A long-period spectrum load fast analysis and calculation method based on artificial neural network
By using an artificial neural network-based method, the problems of high computational resource consumption and low efficiency in long-period spectral load analysis are solved, achieving efficient and automated spectral load response prediction, which is applicable to a variety of complex engineering structures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG AEROSPACE UNIVERSITY
- Filing Date
- 2026-05-06
- Publication Date
- 2026-07-14
Smart Images
Figure CN122389623A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of structural spectral analysis technology, and in particular to a fast calculation method for long-period spectral analysis based on artificial neural networks. Background Technology
[0002] Long-period spectral load analysis is a crucial computational step in structural strength assessment, response prediction, and engineering design, especially suitable for large engineering structures subjected to complex variable amplitude cyclic loads. Traditional methods typically employ finite element analysis to numerically solve for each load cycle in the load spectrum to obtain the structural response under complete spectral load conditions.
[0003] However, when the load spectrum contains tens of thousands or even hundreds of thousands of load cycles, the traditional full-spectrum finite element cycle-by-cycle analysis method will result in a huge amount of computational workload and time, occupying a large amount of computing resources and severely restricting the efficiency of engineering analysis.
[0004] Furthermore, since the load cycles in the spectral load analysis process usually have high similarity and repeatability, the existing technology still performs finite element calculations on each load cycle one by one, resulting in a large number of repeated calculations and significant computational redundancy, which makes it difficult to meet the application requirements of rapid analysis and high-frequency iterative design of complex engineering structures.
[0005] Furthermore, in existing engineering analysis processes, finite element analysis typically lacks effective integration with intelligent prediction algorithms. An automated analysis method capable of rapidly predicting finite element results for long-period spectral load characteristics has not yet been developed, resulting in a still low level of automation and computational efficiency in spectral load analysis.
[0006] Therefore, developing an automated calculation method and system that combines finite element analysis and artificial neural network technology to perform rapid response prediction and batch analysis of long-period spectral loads is of great engineering application value for improving the efficiency of structural spectral load analysis, reducing calculation costs, and enhancing engineering application capabilities. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a fast analysis and calculation method for long-period spectral loads based on artificial neural networks.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: a fast analysis and calculation method for long-period spectral loads based on artificial neural networks, comprising the following steps: Acquire initial load historical data, preprocess the initial load historical data, and generate a standardized load time history file; Rainflow counting processing is performed on the standardized load time history file to generate load cyclic spectrum data; Based on the load cycle spectrum data, representative load cycle samples are selected to construct a training load subset; A finite element model of the target structure is established in the finite element software, and finite element analysis is performed based on the training load subset to obtain training sample result data. Extract the response results corresponding to each analysis node from the training sample result data to construct a neural network training dataset; An artificial neural network model is trained based on the neural network training dataset to obtain a structural response prediction model; The remaining loads that did not participate in the finite element analysis from the load cyclic spectrum data are cyclically input into the structural response prediction model for prediction, and the predicted structural response results are obtained. The accuracy of the predicted structural response results is verified, and the training sample result data and the predicted structural response results are integrated to generate a complete spectral response database and output it.
[0009] Preferably, the steps of acquiring initial load history data, preprocessing the initial load history data, and generating a standardized load time history file are as follows: Read the initial timestamp information of each line in the initial load history data; By comparing the initial timestamp intervals of two adjacent rows of data in the initial load history data, if the initial timestamp intervals are found to be non-increasing, it is determined that there is an anomaly in the data timing order. A sorting algorithm is used to rearrange the data segments that have experienced the data time sequence disorder anomaly according to the initial timestamp information in ascending order to obtain time sequence corrected data. Check whether the load value field of each record in the time-series correction data is empty. If the load value field has an empty value, it is determined that a partial data missing anomaly has occurred. An interpolation algorithm is used to combine the normal values before and after the missing values to complete the load data; A uniform time step is generated based on a preset sampling frequency; Data segments in the complete load data that do not meet the unified time step requirement are resampled to generate the standardized load time history file.
[0010] Preferably, the step of performing rainflow counting processing on the standardized load time history file to generate load cyclic spectrum data specifically includes: Traverse all time-series load points in the standardized load time history file; Identify local maxima and local minima among all time series load points to construct a peak-valley alternation sequence; The four-point rainflow counting method was used to determine the cyclic closure of the peak-valley alternation sequence, and each closed cycle was obtained. Statistically analyze the load amplitude, average load, and frequency of occurrence for each closed loop. The closed loops are rearranged in chronological order to generate the load cyclic spectrum data.
[0011] Preferably, the step of selecting representative load cycle samples based on the load cycle spectrum data and constructing a training load subset specifically includes: The K-means clustering algorithm is used to perform cluster analysis on the load feature space of each load cycle in the load cycle spectrum data to obtain multiple cluster centers; A predetermined number of load cycles are selected from the vicinity of the multiple cluster centers as representative load cycle samples; The representative cyclic load samples are combined to construct the training load subset.
[0012] Preferably, the steps of establishing a finite element model of the target structure in finite element software and performing finite element analysis based on the training load subset to obtain training sample result data are as follows: Establish a geometric model of the target structure; Set the material properties and boundary conditions of the target structural geometric model; Mesh generation is performed on the geometric model of the target structure to obtain the finite element model of the target structure; Obtain the load amplitude and average load corresponding to each representative load cycle sample in the training load subset, and use them as load input parameters to convert them into finite element load boundary conditions; For each of the representative load cycle samples, the finite element solver is called to perform static analysis, and the response results of each analysis node in the finite element model under each of the representative load cycle samples are extracted. The response results are stored according to the correspondence between the finite element load boundary conditions and the response results to generate the training sample result data.
[0013] Preferably, the step of extracting the response results corresponding to each analysis node from the training sample result data and constructing the neural network training dataset specifically includes: Extract the load input parameters corresponding to each representative load cycle sample and the response results corresponding to each analysis node from the training sample result data; Establish a mapping relationship between the load input parameters and the response results of each analysis node; The neural network training dataset is constructed based on the mapping relationship.
[0014] Preferably, the specific steps for training an artificial neural network model based on the neural network training dataset to obtain a structural response prediction model are as follows: Construct an artificial neural network model that includes an input layer, hidden layers, and an output layer; The payload input parameters from the neural network training dataset are input into the input layer, processed by the multi-layer fully connected network structure of the hidden layer, and the predicted response results of each analysis node are output by the output layer. Calculate the loss function between the predicted response and the response in the neural network training dataset; The backpropagation algorithm is used to update the network weights of the artificial neural network model; Training stops when the loss function is less than a preset error threshold, and the structural response prediction model is obtained.
[0015] Preferably, the specific steps for inputting the remaining loads from the load cyclic spectrum data that did not participate in the finite element analysis into the structural response prediction model to obtain the predicted structural response results are as follows: The remaining load cycles that did not participate in the finite element analysis in the load cycle spectrum data are input one by one into the structural response prediction model, and batch neural network forward inference is performed. Output the predicted response results for each of the remaining load cycles; The predicted structural response results are generated by combining the predicted response results corresponding to each remaining load cycle.
[0016] Preferably, the steps of performing accuracy verification on the predicted structural response results, integrating the training sample result data and the predicted structural response results, generating a complete spectral response database and outputting it are as follows: Select verification samples from the remaining load cycles based on error sensitivity; A finite element verification analysis was performed on the verification sample to obtain the finite element verification results; Extract the predicted response result corresponding to the verification sample from the predicted structural response result, and calculate the error between the predicted response result and the finite element verification result; If the error exceeds the preset supplementation threshold, the verification sample is added to the training load subset, and the structural response prediction model is retrained until the accuracy requirements are met. The predicted structural response results that meet the accuracy requirements are merged with the training sample result data to generate the complete spectral response database and output it.
[0017] This invention also provides a fast analysis and calculation system for long-period spectral loads based on artificial neural networks, comprising: The data preprocessing module is used to acquire initial load historical data, preprocess the initial load historical data, and generate a standardized load time history file. The rainflow counting module is used to perform rainflow counting processing on the standardized load time history file to generate load cyclic spectrum data. The sample selection module is used to select representative load cycle samples based on the load cycle spectrum data and construct a training load subset. The finite element analysis module is used to build a finite element model of the target structure in the finite element software, and to perform finite element analysis based on the training load subset to obtain training sample result data. The data construction module is used to extract the response results corresponding to each analysis node in the training sample result data and construct the neural network training dataset; The neural network training module is used to train an artificial neural network model based on the neural network training dataset to obtain a structural response prediction model. The batch prediction module is used to cyclically input the remaining loads in the load cyclic spectrum data that did not participate in the finite element analysis into the structural response prediction model for prediction, and obtain the predicted structural response results. The result output module is used to verify the accuracy of the predicted structural response results, integrate the training sample result data and the predicted structural response results, generate a complete spectral response database and output it.
[0018] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. The computational efficiency is significantly improved. By performing finite element analysis on only a small number of representative load cycles and using artificial neural networks to predict the response of the remaining large-scale load cycles, the number of finite element calculations required for long-period spectral load analysis is greatly reduced.
[0019] 2. Reduce redundant calculations: Effectively utilize the redundancy characteristics of a large number of similar load cycles in long-period spectrum loads to avoid the traditional method of repeatedly solving the finite element problem for all load cycles.
[0020] 3. Ensure analysis accuracy: The artificial neural network model is trained based on finite element samples, and combined with error verification and adaptive sampling mechanism to ensure that the prediction results meet the accuracy requirements of engineering analysis.
[0021] 4. High degree of automation: It realizes fully automated analysis from load data processing, sample screening, finite element analysis, model training to batch prediction, reducing manual intervention.
[0022] 5. Wide range of applications: It can be applied to large-scale spectral response analysis scenarios of various complex engineering structures and has good engineering promotion and application value. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 A schematic diagram of the finite element mesh generation for the corner box component; Figure 3 The Mises stress contour plot of the corner box component under load; Figure 4 Plot the model training loss curve; Figure 5 This is a diagram showing the distribution of axial force residuals. Figure 6 This is a diagram showing the distribution of tangential force residuals. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0025] Please see Figures 1-6 This invention provides a technical solution: a fast analysis and calculation method for long-period spectral loads based on artificial neural networks, comprising the following steps: Acquire historical initial load data, preprocess the historical initial load data, and generate a standardized load time history file; Rainflow counting is performed on the standardized load time history file to generate load cyclic spectrum data; Representative load cycle samples were selected based on load cycle spectrum data to construct a training load subset; A finite element model of the target structure is established in the finite element software, and finite element analysis is performed based on a subset of training loads to obtain training sample result data. Extract the response results corresponding to each analysis node from the training sample result data to construct a neural network training dataset; An artificial neural network model is trained based on a neural network training dataset to obtain a structural response prediction model. The remaining loads that were not involved in the finite element analysis in the load cyclic spectrum data are cyclically input into the structural response prediction model to predict the structural response results. The accuracy of the predicted structural response results is verified, and the training sample results and predicted structural response results are integrated to generate and output a complete spectral response database.
[0026] In this embodiment, the steps of acquiring initial load historical data, preprocessing the initial load historical data, and generating a standardized load time history file are as follows: Read the initial timestamp information of each row in the initial load historical data; compare the initial timestamp intervals of adjacent rows in the initial load historical data; if the initial timestamp interval is detected to be non-increasing, it is determined that there is a data time sequence disorder anomaly; use a sorting algorithm to rearrange the data segments with data time sequence disorder anomalies according to the initial timestamp information in ascending order to obtain time sequence corrected data; check whether the load value field of each row in the time sequence corrected data is empty; if the load value field contains empty values, it is determined that there is a partial data missing anomaly; use an interpolation algorithm to combine the normal values before and after the empty values to complete the data, obtaining complete load data; generate a unified time step according to a preset sampling frequency; resample the data segments in the complete load data that do not meet the unified time step requirements to generate a standardized load time history file.
[0027] Specifically, the initial timestamp information of each row in the initial load historical data is read, and the initial timestamp interval of two adjacent rows is compared. If a non-increasing initial timestamp interval is detected, it is determined that there is a data time-series disorder anomaly. A quicksort algorithm is used to rearrange the data segments with the time-series disorder anomaly in ascending order of the initial timestamp information to obtain time-series corrected data. The load value field of each row in the time-series corrected data is checked to see if it is empty. If there is a null value, it is determined that there is a partial data missing anomaly. A cubic spline interpolation algorithm is used to fill in the missing value with five normal values before and after it. For example, if the first... The first load value is missing; extract the first... to Passing the exam to Ten normal load values were obtained. The missing values were filled in by calculating the estimated values through cubic polynomial fitting to obtain complete load data. A uniform time step of 0.02 seconds was generated based on the inherent sampling clock frequency of 50Hz of the external hardware acquisition device. Data segments in the complete load data that did not meet the uniform time step requirement were resampled. For data segments with a sampling step greater than 0.02 seconds, linear interpolation was used to supplement data points. For data segments with a sampling step less than 0.02 seconds, mean downsampling was performed to merge redundant data points. The time interval between all adjacent load points was fixed at 0.02 seconds. After resampling, the processed data sequence was saved according to a uniform format specification to generate a standardized load time history file.
[0028] In this embodiment, the steps for performing rainflow counting processing on the standardized load time history file to generate load cycle spectrum data are as follows: traversing all time series load points in the standardized load time history file; identifying local maxima and local minima among all time series load points to construct a peak-valley alternation sequence; using the four-point rainflow counting method to perform cycle closure judgment on the peak-valley alternation sequence to obtain each closed cycle; statistically analyzing the load amplitude, average load, and occurrence frequency corresponding to each closed cycle; and rearranging each closed cycle in chronological order to generate load cycle spectrum data.
[0029] Specifically, the process iterates through all time-series load points in the standardized load time history file, comparing the current load point with its preceding and following load points. If the current load point is greater than its preceding and following load points, it is marked as a local maximum; if it is less than its preceding and following load points, it is marked as a local minimum. All intermediate time-series load points that are neither maximum nor minimum are removed. The remaining extreme points are connected in chronological order to construct a peak-valley alternation sequence. A four-point rainflow counting method is used to perform cyclic closure checks on the peak-valley alternation sequence. Four consecutive extreme points are extracted from the peak-valley alternation sequence, and the first load amplitude formed by the first three extreme points and the load amplitude formed by the last three extreme points are calculated. The second load amplitude is calculated. If the first load amplitude is less than or equal to the second load amplitude, and the cycle formed by the first two extreme points is completely contained within the cycle formed by the last two extreme points, the first two extreme points are determined to form a closed cycle. These two extreme points are removed from the peak-valley alternation sequence to obtain each closed cycle. If the closure condition is not met, the extreme point is moved to the next extreme point to continue the judgment until the entire peak-valley alternation sequence is processed. The load amplitude corresponding to each closed cycle is counted. Half of the absolute value of the difference between the maximum and minimum values in the closed cycle corresponds to the load amplitude, and half of the sum of the maximum and minimum values in the closed cycle corresponds to the average load. The number of times each closed cycle appears in the entire sequence is recorded. The closed cycles are rearranged in chronological order to generate load cycle spectrum data.
[0030] In this embodiment, the steps of selecting representative load cycle samples based on load cycle spectrum data and constructing a training load subset are as follows: calling the K-means clustering algorithm to perform cluster analysis on the load feature space of each load cycle in the load cycle spectrum data to obtain multiple cluster centers; selecting a preset number of load cycles from the vicinity of multiple cluster centers as representative load cycle samples; and combining the representative load cycle samples to construct a training load subset.
[0031] Specifically, the K-means clustering algorithm is used to perform cluster analysis on the load feature space of each load cycle in the load cycle spectrum data, extracting the load amplitude and average load of each load cycle as a two-dimensional feature vector, and randomly initializing the number of cluster centers. Number of cluster centers The inflection point of the sum of squared errors is determined by the elbow rule. The sum of squared errors from 1 to 10, when As the sum of squared time errors decreases, the number of cluster centers is set to 5. The Euclidean distance from the two-dimensional feature vector of each load cycle to each cluster center is calculated. Each load cycle is assigned to the cluster containing the cluster center with the smallest distance. The mean of the two-dimensional feature vectors of all load cycles in each cluster is updated as the new cluster center. The assignment and update steps are repeated until the position change of all cluster centers is less than the set convergence threshold of 0.001, resulting in multiple cluster centers. A preset number of load cycles are selected from the vicinity of multiple cluster centers. The Euclidean distance from each load cycle in the cluster corresponding to each cluster center to that cluster center is calculated. The first 20 load cycles are selected as representative load cycle samples in order of distance from near to far. The representative load cycle samples selected from different clusters are spliced and fused to combine the representative load cycle samples and construct a training load subset.
[0032] In this embodiment, the steps of establishing a finite element model of the target structure in finite element software and performing finite element analysis based on a subset of training loads to obtain training sample result data are as follows: establishing a geometric model of the target structure; setting the material properties and boundary conditions of the geometric model of the target structure; performing mesh generation on the geometric model of the target structure to obtain a finite element model of the target structure; obtaining the load amplitude and average load corresponding to each representative load cycle sample in the training load subset, and using them as load input parameters to convert them into finite element load boundary conditions; calling the finite element solver to perform static analysis for each representative load cycle sample, and extracting the response results of each analysis node in the finite element model under each representative load cycle sample; storing the response results according to the correspondence between the finite element load boundary conditions and the response results to generate training sample result data.
[0033] Specifically, a geometric model of the target structure is established, defining its geometric dimensions and topology. Material properties are set, including input elastic modulus, Poisson's ratio, and density constitutive parameters. Boundary conditions are set, and all degrees of freedom at the bottom of the structure are fixed using a fully constrained approach. Mesh generation is performed on the target structure geometric model, which is then discretized using hexahedral solid elements. Pre-defined chamfered and perforated areas are locally meshed to obtain the finite element model of the target structure. The load amplitude and average load corresponding to each representative load cycle sample in the training load subset are obtained. The sum of the load amplitude and average load is used as the upper limit load, and the load amplitude and average load are subtracted. The difference is used as the lower limit load. The upper and lower limit loads are used as load input parameters and converted into finite element load boundary conditions. The finite element load boundary conditions are applied to the preset force surface of the finite element model. Static analysis is performed by calling the finite element solver for each representative load cycle sample. The stress and strain distribution of the finite element model under each representative load cycle sample is calculated. The response results of each analysis node in the finite element model under each representative load cycle sample are extracted. The response results include the Mises equivalent stress value and principal strain value of each analysis node. The response results are stored in a two-dimensional mapping table according to the correspondence between the finite element load boundary conditions and the response results, and training sample result data is generated.
[0034] In this embodiment, the steps of extracting the response results corresponding to each analysis node in the training sample result data and constructing the neural network training dataset are as follows: extracting the load input parameters corresponding to each representative load cyclic sample and the response results corresponding to each analysis node in the training sample result data; establishing the mapping relationship between the load input parameters and the response results of each analysis node; and constructing the neural network training dataset based on the mapping relationship.
[0035] Specifically, the load input parameters corresponding to each representative load cycle sample in the training sample results data are extracted. The load input parameters include the upper and lower limits of the load for each representative load cycle sample. The response results corresponding to each analysis node are extracted. The response results include the Mises equivalent stress and principal strain values for each analysis node. The load input parameters and response results are normalized, and the maximum-minimum scaling method is used to map the load input parameters and response results to an interval, adjusting input and output data of different dimensions to the same numerical range, thus establishing a mapping between the load input parameters and the response results of each analysis node. The relationship is as follows: the normalized upper and lower loads are concatenated to form the input feature vector; the normalized Mises equivalent stress and principal strain values of each analysis node are concatenated to form the target output feature vector; the input feature vector and the target output feature vector are paired to form a supervised learning sample combination; all supervised learning sample combinations are randomly shuffled and split into a preset ratio of 80% as the training set and 20% as the validation set; the training set and validation set are packaged into tensor formats respectively; and a neural network training dataset is constructed based on the mapping relationship and the generated training set and validation set tensors.
[0036] In this embodiment, the steps for training an artificial neural network model based on a neural network training dataset to obtain a structural response prediction model are as follows: constructing an artificial neural network model containing an input layer, a hidden layer, and an output layer; inputting the load input parameters from the neural network training dataset into the input layer, processing them through a multi-layer fully connected network structure in the hidden layer, and outputting the predicted response results of each analysis node by the output layer; calculating the loss function between the predicted response results and the response results in the neural network training dataset; updating the network weights of the artificial neural network model using the backpropagation algorithm; stopping training when the loss function is less than a preset error threshold, thus obtaining the structural response prediction model.
[0037] Specifically, an artificial neural network model is constructed, comprising an input layer, hidden layers, and an output layer. The input layer has one neuron and receives the S-value from the payload input parameters. The hidden layer consists of three fully connected layers connected in series, with 128, 64, and 32 neurons per layer, respectively. A ReLU activation function is applied to the fully connected hidden layers for nonlinear feature transformation. The output layer has three neurons and receives the S-value from the neural network training dataset. Forward propagation feature extraction is performed through the multi-layer fully connected network structure of the hidden layers. The output layer outputs the predicted response results for each analysis node, including predicted values for stress x, stress y, and stress z. The loss function between the predicted response results and the response results in the neural network training dataset is calculated. The mean squared error formula is used to calculate the average of the squared differences between the predicted and true values. The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the network weights at each layer. The Adam optimizer is used to update the network weights of the artificial neural network model based on the calculated gradients. The initial learning rate is set to 0.001. An adaptive learning rate strategy is used to dynamically adjust the learning rate. The model performance is periodically evaluated on the validation set. When the loss function does not decrease after 20 consecutive iterations, an early stopping mechanism is triggered and training is stopped. The network parameters are saved to obtain the structural response prediction model.
[0038] In this embodiment, the specific steps for inputting the remaining load cycles in the load cycle spectrum data that did not participate in the finite element analysis into the structural response prediction model to obtain the predicted structural response results are as follows: inputting the remaining load cycles in the load cycle spectrum data that did not participate in the finite element analysis into the structural response prediction model one by one, and performing batch neural network forward inference; outputting the predicted response results corresponding to each remaining load cycle; and combining the predicted response results corresponding to each remaining load cycle to generate the predicted structural response results.
[0039] Specifically, all residual load cycles not included in the training load subset are extracted from the load cycle spectrum data and treated as residual load cycles not involved in the finite element analysis. The weight matrices and bias vector parameters of each hidden and output layer in the structural response prediction model are read. The residual load cycles not involved in the finite element analysis are scaled using the same maximum and minimum value range boundaries as the training set, converting them into a standard tensor data format that the structural response prediction model can handle. These tensor data format residual load cycles are then input one by one into the structural response prediction model, and batch neural network forward inference is performed. Matrix inner product multiplication and nonlinear activation function transfer operations are performed between the input layer and multiple hidden layers. During the calculation, all weight matrices and bias vector parameters are locked and gradient backpropagation and parameter updates are not performed. The predicted response results corresponding to each residual load cycle are output. The inverse normalization calculation formula is used to map and restore the predicted response results in the normalization interval to values with physical magnitudes, and the predicted Mises equivalent stress and predicted principal strain values of each analysis node are obtained. The predicted response results corresponding to each residual load cycle are combined, and the time axis is aligned by reading the initial timestamp information recorded in the original time series of the residual load cycle. The sequences are spliced in chronological order to construct a continuous stress-strain time series covering all residual load cycles that did not participate in the finite element analysis, and the predicted structural response results are generated.
[0040] In this embodiment, the steps for verifying the accuracy of the predicted structural response results, integrating the training sample results data and the predicted structural response results, generating a complete spectral load response database, and outputting it are as follows: Verification samples are selected from the remaining load cycles according to error sensitivity; finite element verification analysis is performed on the verification samples to obtain finite element verification results; the predicted response results corresponding to the verification samples in the predicted structural response results are extracted, and the error between the predicted response results and the finite element verification results is calculated; if the error exceeds a preset supplementary threshold, the verification samples are added to the training load subset, and the structural response prediction model is retrained until the accuracy requirements are met; the predicted structural response results that meet the accuracy requirements are merged with the training sample results data to generate a complete spectral load response database and output it.
[0041] Specifically, from the remaining load cycles, verification samples are selected based on error sensitivity. The activation layer partial derivatives of the remaining load cycles not involved in the finite element analysis are calculated during forward propagation of the structural response prediction model. Load cycles with activation layer partial derivative absolute values exceeding twice the average value are selected as high-sensitivity samples. From these high-sensitivity samples, 50 load cycle data are randomly selected as verification samples. Finite element verification analysis is performed on these verification samples. The upper and lower limit loads corresponding to these 50 verification samples are then re-imported into the finite element software as boundary conditions for static calculations, yielding the finite element verification results. The predicted response results corresponding to the verification samples are extracted from the predicted structural response results, and the correlation between the predicted response results and the finite element verification results is calculated. Error is the absolute difference between the predicted stress value and the actual stress value calculated by the finite element method. If the error exceeds a preset supplementary threshold, for example, if the difference is greater than 5% of the yield strength, the accuracy is deemed unsatisfactory. The verification samples with excessive errors are added to the training load subset. Some network weights of the last hidden layer in the structural response prediction model are reset. The structural response prediction model is retrained using the expanded training load subset until the error on all verification samples is less than 5%, meeting the accuracy requirements. The predicted structural response results that meet the accuracy requirements are merged with the training sample results. According to the time sequence number of the original load, the stress and strain data of all analysis nodes are merged into a complete two-dimensional matrix to generate a complete spectral load response database and output it.
[0042] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A fast analysis and calculation method for long-period spectral loads based on artificial neural networks, characterized in that, Includes the following steps: Acquire initial load history data, preprocess the initial load history data, and generate a standardized load time history file; Rainflow counting processing is performed on the standardized load time history file to generate load cyclic spectrum data; Based on the load cycle spectrum data, representative load cycle samples are selected to construct a training load subset; A finite element model of the target structure is established in the finite element software, and finite element analysis is performed based on the training load subset to obtain training sample result data. Extract the response results corresponding to each analysis node from the training sample result data to construct a neural network training dataset; An artificial neural network model is trained based on the neural network training dataset to obtain a structural response prediction model; The remaining loads that did not participate in the finite element analysis from the load cyclic spectrum data are cyclically input into the structural response prediction model for prediction, and the predicted structural response results are obtained. The accuracy of the predicted structural response results is verified, and the training sample result data and the predicted structural response results are integrated to generate a complete spectral response database and output it.
2. The method for fast analysis and calculation of long-period spectral loads based on artificial neural networks according to claim 1, characterized in that, The specific steps for obtaining initial load history data, preprocessing the initial load history data, and generating a standardized load time history file are as follows: Read the initial timestamp information of each line in the initial load history data; By comparing the initial timestamp intervals of two adjacent rows of data in the initial load history data, if the initial timestamp intervals are found to be non-increasing, it is determined that there is an anomaly in the data timing order. A sorting algorithm is used to rearrange the data segments that have experienced the data time sequence disorder anomaly according to the initial timestamp information in ascending order to obtain time sequence corrected data. Check whether the load value field of each record in the time-series correction data is empty. If the load value field has an empty value, it is determined that a partial data missing anomaly has occurred. An interpolation algorithm is used to combine the normal values before and after the missing values to complete the load data; A uniform time step is generated based on a preset sampling frequency; Data segments in the complete load data that do not meet the unified time step requirement are resampled to generate the standardized load time history file.
3. The method for rapid analysis and calculation of long-period spectral loads based on artificial neural networks according to claim 1, characterized in that, The specific steps for performing rainflow counting processing on the standardized load time history file to generate load cyclic spectrum data are as follows: Traverse all time-series load points in the standardized load time history file; Identify local maxima and local minima among all time series load points to construct a peak-valley alternation sequence; The four-point rainflow counting method was used to determine the cyclic closure of the peak-valley alternation sequence, and each closed cycle was obtained. Statistically analyze the load amplitude, average load, and frequency of occurrence for each closed loop. The closed loops are rearranged in chronological order to generate the load cyclic spectrum data.
4. The method for rapid analysis and calculation of long-period spectral loads based on artificial neural networks according to claim 1, characterized in that, The specific steps for selecting representative load cycle samples and constructing a training load subset based on the load cycle spectrum data are as follows: The K-means clustering algorithm is used to perform cluster analysis on the load feature space of each load cycle in the load cycle spectrum data to obtain multiple cluster centers; A predetermined number of load cycles are selected from the vicinity of the multiple cluster centers as representative load cycle samples; The representative cyclic load samples are combined to construct the training load subset.
5. The method for rapid analysis and calculation of long-period spectral loads based on artificial neural networks according to claim 1, characterized in that, The specific steps for establishing a finite element model of the target structure in finite element software and performing finite element analysis based on the training load subset to obtain training sample result data are as follows: Establish a geometric model of the target structure; Set the material properties and boundary conditions of the target structural geometric model; Mesh generation is performed on the geometric model of the target structure to obtain the finite element model of the target structure; Obtain the load amplitude and average load corresponding to each representative load cycle sample in the training load subset, and use them as load input parameters to convert them into finite element load boundary conditions; For each of the representative load cycle samples, the finite element solver is called to perform static analysis, and the response results of each analysis node in the finite element model under each of the representative load cycle samples are extracted. The response results are stored according to the correspondence between the finite element load boundary conditions and the response results to generate the training sample result data.
6. The method for fast analysis and calculation of long-period spectral loads based on artificial neural networks according to claim 1, characterized in that, The specific steps for extracting the response results corresponding to each analysis node from the training sample result data and constructing the neural network training dataset are as follows: Extract the load input parameters corresponding to each representative load cycle sample and the response results corresponding to each analysis node from the training sample result data; Establish a mapping relationship between the load input parameters and the response results of each analysis node; The neural network training dataset is constructed based on the mapping relationship.
7. The method for rapid analysis and calculation of long-period spectral loads based on artificial neural networks according to claim 1, characterized in that, The specific steps for training an artificial neural network model based on the aforementioned neural network training dataset to obtain a structural response prediction model are as follows: Construct an artificial neural network model that includes an input layer, hidden layers, and an output layer; The payload input parameters from the neural network training dataset are input into the input layer, processed by the multi-layer fully connected network structure of the hidden layer, and the predicted response results of each analysis node are output by the output layer. Calculate the loss function between the predicted response and the response in the neural network training dataset; The backpropagation algorithm is used to update the network weights of the artificial neural network model; Training stops when the loss function is less than a preset error threshold, and the structural response prediction model is obtained.
8. The method for rapid analysis and calculation of long-period spectral loads based on artificial neural networks according to claim 1, characterized in that, The specific steps for cyclically inputting the remaining loads from the load cyclic spectrum data that did not participate in the finite element analysis into the structural response prediction model to obtain the predicted structural response are as follows: The remaining load cycles that did not participate in the finite element analysis in the load cycle spectrum data are input one by one into the structural response prediction model, and batch neural network forward inference is performed. Output the predicted response results for each of the remaining load cycles; The predicted structural response results are generated by combining the predicted response results corresponding to each remaining load cycle.
9. The method for rapid analysis and calculation of long-period spectral loads based on artificial neural networks according to claim 1, characterized in that, The specific steps for verifying the accuracy of the predicted structural response results, integrating the training sample results data and the predicted structural response results, generating a complete spectral response database, and outputting it are as follows: Select verification samples from the remaining load cycles based on error sensitivity; A finite element verification analysis was performed on the verification sample to obtain the finite element verification results; Extract the predicted response result corresponding to the verification sample from the predicted structural response result, and calculate the error between the predicted response result and the finite element verification result; If the error exceeds the preset supplementation threshold, the verification sample is added to the training load subset, and the structural response prediction model is retrained until the accuracy requirements are met. The predicted structural response results that meet the accuracy requirements are merged with the training sample result data to generate the complete spectral response database and output it.
10. A fast analysis and calculation system for long-period spectral carriers based on artificial neural networks, used to implement the fast analysis and calculation method for long-period spectral carriers based on artificial neural networks as described in any one of claims 1-9, characterized in that, include: The data preprocessing module is used to acquire initial load historical data, preprocess the initial load historical data, and generate a standardized load time history file. The rainflow counting module is used to perform rainflow counting processing on the standardized load time history file to generate load cyclic spectrum data. The sample selection module is used to select representative load cycle samples based on the load cycle spectrum data and construct a training load subset. The finite element analysis module is used to build a finite element model of the target structure in the finite element software, and to perform finite element analysis based on the training load subset to obtain training sample result data. The data construction module is used to extract the response results corresponding to each analysis node in the training sample result data and construct the neural network training dataset; The neural network training module is used to train an artificial neural network model based on the neural network training dataset to obtain a structural response prediction model. The batch prediction module is used to cyclically input the remaining loads in the load cyclic spectrum data that did not participate in the finite element analysis into the structural response prediction model for prediction, and obtain the predicted structural response results. The result output module is used to verify the accuracy of the predicted structural response results, integrate the training sample result data and the predicted structural response results, generate a complete spectral response database and output it.