Pipe truss structure fatigue analysis parameter optimization method and system based on deep learning

Abstract

The present application relates to the technical field of structural fatigue evaluation, in particular to a pipe truss structure fatigue analysis parameter optimization method and system based on deep learning, the steps are as follows: a deep learning agent model is constructed by fusing a multi-scale structure graph encoder, a working condition parameter encoder and a damage dimension curve decoder; the model is trained by using a high-fidelity training data set; for a target truss structure and a load working condition, the damage dimension relationship curve of the target truss structure is predicted by the model; the optimal dimension of the rain flow counting matrix is determined by analyzing the convergence characteristics of the curve; and the multi-working condition damage values are predicted by the model, and the dominant working condition is screened out based on the damage contribution percentage. The present application replaces the traditional iterative simulation with the deep learning agent model, realizes the rapid optimization of the key parameters in the fatigue analysis, significantly improves the efficiency and accuracy of the large truss structure fatigue evaluation, and improves the generalization ability of the model by fusing the multi-scale structure graph encoder to capture the physical effects of the key hot spots.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural fatigue evaluation, in particular to a pipe truss structure fatigue analysis parameter optimization method and system based on deep learning. BACKGROUND

[0002] Large pipe truss structures, such as wind towers, offshore platforms, etc., bear complex alternating loads such as wind and waves during long-term service, and fatigue damage is a key factor affecting their safety and service life. Currently, when engineering fatigue evaluation of such structures is performed, the following technical process is usually adopted: first, stress time history data of key parts of the structure are obtained through finite element simulation, then the rainflow counting method is used to convert the stress time history into a rainflow counting matrix, and finally the cumulative fatigue damage is calculated in combination with the S-N curve and the Miner criterion. S - N

[0003] However, in the above traditional evaluation method, the following defects exist: 1. The dimension of the rainflow counting matrix lacks scientific basis; the matrix dimension (i.e. the number of discrete intervals of load amplitude and mean value) directly affects the calculation accuracy of fatigue damage, and at present the dimension is mostly set by engineers according to experience or recommended values in specifications, which has great subjectivity; if the dimension is too low, the load cycle information will be lost and the damage calculation will be conservative; if the dimension is too high, it will significantly increase the data storage and calculation burden; in order to seek a reasonable dimension, the traditional method needs to repeatedly perform the complete fatigue simulation and calculation process for different dimensions, which is time-consuming and difficult to be widely used in practical engineering. 2. The calculation efficiency of multi-condition analysis is low; according to the relevant specification requirements, fatigue evaluation needs to consider hundreds of design load conditions formed by combinations of wind speed, wind direction, running state, etc.; if full-process simulation and damage calculation are performed for each condition, the calculation cost will increase dramatically, which becomes the main bottleneck restricting the efficiency of structural optimization design and safety evaluation.

[0004] In summary, the traditional evaluation method relies on experience and lacks objective basis when determining the dimension of the rainflow counting matrix, and has the problem of low calculation efficiency when facing numerous design load conditions, resulting in low efficiency and insufficient reliability in the fatigue evaluation process. SUMMARY

[0005] The present application aims to provide a pipe truss structure fatigue analysis parameter optimization method and system based on deep learning to solve the technical problems pointed out in the background.

[0006] The present application is implemented by the following technical solution: a pipe truss structure fatigue analysis parameter optimization method based on deep learning, which is suitable for determining the optimal dimension of the rainflow counting matrix and screening design load conditions, including the following steps:

[0007] ​A deep learning agent model is constructed, which comprises a multi-scale structural graph encoder, a working condition parameter encoder, and a damage dimension curve decoder. The multi-scale structural graph encoder is used to encode the global truss topology information and local hot spot geometric information of the truss structure into a structural feature vector. The working condition parameter encoder is used to encode the design load working condition parameters into a load feature vector. The damage dimension curve decoder is used to fuse the structural feature vector and the load feature vector, and output a damage dimension prediction curve vector.

[0008] A high-fidelity training data set is obtained, and the deep learning agent model is trained using the high-fidelity training data set to obtain a trained agent model.

[0009] For the target truss structure to be analyzed and the target design load working condition, the global truss topology information and local hot spot geometric information encoding of the target truss structure, and the design load working condition parameters of the target design load working condition are determined.

[0010] The global truss topology information, local hot spot geometric information encoding, and design load working condition parameters are input into the trained agent model to obtain a damage dimension prediction curve vector.

[0011] The determination of the rainflow counting matrix dimension includes:

[0012] The optimal dimension of the rainflow counting matrix is determined by analyzing the convergence characteristics of the damage dimension prediction curve vector.

[0013] The selection of the design load working condition includes:

[0014] The dominant working condition is determined by analyzing the damage contribution percentage of each design load working condition.

[0015] According to a preferred embodiment, the multi-scale structural graph encoder comprises:

[0016] A global graph neural network module is used to encode the overall mechanical response characteristics contained in the global truss topology information to obtain a global feature vector,

[0017] A local geometric encoder module is used to encode the stress concentration effect contained in the local hot spot geometric information to obtain a local feature vector.

[0018] A fusion module is used to fuse the global feature vector and the local feature vector to obtain the structural feature vector.

[0019] According to a preferred embodiment, the damage dimension curve decoder is based on a Transformer decoder architecture, which includes a set of learnable dimension query embeddings, the number of which NEach of the dimension query embeddings corresponds to one of the preset rainflow counting dimensions.

[0020] According to a preferred embodiment, the obtaining of the damage dimension prediction curve vector specifically comprises: receiving a context vector formed by fusing the structure feature vector and the load feature vector as memory, processing a plurality of dimension query embeddings in parallel through a multi-head attention mechanism, and mapping each dimension query embedding to a corresponding output through a shared regression head to obtain an accumulated fatigue damage value, and outputting one N dimension damage dimension prediction curve vector. N

[0021] According to a preferred embodiment, the obtaining of the high-fidelity training data set comprises:

[0022] creating a plurality of parameterized truss structure samples corresponding to working condition parameter samples ;

[0023] performing high-fidelity physical simulation processing on each group of samples , including obtaining original fatigue load time series through load simulation, and establishing a multi-scale finite element analysis model to obtain a load hot spot stress curve;

[0024] performing rainflow counting and fatigue damage accumulation calculation for N preset rainflow counting matrix dimensions, respectively, to obtain real accumulated fatigue damage corresponding to each dimension;

[0025] generating a damage dimension relationship curve vector based on the real accumulated fatigue damage value, and pairing it with the corresponding sample to construct a high-fidelity training data set.

[0026] According to a preferred embodiment, the working condition parameter samples include working condition types, wind speeds, wind seeds, and inflow angle parameters.

[0027] According to a preferred embodiment, in the training process of the deep learning agent model, the minimization of the difference between the accumulated fatigue damage prediction value and the real accumulated fatigue damage value is taken as the target, and the expression of the objective function is as follows:

[0028]

[0029] In the above formula, is an input sample, is a damage dimension relationship curve vector label, is a high-fidelity training data set, represents a weighting coefficient of the dimension for adjusting the sensitivity of the convergence region, represents the ​The cumulative fatigue damage prediction value in each dimension Indicates the first The true value of cumulative fatigue damage in each dimension.

[0030] According to a preferred embodiment, determining the optimal dimension of the rainflow counting matrix includes:

[0031] Set a convergence threshold;

[0032] The relative rate of change of cumulative fatigue damage prediction values ​​between adjacent points on the damage dimension prediction curve is calculated using the following expression:

[0033]

[0034] In the above formula, This represents the relative rate of change of the predicted cumulative fatigue damage. Indicates the first The cumulative fatigue damage prediction value in each dimension;

[0035] Based on the relative rate of change of the predicted cumulative fatigue damage value The optimal dimension of the rainflow counting matrix is ​​determined by the following expression:

[0036]

[0037] In the above formula, The optimal dimension for representing the rainflow counting matrix. This represents the convergence threshold.

[0038] According to a preferred embodiment, the step of screening the design load condition specifically includes:

[0039] Extract the cumulative fatigue damage prediction value in the optimal dimension or preset dimension from the prediction curve vector of each damage dimension;

[0040] The percentage of damage contribution for each design load condition is calculated based on the cumulative fatigue damage prediction value.

[0041] The dominant load condition is determined from each design load condition based on the percentage of damage contribution.

[0042] This invention also provides a deep learning-based parameter optimization system for fatigue analysis of tubular truss structures. This system is used to execute the deep learning-based parameter optimization method for fatigue analysis of tubular truss structures as described above. The system includes:

[0043] The model construction module is configured to construct a deep learning agent model, which comprises a multi-scale structure graph encoder, a working condition parameter encoder and a damage dimension curve decoder.

[0044] The model training module is configured to obtain a high-fidelity training data set and train the deep learning agent model by using the high-fidelity training data set to obtain a trained agent model.

[0045] The data acquisition module is configured to determine the global truss topology information and the local hot spot geometric information code of the target truss structure and the design load working condition parameter of the target design load working condition.

[0046] The prediction module is configured to input the global truss topology information, the local hot spot geometric information code and the design load working condition parameter into the trained agent model to obtain a damage dimension prediction curve vector.

[0047] The first application module is configured to determine the optimal dimension of the rain flow counting matrix by analyzing the convergence characteristics of the damage dimension prediction curve vector.

[0048] The second application module is configured to determine the dominant working condition by analyzing the damage contribution percentage of each design load working condition.

[0049] The technical scheme of the pipe truss structure fatigue analysis parameter optimization method and system based on deep learning provided by the application has at least the following advantages and beneficial effects: (1) The application converts the traditional simulation process requiring multiple iterations into a one-time model forward inference by constructing and training a deep learning surrogate model, realizes the rapid completion of dimension optimization and working condition selection, greatly shortens the analysis period, and provides strong support for the rapid iteration and optimization of engineering design; (2) The application predicts the complete damage dimension relationship curve at one time, and automatically determines the optimal dimension based on the preset convergence criterion, so that the dimension selection is changed from experience setting to data-driven convergence optimization, which significantly improves the accuracy and reliability of fatigue evaluation; (3) The application uses the trained surrogate model to predict the cumulative damage of each working condition in parallel, and quantifies the influence degree of each working condition through the damage contribution percentage, quickly selects one or more dominant working conditions with the largest contribution, which not only greatly reduces the number of working conditions that need to be analyzed in detail, but also ensures that fatigue evaluation is still focused on the most critical load condition, realizing the simplification of the analysis process; (4) The multi-scale structure graph encoder adopted by the application combines global graph neural networks and local geometric encoders, which can capture the overall mechanical properties and local stress concentration effects of the structure at the same time. This design makes the model not limited to a specific topology or geometry, and can be applied to pipe truss structures of different types and sizes, and has strong generalization ability, unlike black box models for a single structure. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The overall flowchart of the pipe truss structure fatigue analysis parameter optimization method provided for the embodiment 1 of the application is shown in the figure;

[0051] Figure 2 The architecture diagram of the deep learning surrogate model provided for the embodiment 1 of the application is shown in the figure;

[0052] Figure 3 The high-fidelity training data set generation diagram provided for the embodiment 1 of the application is shown in the figure;

[0053] Figure 4 The parameter optimization diagram provided for the embodiment 1 of the application is shown in the figure;

[0054] Figure 5 The damage dimension prediction curve and optimal dimension determination diagram provided for the embodiment 1 of the application is shown in the figure;

[0055] Figure 6 The dominant working condition screening flowchart provided for the embodiment 1 of the application is shown in the figure;

[0056] Figure 7 The cumulative fatigue damage contribution percentage diagram of each working condition provided for the embodiment 1 of the application is shown in the figure;

[0057] Figure 8A structural block diagram of a pipe truss structure fatigue analysis parameter optimization system provided for Embodiment 2 of the present application. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all. The components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0059] Embodiment 1

[0060] The embodiments of the present application provide a pipe truss structure fatigue analysis parameter optimization method based on deep learning, which is suitable for determining the optimal dimension of a rainflow counting matrix and screening design load conditions, Figure 1 For the overall flowchart of the pipe truss structure fatigue analysis parameter optimization method, see Figure 1 The pipe truss structure fatigue analysis parameter optimization method includes the following steps:

[0061] Step S1, constructing and training a deep learning agent model;

[0062] In this embodiment, this step aims to establish an agent model capable of quickly and accurately predicting damage dimension curves to replace high-fidelity physical simulation; specifically including:

[0063] Step S101, model construction;

[0064] First, a deep learning agent model as shown in Figure 2 is constructed, which is composed of two encoders and one decoder, which are respectively:

[0065] Multi-scale structure graph encoder: for comprehensively capturing the mechanical characteristics of the truss structure; specifically, the multi-scale structure graph encoder further comprises a global graph neural network module, a local geometry encoder module, and a fusion module; wherein the global graph neural network module is used to take the entire tower tube rod topological structure (node coordinates, rod connection relationship) as input, learn the overall mechanical response characteristics of the structure through a multi-layer graph convolution network, and output a global feature vector; the local geometry encoder module focuses on stress concentration areas, such as pipe node welds, and inputs local geometric parameters of these hot spots, such as pipe diameter, wall thickness, chord and web angle, etc., encodes the local stress concentration effect through a multi-layer perceptron, and outputs a local feature vector; the fusion module combines the global and local feature vectors to form the final structure feature vector through splicing operation and the like. It should be noted that by fusing the global graph neural network and the local geometry encoder, the multi-scale structure graph encoder can simultaneously capture the overall mechanical characteristics and local stress concentration effect of the structure. This design makes the model not limited to a specific topology or geometric form, and can be applied to different types and sizes of pipe truss structures, and has strong generalization ability, rather than a black box model for a single structure.

[0066] Working condition parameter encoder: this part is a fully connected neural network; the input is a parameter vector of a specific design working condition, which in this embodiment includes: working condition type, average wind speed, wind seed and inflow angle, and the working condition parameter encoder is used to map these parameters into a load feature vector.

[0067] Damage dimension curve decoder: adopts a decoder architecture based on Transformer; the core is a set of learnable dimension query embeddings, the number of embeddings N corresponds to the number of dimensions of the rainflow counting matrix; the decoder receives a context vector generated by fusing the structure feature vector and the load feature vector as memory, and through a multi-head attention mechanism, each dimension query embedding interacts with the context vector in parallel to learn to focus on the features related to a specific dimension; finally, the output corresponding to each query embedding is mapped to a scalar through a shared regression head, i.e. the predicted cumulative fatigue damage value under that dimension; finally, the damage dimension curve decoder outputs a N dimensional damage dimension prediction curve vector.

[0068] Step S102, high-fidelity training data set preparation;

[0069] In order to train the deep learning agent model constructed in step S101, a high-fidelity data set needs to be prepared, as shown in Figure 3 , the specific process is as follows:

[0070] Through parameterized modeling, a batch of truss structure samples with different geometric sizes and topological forms are generated Simultaneously, by combining different operating condition parameters, a batch of operating condition parameter samples is generated. ;

[0071] For each group of samples Perform high-fidelity physical simulation processing, including obtaining the original fatigue load time series through load simulation and establishing a multi-scale finite element analysis model to obtain the load hot spot stress curve;

[0072] against N For each of the preset rainflow counting matrix dimensions, rainflow counting and fatigue damage accumulation calculation are performed to obtain the actual cumulative fatigue damage corresponding to each dimension.

[0073] Based on the actual cumulative fatigue damage values, a damage dimension relationship curve vector is generated and paired with the corresponding samples to construct a high-fidelity training dataset.

[0074] Step S103: Model training;

[0075] The high-fidelity training dataset constructed in step S102 is used to supervise the training of the deep learning proxy model constructed in step S101. Specifically, in some embodiments of this example, during the training of the deep learning proxy model, the objective is to minimize the difference between the predicted cumulative fatigue damage value and the actual cumulative fatigue damage value. The expression of the objective function is as follows:

[0076]

[0077] In the above formula, For the input sample, Labels for the damage dimension relationship curve vector. For high-fidelity training datasets, The first term represents the adjustment of the sensitivity of the convergence region. Weighting coefficients for each dimension Indicates the first The cumulative fatigue damage prediction value in each dimension Indicates the first The true value of cumulative fatigue damage in each dimension.

[0078] It should be noted that by constructing and training a deep learning proxy model, the traditional iterative simulation process is transformed into a one-time model forward inference, which can quickly complete dimensional optimization and working condition screening, greatly shortening the analysis cycle and providing strong support for rapid iteration and optimization of engineering design.

[0079] Step S2: Apply the trained agent model to optimize parameters;

[0080] See Figure 4As shown, for the target truss structure to be analyzed and the target design load condition, firstly, the global truss topology information and local hotspot geometric information encoding of the target truss structure, as well as the design load condition parameters of the target design load condition, are determined; then, the global truss topology information, local hotspot geometric information encoding and design load condition parameters are input into the trained surrogate model to obtain the damage dimension prediction curve vector.

[0081] Regarding the determination of the dimensions of the rainflow counting matrix, in this specific embodiment, it is determined by analyzing the convergence characteristics of the damage dimension prediction curve vector. See [link to relevant documentation]. Figure 5 As shown, Figure 5 The first curve on the right corresponds to the damage dimension prediction curve under the DLC1.2 load condition, the second curve corresponds to the damage dimension prediction curve under the DLC2.4 load condition, and the third curve corresponds to the total damage dimension prediction curve for all load conditions. Figure 5 The prediction curves for each damage dimension clearly demonstrate the predicted cumulative fatigue damage values. With dimension Physical properties that increase but gradually decrease and tend to converge, and fatigue life as a function of dimension. The characteristic of gradually increasing with addition; the specific steps to determine the optimal dimension are as follows:

[0082] First, set a convergence threshold;

[0083] Then, the relative rate of change of cumulative fatigue damage prediction values ​​between adjacent dimension points on the damage dimension prediction curve is calculated, as shown in the following expression:

[0084]

[0085] In the above formula, This represents the relative rate of change of the predicted cumulative fatigue damage. Indicates the first The cumulative fatigue damage prediction value in each dimension;

[0086] Finally, based on the relative rate of change of the cumulative fatigue damage prediction value... The optimal dimension of the rainflow counting matrix is ​​determined by the following expression:

[0087]

[0088] In the above formula, The optimal dimension for representing the rainflow counting matrix. This represents the convergence threshold. It is the threshold value relative to the rate of change of the predicted cumulative fatigue damage. First time less than the convergence threshold At times, for example, in dimensions hour, less than the convergence threshold If the curve is considered to have substantially converged at this point, the dimension is the optimal dimension of the structure under such working conditions.

[0089] It should be noted that by predicting the complete damage dimension relationship curve at one time and automatically determining the optimal dimension based on the preset convergence criterion, the dimension selection is changed from empirical setting to data-driven convergence optimization, which significantly improves the accuracy and reliability of fatigue evaluation.

[0090] Regarding the screening of design load conditions, in the embodiment, the dominant condition is determined by analyzing the damage contribution percentages of each design load condition, as shown in Figure 6 and Figure 7 The steps are as follows:

[0091] First, from each of the damage dimension prediction curve vectors, the cumulative fatigue damage prediction value at the optimal dimension or the preset dimension is extracted, and the total damage is calculated, as follows:

[0092]

[0093] In the above formula, represents the number of design load conditions, represents the cumulative fatigue damage prediction value of the i-th design load condition; Secondly, based on the cumulative fatigue damage prediction value, the damage contribution percentage of each design load condition is calculated, as follows:

[0094]

[0095] In the above formula,

[0096] represents the damage contribution percentage of the i-th design load condition; Finally, according to the damage contribution percentage, the dominant condition is determined from each design load condition, and the design load condition with a damage contribution percentage greater than the preset screening threshold is determined as the dominant condition.

[0097] As can be seen from , the damage contribution percentage of the normal power generation condition DLC1.2 is greater than 95%, and the total contribution of all other conditions is less than 5%. If the preset screening threshold is 95%, then since the damage contribution percentage of the normal power generation condition DLC1.2 is greater than the preset screening threshold, the normal power generation condition DLC1.2 is determined as the dominant condition.

[0098] Figure 7

[0099] ​​It should be noted that by using the trained agent model, the cumulative damage of each working condition can be predicted in parallel, and the influence degree of each working condition can be quantified by damage contribution percentage, and one or more dominant working conditions with the largest contribution can be quickly screened out. This way not only greatly reduces the number of working conditions that need to be analyzed in detail, but also ensures that fatigue evaluation is still focused on the most critical load conditions, simplifying the analysis process.

[0100] Embodiment 2

[0101] Based on the technical solutions provided in Embodiment 1, an embodiment of the present application provides a pipe truss structure fatigue analysis parameter optimization system based on deep learning, which is used to execute the pipe truss structure fatigue analysis parameter optimization method based on deep learning as described in Embodiment 1, as shown in Figure 8 The system includes:

[0102] A model construction module is configured to construct a deep learning agent model, which includes a multi-scale structure graph encoder, a working condition parameter encoder, and a damage dimension curve decoder. The multi-scale structure graph encoder is configured to encode global truss topology information and local hot spot geometric information of the truss structure into a structure feature vector. The working condition parameter encoder is configured to encode design load working condition parameters into a load feature vector. The damage dimension curve decoder is configured to fuse the structure feature vector and the load feature vector and output a damage dimension prediction curve vector.

[0103] A model training module is configured to obtain a high-fidelity training data set and train the deep learning agent model using the high-fidelity training data set to obtain a trained agent model.

[0104] A data acquisition module is configured to determine global truss topology information and local hot spot geometric information encoding of a target truss structure to be analyzed and design load working condition parameters of a target design load working condition.

[0105] A prediction module is configured to input the global truss topology information, local hot spot geometric information encoding, and design load working condition parameters into the trained agent model to obtain a damage dimension prediction curve vector.

[0106] A first application module is configured to determine the optimal dimension of a rainflow counting matrix by analyzing the convergence characteristics of the damage dimension prediction curve vector.

[0107] A second application module is configured to determine a dominant working condition by analyzing the damage contribution percentage of each design load working condition.

[0108] The functions of the modules of the deep learning-based pipe truss structure fatigue analysis parameter optimization system of the embodiment are the same as the explanation of the deep learning-based pipe truss structure fatigue analysis parameter optimization method, and the technical effects are the same, and will not be repeated here.

[0109] The above is only the preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based parameter optimization method for fatigue analysis of tubular truss structures, applicable to determining the optimal dimension of the rainflow counting matrix and selecting design load cases, characterized in that... Includes the following steps: A deep learning proxy model is constructed, which includes a multi-scale structural graph encoder, a load condition parameter encoder, and a damage dimension curve decoder. The multi-scale structural graph encoder is used to encode the global truss topology information and local hotspot geometric information of the truss structure into structural feature vectors. The load condition parameter encoder is used to encode the design load condition parameters into load feature vectors. The damage dimension curve decoder is used to fuse the structural feature vectors and load feature vectors and output the damage dimension prediction curve vector. Obtain a high-fidelity training dataset and use the high-fidelity training dataset to train the deep learning agent model to obtain the trained agent model. For the target truss structure to be analyzed and the target design load condition, determine the global truss topology information and local hot spot geometric information encoding of the target truss structure, as well as the design load condition parameters of the target design load condition; The global pole system topology information, local hotspot geometric information, and design load condition parameters are encoded and input into the trained surrogate model to obtain the damage dimension prediction curve vector. The steps for determining the dimensions of the rainflow counting matrix include: By analyzing the convergence characteristics of the damage dimension prediction curve vector, the optimal dimension of the rainflow counting matrix is ​​determined. The screening steps for the design load conditions include: The dominant load condition is determined by analyzing the percentage of damage contribution for each of the aforementioned design load conditions. The damage dimension curve decoder is based on the Transformer decoder architecture and includes a set of learnable dimension query embeddings, the number of which is... N The number of rainflow counting dimensions is the same as the preset number, and each of the dimension queries corresponds to a preset rainflow counting dimension; The acquisition of the damage dimension prediction curve vector specifically includes: receiving a context vector formed by fusing the structural feature vector and the load feature vector as memory, and processing it in parallel through a multi-head attention mechanism. N Each dimension query embedding is used, and the output corresponding to each dimension query embedding is mapped to a cumulative fatigue damage value through a shared regression head, resulting in a single output. N The damage dimension prediction curve vector.

2. The method for optimizing fatigue analysis parameters of tubular truss structures based on deep learning as described in claim 1, characterized in that, The multi-scale structure graph encoder includes: The global graph neural network module is used to encode the overall mechanical response characteristics contained in the global pole system topology information to obtain a global feature vector. The local geometry encoder module is used to encode the stress concentration effect contained in the local hotspot geometric information to obtain local feature vectors; The fusion module is used to fuse global feature vectors and local feature vectors to obtain the structural feature vector.

3. The method for optimizing fatigue analysis parameters of tubular truss structures based on deep learning as described in claim 1, characterized in that, The acquisition of the high-fidelity training dataset includes: Create several parametric truss structure samples Operating Parameter Sample ; For each group of samples Perform high-fidelity physical simulation processing, including obtaining the original fatigue load time series through load simulation and establishing a multi-scale finite element analysis model to obtain the load hot spot stress curve; against N For each of the preset rainflow counting matrix dimensions, rainflow counting and fatigue damage accumulation calculation are performed to obtain the actual cumulative fatigue damage corresponding to each dimension. Based on the actual cumulative fatigue damage values, a damage dimension relationship curve vector is generated and paired with the corresponding samples to construct a high-fidelity training dataset.

4. The method for optimizing fatigue analysis parameters of tubular truss structures based on deep learning as described in claim 3, characterized in that, The sample of operating parameters includes operating condition type, wind speed, wind seed, and inflow angle parameters.

5. The method for optimizing fatigue analysis parameters of tubular truss structures based on deep learning as described in claim 1, characterized in that, During the training of the deep learning proxy model, the objective is to minimize the difference between the predicted cumulative fatigue damage and the actual cumulative fatigue damage. The expression for the objective function is as follows: In the above formula, For the input sample, Labels for the damage dimension relationship curve vector. For high-fidelity training datasets, The first term represents the adjustment of the sensitivity of the convergence region. Weighting coefficients for each dimension Indicates the first The cumulative fatigue damage prediction value in each dimension Indicates the first The true value of cumulative fatigue damage in each dimension.

6. The method for optimizing fatigue analysis parameters of tubular truss structures based on deep learning as described in claim 1, characterized in that, Determining the optimal dimension of the rainflow counting matrix includes: Set a convergence threshold; The relative rate of change of cumulative fatigue damage prediction values ​​between adjacent points on the damage dimension prediction curve is calculated using the following expression: In the above formula, This represents the relative rate of change of the predicted cumulative fatigue damage. Indicates the first The cumulative fatigue damage prediction value in each dimension; Based on the relative rate of change of the predicted cumulative fatigue damage value The optimal dimension of the rainflow counting matrix is ​​determined by the following expression: In the above formula, The optimal dimension for representing the rainflow counting matrix. This represents the convergence threshold.

7. The method for optimizing fatigue analysis parameters of tubular truss structures based on deep learning as described in claim 1, characterized in that, The specific steps for selecting the design load conditions include: Extract the cumulative fatigue damage prediction value in the optimal dimension or preset dimension from the prediction curve vector of each damage dimension; The percentage of damage contribution for each design load condition is calculated based on the cumulative fatigue damage prediction value. The dominant load condition is determined from each design load condition based on the percentage of damage contribution.

8. A deep learning-based parameter optimization system for fatigue analysis of tubular truss structures, characterized in that, The system is used to execute the deep learning-based parameter optimization method for fatigue analysis of tubular truss structures as described in any one of claims 1 to 7, and the system comprises: The model building module is used to build a deep learning proxy model. The deep learning proxy model includes a multi-scale structural graph encoder, a load condition parameter encoder, and a damage dimension curve decoder. The multi-scale structural graph encoder is used to encode the global truss topology information and local hotspot geometric information of the truss structure into structural feature vectors. The load condition parameter encoder is used to encode the design load condition parameters into load feature vectors. The damage dimension curve decoder is used to fuse the structural feature vectors and load feature vectors and output the damage dimension prediction curve vector. The model training module is used to acquire a high-fidelity training dataset and use the high-fidelity training dataset to train the deep learning agent model to obtain the trained agent model. The data acquisition module is used to determine the global truss topology information and local hot spot geometric information encoding of the target truss structure and the design load condition parameters of the target design load condition for the target truss structure to be analyzed and the target design load condition. The prediction module is used to encode the global pole system topology information, local hot spot geometric information and design load condition parameters and input them into the trained surrogate model to obtain the damage dimension prediction curve vector; The first application module is used to determine the optimal dimension of the rainflow counting matrix by analyzing the convergence characteristics of the damage dimension prediction curve vector. The second application module is used to determine the dominant load condition by analyzing the percentage of damage contribution for each of the design load conditions.

Images