 000-66 + 350 (chainage number) are analyzed and predicted. The results show that the measured values of FPI and TPI in the collapse section are significantly smaller, . with evaluation conclusion of "positive" based on three prediction criteria. The related research results provide a new method for big data machine learning in TBM field, and create favorable conditions for the realization of advanced geological early warning.
2021 Vol. 42 (3): 287-301 [
Abstract
] (
243
)
HTML
(1 KB)
PDF
(0 KB) (
155
)
302
A CNN-based approach for optimizing support selection of meshfree methods
王 东东
In meshfree methods, the discretized particles or nodes are inter-related by the supports of meshfree shape functions, and thus the support sizes of meshfree shape functions have a significant impact on the meshfree solution accuracy. However, due to the complexity of meshfree shape functions, currently there is still no general guidelines for meshfree support selection, and in practice the meshfree support size is often picked up by trial and error and a uniform support size is set for all particles which cannot ensure the meshfree solution accuracy. In this work, a machine learning approach, namely, the convolutional neural network (CNN), is introduced to optimize the meshfree support selection. It is shown that there is a strong similarity and linkage between the CNN receptive fields and meshfree supports, and accordingly a meshfree intrinsic CNN framework is proposed for meshfree analysis, where the network architecture and the corresponding determination of hyperparameters, such as the convolution kernel size, receptive field size, and their relationships with the number of meshfree nodes, the monomial basis function order and the support size, are particularly detailed for meshfree methods. Subsequently, within the proposed meshfree intrinsic CNN framework, several networks are designed to facilitate the meshfree support selection and solution prediction, in which a single objective like meshfree support selection, or multi-objectives like both meshfree support selection and solution prediction can be realized. Numerical results systematically demonstrate that the proposed meshfree intrinsic CNN framework is very effective to optimize the meshfree support selection with improved solution accuracy.
2021 Vol. 42 (3): 302-319 [
Abstract
] (
224
)
HTML
(1 KB)
PDF
(0 KB) (
144
)
334
Phase Field Approach to Cyclic Plasticity and Fatigue Analysis
2021 Vol. 42 (3): 334-344 [
Abstract
] (
301
)
HTML
(1 KB)
PDF
(0 KB) (
150
)
News
Download
Download
Download
Links
Links
Copyright © Editorial Board of
Supported by:
Beijing Magtech
