Pipeline structure stress proxy model construction method based on deep learning

Abstract

The invention discloses a deep learning-based pipeline structure stress proxy model construction method, which belongs to the field of engineering mechanics, energy transmission and artificial intelligence, and comprises the following steps of: firstly, establishing a numerical model of a pipeline, simulating different environmental conditions by applying different loads, and solving to generate a large amount of simulation data; then, feature unit extraction is carried out on the simulation data so as to compress feature dimensions; then, a deep learning initial proxy model of pipeline stress is constructed, and initial training is carried out through finite element simulation data; and finally, transfer learning is carried out based on the initial agent model, the transfer learning is divided into two stages, enhanced actual measurement sample data is used for training in the pre-training stage, and accurate actual measurement samples are used for further training and fine adjustment in the strengthening stage, so that the generalization ability of the model is improved. According to the method, the stress distribution can be rapidly and accurately solved by constructing the stress agent model, the generalization ability is high, and the problem that a traditional physical modeling method is insufficient in real-time performance and precision in stress monitoring and prediction can be solved.

Description

Technical Field

[0001] The present invention belongs to the fields of engineering mechanics, energy transmission and artificial intelligence, and specifically relates to a method for constructing a pipeline structure stress proxy model based on deep learning. Background Art

[0002] With the rapid development of urban gas industry, the construction of urban natural gas pipeline network is accelerating. However, due to the high population density and complex underground pipelines in urban areas, the probability of buried gas pipeline accidents has increased significantly. Gas pipeline accidents can cause casualties and serious property losses in cities and towns. Therefore, stress monitoring of buried pipelines is particularly important.

[0003] Pipelines are in a complex external environment for a long time, and their stress influencing factors are numerous and complex, making pipeline stress monitoring and maintenance more difficult. Traditional finite element analysis technology discretizes the pipeline structure, decomposes it into multiple finite element units, and performs numerical simulations based on boundary conditions, loads, and material properties. Although finite element analysis technology can accurately simulate the stress distribution of pipelines under different working conditions, its calculation process is complex and the amount of calculation is large, making it difficult to achieve real-time monitoring and prediction.

[0004] With the development of artificial intelligence technology, stress monitoring and prediction methods based on deep learning have gradually become a research hotspot in the field of gas pipeline monitoring. Deep learning models have powerful feature extraction capabilities and can learn and model nonlinear relationships between complex high-dimensional data, thereby effectively processing time series data and spatial features. Using deep learning models for stress monitoring and prediction can greatly improve computing efficiency and achieve real-time stress monitoring and prediction. There is no method in the prior art for pipeline structure stress monitoring and prediction using deep learning models. Summary of the invention

[0005] The purpose of the present invention is to solve the above-mentioned deficiencies in the prior art and to provide a method for constructing a pipeline structural stress proxy model based on deep learning, so as to realize real-time and accurate monitoring and risk prediction of the overall stress of the pipeline. The specific steps include modeling the pipeline using finite element simulation technology, generating simulation data, and performing feature extraction and data preprocessing. Subsequently, a deep learning proxy model is constructed and pre-trained using simulation data. Finally, transfer learning technology is introduced to perform secondary training and fine-tuning of the model using a small amount of measured data and its enhanced data, thereby enhancing the generalization ability and adaptability of the model. The stress proxy model constructed by this method has efficient and real-time stress solution capabilities, can be used for real-time stress monitoring and prediction of pipelines, and helps to improve the safety and stability of pipeline operation.

[0006] The present invention solves the technical problem by adopting the following technical solution: a method for constructing a pipeline structure stress proxy model based on deep learning, comprising the following steps:

[0007] Step 1: Establish a pipeline numerical model through finite element analysis tools, divide the pipeline structure into multiple finite elements through discretization, simulate the mechanical characteristics of the pipeline, set the pipeline geometry, material properties, soil properties and boundary conditions, apply different loads to simulate various environmental conditions, and solve the output stress simulation data;

[0008] Step 2: Extract features from the simulation data; perform correlation analysis on the finite elements of the pipeline structure using the Pearson correlation coefficient, split the three-dimensional structure into multiple clusters with similar stress changes, and select representative finite elements from each cluster to form a feature unit set, so as to compress the data feature dimension and reduce the number of sensors;

[0009] Step 3: Construct a deep learning preliminary stress proxy model, use the Transformer deep neural network to train the characteristic unit set and stress simulation data of step 2, and preliminarily establish a nonlinear mapping relationship between the characteristic unit set and the pipeline stress value;

[0010] Step 4: Based on the preliminary stress proxy model constructed in step 3, transfer learning is used to optimize the model; data enhancement is performed on the pipeline stress data measured on site, and the enhanced data and the original precise measured samples are used for training and fine-tuning in turn to make the model have better generalization ability and adaptability, thereby improving the application effect of the model in the actual environment.

[0011] The advantages of the present invention compared with the prior art are:

[0012] (1) The present invention calculates the Pearson correlation coefficient, performs correlation analysis on the finite elements of the pipeline structure, selects representative finite elements to construct a set of characteristic elements, and only needs to arrange sensors at the pipeline positions corresponding to the characteristic elements for monitoring. Finally, the monitoring data of some characteristic elements are selected as the input of the deep learning model. This can not only ensure a high model solution accuracy, but also effectively reduce the number of sensors, compress the data dimension, and improve the reliability and economy of the system.

[0013] (2) The pipeline stress proxy model is constructed using the Transformer architecture of deep learning, which can achieve rapid and accurate solution of pipeline stress distribution and improve the efficiency of pipeline stress monitoring and prediction. Through this proxy model, pipeline stress changes can be quickly captured to achieve real-time monitoring and abnormal warning. The accuracy of this invention is not inferior to that of the traditional finite element method, but the real-time performance and stress calculation efficiency are greatly improved, meeting the needs of rapid detection of pipeline stress.

[0014] (3) Based on the proxy model preliminarily trained with simulation data, the present invention introduces data enhancement and transfer learning techniques, and further optimizes the model using a small amount of measured data, thereby improving the model's adaptability and generalization capabilities to actual working conditions and achieving accurate prediction of stress distribution under complex actual environments. The present invention not only overcomes the limitation of poor adaptability of traditional simulation data models, but also effectively solves the problem of insufficient measured training data caused by the difficulty in obtaining pipeline data at gas stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A block diagram of a pipeline structure stress proxy model construction method based on deep learning in the present invention;

[0016] Figure 2 Select a flow chart for the feature unit;

[0017] Figure 3 This is the structural diagram of the Transformer model;

[0018] Figure 4 Flowchart of transfer learning for the stress agent model. DETAILED DESCRIPTION

[0019] The present invention is described in further detail below in conjunction with the accompanying drawings.

[0020] The present invention relates to a pipeline structure stress proxy model construction method based on deep learning, the structural block diagram is as follows Figure 1 As shown, it includes finite element modeling to obtain simulation data, feature unit extraction, preliminary construction of a stress proxy model based on a Transformer deep learning model, and the construction of a final pipeline stress proxy model by introducing data enhancement and transfer learning. The method of the present invention can realize the construction of a pipeline structure stress proxy model and achieve rapid and accurate stress distribution monitoring and prediction of gas pipelines under complex working conditions.

[0021] like Figure 1 As shown, a pipeline structure stress proxy model construction method based on deep learning of the present invention comprises the following steps:

[0022] Step 1: Use finite element analysis tools to perform three-dimensional fine modeling of the pipeline, divide the pipeline structure into multiple finite elements through discretization, simulate the mechanical characteristics of the pipeline, set the pipeline geometry, material properties, soil properties and boundary conditions, apply different loads to simulate various environmental conditions, and solve the output stress simulation data. The specific implementation is as follows:

[0023] Step 1.1 Use the finite element analysis tool to perform three-dimensional fine modeling of the pipeline, and divide the pipeline structure into several finite elements through the discretization method. Based on the viscoelastic theory, calculate the Prony series expression of the relaxation modulus of the pipeline material as the viscoelastic mechanical model of the pipeline. The foundation soil is considered as a soft soil base, and has the property that the compressive yield strength is much greater than the tensile yield strength. The Drucker-Prager model is used to simulate the soil properties. Select the appropriate pipe series and soil units, set the pipeline geometry, material performance parameters, soil properties and cover thickness. Divide the pipeline structure into several finite elements through the discretization method.

[0024] Step 1.2 sets constraints and boundary conditions, considers the interaction between the pipeline and the soil, and considers friction and relative displacement. The pipe-soil contact is considered as surface-to-surface contact, and the contact mode and friction coefficient are set. Internal pressure is applied to the inner surface of the pipeline, ground load is applied to the soil surface directly above the pipeline, symmetric constraints are applied to the sides of the pipeline and the remaining soil surface, and a gravity factor is added to the whole. According to the Boussinesq theory, the static load calculation formula for buried pipelines is:

[0025] ;

[0026] Where: is the static load, in units of ; is the ground force, in units of ; is the vertical distance from the load to the top of the pipe, in units of ; is the length between the load point and the top of the pipe, in units of ; is the width between the load action point and the top of the pipe, in units of ; is the height between the load action point and the top of the pipe, in units of .

[0027] The calculation formula for pipeline dynamic load is:

[0028] ;

[0029] Where: For dynamic load ; is the dynamic load concentration adjustment coefficient, which is taken as 2.0.

[0030] The total load is the sum of the static load and the dynamic load.

[0031] Apply constraints and loads. The left and right sides of the soil are restricted to the x-direction only, the bottom is restricted to the y-direction displacement, a load is applied to the upper surface of the soil, the internal pressure of the pipe is set, and gravity is applied at the same time. .

[0032] Step 1.3 applies different constraints and load combinations to simulate different environmental conditions, performs finite element solution, records the stress change data of each finite element, and outputs a large amount of stress distribution simulation data.

[0033] Step 2: Extract features from the simulation data, perform correlation analysis on the finite element of the pipeline structure through the Pearson correlation coefficient, and construct the pipeline finite element unit correlation matrix , as shown below:

[0034] ;

[0035] Correlation coefficient The value range of is [-1,1], The larger the absolute value of is, the stronger the correlation is. The closer it is to 0, the weaker the correlation. When the finite element and finite element There is a significant correlation between them. The finite elements with obvious correlation to the load in the three-dimensional structure are classified into the same cluster, and representative finite elements are selected from each cluster to form a characteristic element set to compress the data feature dimension and reduce the number of sensors. The characteristic element extraction flow chart is shown in Figure 2 As shown, the specific implementation is as follows:

[0036] Step 2.1: From the overall structure, select the finite element that cannot be installed with the sensor, and select the element with the largest absolute stress value among the remaining finite elements. The large absolute stress value indicates that the finite element is sensitive to the external load. Arranging the measuring point at this position is conducive to sensing the small changes of the external load and the structural response.

[0037] Step 2.2 A finite element set of sub-cycles The finite element with the largest absolute value of stress is taken as the characteristic element ,in is the number of cycles. Other units and The correlation coefficient, j ,in for The number of finite elements in , if the correlation coefficient 0.95, then Add to collection ;

[0038] Step 2.3 Zhongyu The repeated finite elements are deleted to obtain the remaining finite element set, which is used as the finite element set for the next cycle.

[0039] Step 2.3 Repeat steps 2.2 and 2.3 until the number of remaining finite element sets is 0, then the feature element extraction is completed, and the loop is executed. Finally, tens of thousands of finite elements can be divided into Cluster , and get the characteristic unit set:

[0040] ;

[0041] Step 3 builds a preliminary deep learning proxy model, uses the Transformer deep neural network to train the feature unit set and simulation data of step 2, and preliminarily establishes a nonlinear mapping relationship between the feature unit set and the pipeline stress value. The specific implementation is as follows:

[0042] Step 3.1 Plan the data samples, balance, randomly shuffle and normalize the sample data, and divide them into non-overlapping training sets (70%), validation sets (20%) and test sets (10%).

[0043] Step 3.2 builds a stress proxy model based on the Transformer model, evaluates the model performance through a series of comparative tests, and determines the model's hyperparameters, including the number of input feature units, the number of attention heads, the number of network layers, and the feature vector dimension, in order to take into account both model performance and generalization ability.

[0044] Step 3.3 uses the training set divided in step 3.1 to train the model, uses the validation set to judge the convergence of the neural network, uses the test machine to evaluate the accuracy of the converged model, and finally selects the model with better accuracy as the preliminary stress proxy model.

[0045] The structure of the Transformer model is as follows Figure 3 As shown in the figure, it consists of N encoders and N decoders. The stress data of the feature unit set is used as the model input, which is first sent to the encoder after input embedding and position encoding. In each encoder layer, multi-dimensional feature extraction and nonlinear transformation are performed through a multi-head self-attention mechanism and a feedforward neural network. The encoder output of the last layer is passed as input to all decoders.

[0046] The decoder structure is similar to the encoder, but a masked multi-head self-attention mechanism is added to ensure that only the elements before the current position are focused on when generating a sequence to avoid information leakage. The output of the final decoder is linearly transformed and the Softmax layer is used to obtain the final model output.

[0047] In order to optimize the training efficiency of multi-layer networks, residual connections and layer normalization are added after each decoder layer to address the challenges in deep network training. Residual connections add the input directly to the output, combined with layer normalization, to normalize the output, thereby losslessly retaining the previous layer information and significantly improving model performance.

[0048] The Transformer model proposes a new position encoding mechanism to capture the time series information between input data and output data, which is expressed using the following formula:

[0049] ;

[0050] ;

[0051] Where pos is the length of the index sequence, i is the index of the dimension, is the word vector dimension, and PE represents the position of the sequence.

[0052] The self-attention mechanism allows the neural network to automatically pay attention to other positions related to the current position when processing different positions, and give these positions different weights. The core process is to obtain the attention weights by calculating the dot product of the query matrix Q and the key-value matrix K, and then use these weights to weight the value matrix V to obtain the self-attention output. The calculation formula is as follows:

[0053] ;

[0054] in, Indicates transposing the matrix. represents the word embedding dimension, Represents the attention calculation value.

[0055] The multi-head attention mechanism is to perform multiple groups of self-attention processing on the original input sequence, and then piece together the results of each group of self-attention. The calculation formula is as follows:

[0056] ;

[0057] in Indicates The value of group attention, , , Respectively represent The linear transformation matrix corresponding to the group attention Q, K, V, Represents the output value of the multi-head attention mechanism, It means to splice multiple attentions. Represents the linear transformation matrix during the splicing process.

[0058] The Transformer model can effectively capture the interdependence of finite elements under load and establish the relationship between each finite element, greatly improving the model accuracy and efficiency of traditional neural networks.

[0059] Step 4 uses transfer learning to optimize the model based on the preliminary stress proxy model constructed in step 3. Data enhancement is performed on the pipeline stress data measured on site, and the enhanced data and the original accurate measured samples are used for training and fine-tuning in turn to make the model have better generalization ability and adaptability, and improve the application effect of the model in the actual environment. The process is as follows: Figure 4 As shown, the specific implementation is as follows:

[0060] Step 4.1: According to the feature unit set extracted in step 2, three-dimensional strain gauges are arranged at the corresponding positions of the pipeline. The strain gauge monitoring data at the corresponding positions of the feature units determined in step 3 are used as the input of the stress proxy model. In addition, 6 high stress points are selected as fine-tuning points for transfer learning, and 6 assessment points are used to evaluate the generalization performance of the model.

[0061] Step 4.2 Before using the data of each measuring point, it is necessary to pre-process and clean the data. Use Fourier transform to draw the spectrum diagram for spectrum analysis, design Fourier filter and smoothing filter to filter out noise interference and unnecessary detail oscillation in stress acquisition data. Based on the symmetry of sample data, symmetry operation is applied to the sample to generate approximate samples, which can expand the sample size and provide prior knowledge for the model to improve its adaptability to actual working conditions.

[0062] Step 4.3 Transfer learning is divided into two stages. In the pre-training stage, the approximate sample set generated by data enhancement is used to train the preliminary proxy model constructed in step 3, so that the model gradually adapts to the environmental conditions of the real scene, learns and matches the distribution law of stress changes in the actual field. In the reinforcement stage, precise measured data is used for fine-tuning training to improve the generalization ability of the model, enhance the model's adaptability to actual application scenarios and prediction accuracy. Specific fine-tuning methods include dynamically adjusting the learning rate, introducing new network layers, freezing some network layers and the number of hidden layer nodes, etc. Finally, the measured data of the assessment points are used for model verification and performance evaluation to ensure the adaptability and accurate prediction ability of the final model to the real scene. The model that finally meets the evaluation indicators is used as the final stress proxy model.

[0063] In summary, the present invention designs a pipeline structure stress proxy model based on deep learning. First, a finite element data model of the pipeline is constructed, and feature extraction and dimensionality reduction are performed on the simulation data. The nonlinear mapping relationship between stress distribution and multi-physical fields is realized by combining a deep neural network based on the Transformer architecture. Finally, data enhancement and transfer learning techniques are introduced to optimize the model to improve the adaptability and generalization ability of the model. The method of the present invention can realize the construction of a pipeline structure stress proxy model based on deep learning, and the model can realize rapid and accurate stress distribution monitoring and prediction of gas pipelines under actual complex working conditions.

[0064] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in this field.

[0065] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for constructing a pipeline structure stress proxy model based on deep learning, characterized in that: The steps include: Step 1: Establish a pipeline numerical model through finite element analysis tools, divide the pipeline structure into multiple finite elements through discretization, simulate the mechanical characteristics of the pipeline, set the pipeline geometry, material properties, soil properties and boundary conditions, apply different loads to simulate various environmental conditions, and solve the output stress simulation data; Step 2: Extract features from simulation data; The correlation analysis of the finite elements of the pipeline structure is carried out through the Pearson correlation coefficient, and the three-dimensional structure is split into multiple clusters with similar stress changes. Representative finite elements are selected from each cluster to form a characteristic unit set to compress the data feature dimension and reduce the number of sensors. Step 3: Construct a deep learning preliminary stress proxy model, use the Transformer deep neural network to train the characteristic unit set and stress simulation data of step 2, and preliminarily establish a nonlinear mapping relationship between the characteristic unit set and the pipeline stress value; Step 4: Based on the preliminary stress proxy model constructed in step 3, transfer learning is used to optimize the model; data enhancement is performed on the pipeline stress data measured on site, and the enhanced data and the original precise measured samples are used for training and fine-tuning in turn to make the model have better generalization ability and adaptability, thereby improving the application effect of the model in the actual environment.

2. According to a method for constructing a pipeline structure stress proxy model based on deep learning according to claim 1, it is characterized in that: The step 1 comprises: Step 1.1: Use finite element analysis tools to perform three-dimensional fine finite element modeling on the buried pipeline, discretize the pipeline into several finite elements, select appropriate element types to model the pipeline and surrounding soil, accurately simulate their contact relationship, and divide the grid; Step 1.2, set constraints and boundary conditions, apply internal pressure to the inner surface of the pipeline, apply ground load to the soil surface directly above the pipeline, apply symmetric constraints to the sides of the pipeline and the remaining soil surfaces, add gravity factors to the whole, consider friction and relative displacement, regard the pipe-soil contact as surface-to-surface contact, set the contact mode and friction coefficient, and simulate the combination of dynamic load and static load; Step 1.3: Apply different constraints and load combinations to simulate different environmental conditions, calculate and record the equivalent stress of each finite element, and output stress simulation data.

3. According to a method for constructing a pipeline structure stress proxy model based on deep learning according to claim 2, it is characterized in that: The step 2 performs correlation analysis on the finite elements of the pipeline structure based on the Pearson correlation coefficient. According to the constructed finite element correlation matrix, the finite elements with similar stress direction changes in the pipeline model are classified into the same cluster. Representative finite elements are selected in each cluster as monitoring points, i.e., characteristic elements. The characteristic elements are sensitive to the magnitude and direction of pipeline stress and have a strong correlation with the remaining finite elements in response.

4. According to a method for constructing a pipeline structure stress proxy model based on deep learning according to claim 3, it is characterized in that: The step 3 comprises: Step 3.1: Data sample planning: balance, randomly shuffle and normalize the sample data, and divide it into non-overlapping training sets, validation sets and test sets; Step 3.2: Use the Transformer architecture to build a deep learning agent model. Through a series of comparative experiments, the hyperparameters are set one by one, including the number of attention heads, the number of network layers, and the dimension of feature vectors, to balance the model performance and generalization ability. Step 3.3: Use the training set divided in step 3.1 to train the model, use the validation set to judge the convergence of the neural network, use the test set to test the model accuracy of the converged model, and select the model with better accuracy as the preliminary stress proxy model.

5. The method for constructing a pipeline structure stress proxy model based on deep learning according to claim 4 is characterized in that: The step 4 comprises: Step 4.1: Arrange three-dimensional strain gauges at the position of pipeline characteristic units, where the monitoring data of the strain gauges at the pipeline position corresponding to the characteristic units selected in step 3.2 are used as the input of the Transformer model. In addition, multiple high stress points are selected as fine-tuning points for transfer learning, and multiple assessment points are used to evaluate the generalization performance of the model. Step 4.2: Before using the data of each measuring point, it is necessary to draw a spectrum diagram through Fourier transform, perform data cleaning and preprocessing to improve the quality and availability of the data; and generate approximate samples through symmetric data enhancement methods to expand the sample size and provide prior knowledge for the model; Step 4.3, transfer learning is divided into two stages. The pre-training stage uses the approximate sample set generated by data enhancement to train the model so that the model gradually adapts to the environmental conditions of the real scene. The reinforcement stage uses precise measured data for fine-tuning training to enhance the model's adaptability to actual application scenarios and prediction accuracy, and uses measured assessment point data to verify the model and evaluate its performance.

6. The method for constructing a pipeline structure stress proxy model based on deep learning according to claim 5, characterized in that: The Transformer model consists of N encoders and N decoders; the input is sent to the encoder after input embedding and position encoding. In each encoder layer, multi-dimensional feature extraction and nonlinear transformation are performed through a multi-head self-attention mechanism and a feedforward neural network. The encoder output of the last layer is passed as input to all decoders. The decoder structure is similar to that of the encoder, but a masked multi-head self-attention mechanism is additionally added. The final output of the decoder is subjected to a linear transformation and a Softmax layer to obtain the final model output. Residual connections and layer normalization are added after each decoder layer to standardize the output and retain the previous layer information losslessly to improve model performance.

7. The method for constructing a pipeline structure stress proxy model based on deep learning according to claim 5, characterized in that: Transfer learning is used to fine-tune the model by adjusting the learning rate, introducing new network layers, freezing certain layers, or hiding network nodes to improve the model's stress prediction capabilities under various working conditions.

Images