Structural global dynamic response prediction method and related equipment

Abstract

This invention provides a method and related equipment for predicting the global dynamic response of a structure, comprising: acquiring a real-world dataset of the structure in its current state; processing multiple sparse sensor data based on a preset finite element model to output a simulation dataset, the simulation dataset including multiple pairs of sparse input-global output data; pre-training a blank neural network model based on the simulation dataset to obtain a virtual model; performing physical correction processing on the virtual model to obtain a real-world model; processing the sensor data of the current structure using the real-world model to determine the global dynamic response field data of the current structure; and predicting the risk deformation of the current structure based on the global dynamic response field data and executing corresponding risk warning operations. Through the above method steps, global dynamic response prediction of large and complex structures can be achieved with only a small amount of sensor data, reducing sensor deployment costs and improving the accuracy and precision of structural risk monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent testing, and in particular to a method, apparatus, electronic device and storage medium for predicting the dynamic response of a structure across its entire domain. Background Technology

[0002] Currently, health assessments of large-scale infrastructure primarily rely on two independent technological paradigms: direct monitoring based on physical sensors and simulation analysis based on numerical models. However, both paradigms suffer from fundamental and insurmountable limitations, preventing the achievement of comprehensive, accurate, and real-time assessments of structural condition.

[0003] In existing technologies, large-scale sensor arrays are typically deployed on structures to acquire response data under different operating conditions for damage identification and risk assessment. However, in complex structures such as bridges, tunnels, and large buildings, the number of sensors that can be deployed is limited by cost, energy consumption, and maintenance conditions, often resulting in only a limited amount of sparse sensor data being acquired in practical applications. This makes it difficult for monitoring data to comprehensively reflect the structure's full-domain response, severely restricting the accuracy and reliability of structural risk prediction and early warning.

[0004] On the other hand, the finite element model, as a high-fidelity numerical simulation method, can theoretically output the dynamic response field across the entire domain. However, this type of model is highly dependent on idealized input parameters and boundary conditions, resulting in a "simulation-reality gap" between the results and the actual structure. Traditional finite element model update methods often require a large amount of high-precision measured data for back-correction, which contradicts the reality of sparse sensor data, making it difficult to meet the dual requirements of real-time performance and accuracy in practical engineering.

[0005] Therefore, existing methods for predicting the global dynamic response of structures cannot accurately reconstruct the global dynamic response of structures and achieve real-time risk warning when relying solely on sparse sensor data. Summary of the Invention

[0006] This invention provides a method for predicting the global dynamic response of a structure, in order to solve the problem that existing methods for predicting the global dynamic response of a structure cannot accurately reconstruct the global dynamic response of a structure and achieve real-time risk warning when relying solely on sparse sensor data.

[0007] In a first aspect, the present invention provides a method for predicting the dynamic response of a structure across its entire domain, the method comprising the following steps:

[0008] Obtain the real-world dataset of the structure in the current state, wherein the real-world dataset contains multiple sparse sensor data corresponding to the current structure;

[0009] Based on a preset finite element model, the multiple sparse sensing data are processed to output a simulation dataset, which includes multiple pairs of sparse input-global output data pairs.

[0010] Based on the simulation dataset, the blank neural network model is pre-trained to obtain a virtual model;

[0011] The virtual model is then subjected to physical correction processing to obtain the real model;

[0012] By processing the sensing data of the current structure using a realistic model, the full-domain dynamic response field data of the current structure can be determined.

[0013] Based on the global dynamic response field data, the risk deformation of the current structure is predicted, and the corresponding risk warning operation is executed.

[0014] Optionally, obtaining the real-world dataset of the structure in the current state includes:

[0015] The preset sensor array is deployed at vulnerable locations in the structure, and the preset sensor array includes an acceleration sensor, a strain sensor, a displacement sensor, and an environmental sensor.

[0016] The sparse sensing data of the vulnerable locations of the structure in the current state are collected by the preset sensor array to obtain the actual dataset of the structure in the current state. The sparse sensing data corresponding to the vulnerable locations are used to maximize the information reflecting the overall state of the structure.

[0017] Optionally, the step of processing the multiple sparse sensing data based on a preset finite element model to output a simulation dataset includes:

[0018] Determine the virtual load case data required for the current simulation;

[0019] The virtual load condition data is loaded into the preset finite element model, and the multiple sparse sensor data are processed by the finite element algorithm to determine multiple corresponding finite element response outputs.

[0020] Multiple sparse sensing data are paired with multiple corresponding finite element response outputs to obtain multiple sparse input-global output data pairs;

[0021] A simulation dataset is constructed based on multiple sparse input-global output data pairs.

[0022] Optionally, the step of pre-training the blank neural network model based on the simulation dataset to obtain a virtual model includes:

[0023] The sparse sensing data in the simulation dataset is input into the blank neural network model to obtain the training global output;

[0024] The finite element response output corresponding to the sparse sensing data is used as a supervision label. The training global output is compared with the finite element response output according to the time-frequency domain contrast regularization fidelity loss function to obtain the calculation result. The time-frequency domain contrast regularization fidelity loss function is constructed based on the time-domain morphological fidelity data, frequency-domain robustness feature data and multi-scale physical consistency data corresponding to the sparse sensing data of the current structure.

[0025] With minimizing the computational result as the optimization objective, the blank neural network model is iteratively trained, and a virtual model is obtained after the iterative training is completed.

[0026] Optionally, the step of performing physical correction processing on the virtual model to obtain the real model includes:

[0027] Acquire sparse sensing data of the current structure;

[0028] Based on the sparse sensing data of the current structure, the output layer and correction layer of the virtual model are adjusted and trained to determine the correction parameters corresponding to the output layer and correction layer of the virtual model.

[0029] Based on the correction parameters, the virtual model is corrected and trained. After the training converges, a real model is obtained. The real model is used to output physical characteristic data reflecting the current structure.

[0030] Optionally, the step of processing the sensing data of the current structure using a real-world model to determine the global dynamic response field data of the current structure includes:

[0031] The sparse sensing data of the current real-time structure is input into the reality model for nonlinear mapping inference to determine the global dynamic response field data corresponding to the current real-time structure. The global dynamic response field data is used to reflect the dynamic response distribution of the current structure at different locations and at different times.

[0032] Optionally, the step of predicting the risk deformation of the current structure based on the global dynamic response field data and executing the corresponding risk warning operation includes:

[0033] The temporal and spatial characteristics of the global dynamic response field data are calculated to determine the deformation trend of the current structure at different locations.

[0034] The deformation trend is compared with the preset risk trend to identify potential deformation areas;

[0035] Based on the potential deformation area, corresponding risk warning information is generated, and corresponding risk warning operations are triggered.

[0036] Secondly, the present invention also provides a structure-wide dynamic response prediction device, the structure-wide dynamic response prediction device comprising:

[0037] The first acquisition module is used to acquire the real dataset of the structure in the current state, wherein the real dataset contains multiple sparse sensing data corresponding to the current structure;

[0038] The first processing module is used to process the multiple sparse sensing data based on a preset finite element model and output a simulation dataset, which includes multiple pairs of sparse input-global output data pairs.

[0039] The first training module is used to pre-train the blank neural network model based on the simulation dataset to obtain a virtual model;

[0040] The second processing module is used to perform physical correction processing on the virtual model to obtain the real model;

[0041] The first determining module is used to process the sensing data of the current structure through a real-world model to determine the global dynamic response field data of the current structure.

[0042] The first execution module is used to predict the risk deformation of the current structure based on the global dynamic response field data, and to execute the corresponding risk warning operation.

[0043] Thirdly, the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the structural global dynamic response prediction method provided by the present invention.

[0044] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the structure-wide dynamic response prediction method provided by the invention.

[0045] This invention acquires a real-world dataset of the structure in its current state; based on a pre-defined finite element model, it processes multiple sparse sensor data to output a simulation dataset, which includes multiple pairs of sparse input-global output data; based on the simulation dataset, it pre-trains a blank neural network model to obtain a virtual model; it performs physical correction processing on the virtual model to obtain a real-world model; it processes the sensor data of the current structure using the real-world model to determine the global dynamic response field data of the current structure; based on the global dynamic response field data, it predicts the risk deformation of the current structure and executes corresponding risk warning operations. Through these steps, it is possible to predict the global dynamic response of large and complex structures using only a small amount of sensor data, reducing sensor deployment costs and improving the accuracy and precision of structural risk monitoring. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of a structural global dynamic response prediction method provided in an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of another structural global dynamic response prediction device provided in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] like Figure 1 As shown, Figure 1 This is a flowchart of a structural global dynamic response prediction method provided by an embodiment of the present invention. The structural global dynamic response prediction method includes the following steps:

[0052] 101. Obtain the real-world dataset of the structure in the current state.

[0053] In this embodiment of the invention, the above-mentioned structural global dynamic response prediction method can be applied to a structural global dynamic response prediction platform. The structural global dynamic response prediction platform has functions such as global dynamic response prediction data processing, global dynamic response prediction data transmission and reception, and global dynamic response prediction data memory storage. It can be built based on a server or server cluster. The server or server cluster can be an electronic device with global dynamic response prediction data processing capabilities.

[0054] The aforementioned current state can refer to the physical state of a structure at a specific moment or under specific operating conditions. It can be understood that this state can be determined by the external environmental loads on which the structure is located, such as vibration and wind, as well as the internal service conditions, such as aging and temperature effects. For example, the stress condition of a bridge under traffic load during peak hours, and the dynamic response of a building under strong winds, can both be regarded as the current state of a structure.

[0055] The aforementioned structures can refer to engineering objects that require full-domain dynamic response monitoring and prediction, including typical civil engineering infrastructure such as bridges, tunnels, large buildings, and offshore platforms.

[0056] The aforementioned real-world dataset may include, but is not limited to, multiple sparse sensor data corresponding to the current structure. Generally, it can be a set of actual monitoring data collected by sensors deployed on the structure, including the outputs of various sensors, such as vibration signals output by accelerometers, local strain values ​​collected by strain gauges, structural displacements collected by displacement sensors, and environmental parameters such as temperature, humidity, and wind speed. It can be understood that the data in the real-world dataset all come from specific physical measurements and can reflect the dynamic response of the structure and external environmental conditions.

[0057] The aforementioned sparse sensing data refers to data extracted from the real-world dataset and obtained by deploying a limited number of sensors. This data, collected by deploying sensors at key points throughout the structure, covers only a portion of the structure's nodes. For example, on a long-span bridge, data obtained by deploying accelerometers and strain gauges only at the mid-span, quarter-span, and tower top constitutes sparse sensing data. Therefore, although the quantity of sparse sensing data is limited, its content can be used to represent key information representing the overall dynamic characteristics of the structure.

[0058] In one possible embodiment, the above-mentioned structural global dynamic response prediction platform collects real-world datasets by using a limited number of sensors deployed at key locations when the structure is in a specific operating condition.

[0059] 102. Based on the preset finite element model, process multiple sparse sensing data and output a simulation dataset.

[0060] In this embodiment of the invention, the aforementioned preset finite element model can refer to the high-fidelity numerical model established by the above-mentioned structural global dynamic response prediction platform based on the current structural design drawings, material parameters and actual boundary conditions. The aforementioned preset finite element model can simulate the dynamic response of the structure under virtual working conditions such as traffic load, wind load or earthquake action through numerical calculation. That is, it calculates based on a small amount of sparse sensor data, outputs a large amount of corresponding dynamic response data and uses it as simulation data. This process can serve as the basis for generating simulation data.

[0061] In one possible embodiment, the above-mentioned structure global dynamic response prediction platform loads the set virtual load conditions onto the preset finite element model, extracts response data at the nodes corresponding to the sparse sensing data locations, and generates dynamic response outputs across the entire domain. This allows a correspondence to be established between a limited number of sensor point data and the global response results, forming a pairing of sparse inputs and global outputs.

[0062] The aforementioned simulation dataset may include, but is not limited to, multiple pairs of sparse input-global output data pairs. Specifically, each set of data samples contains the temporal response of sparse sensing nodes and its corresponding global dynamic response field. These samples are repeatedly generated under various virtual load conditions, which can construct multiple sets of samples covering different states, thereby providing a source of training data for the pre-training of neural network models.

[0063] In another possible embodiment, the above-mentioned structural global dynamic response prediction platform applies virtual load conditions to the finite element model and extracts response results at the nodes corresponding to the sensor deployment locations. At the same time, it outputs a dynamic response field covering the entire domain. By pairing sparse inputs with global outputs, multiple sets of data samples are formed, thereby constructing a simulation dataset to provide training data for the pre-training of subsequent neural network models.

[0064] 103. Based on the simulation dataset, the blank neural network model is pre-trained to obtain a virtual model.

[0065] In this embodiment of the invention, the above-mentioned blank neural network model can refer to a neural network structure that has not been trained with any data before pre-training, and its parameters are all in the initial state. Specifically, the above-mentioned blank neural network model can adopt a spatiotemporal neural network structure (such as spatiotemporal U-Net) and has the ability to extract features and perform nonlinear mapping on the input sparse data. It is understood that in the initial state, that is, before training, the above-mentioned blank neural network model does not have an effective structure-wide response prediction capability.

[0066] In this embodiment, a blank neural network model can be initially trained based on a simulation dataset generated from a finite element model. Sparse inputs are used as input layer data of the model, and global response outputs are used as supervision labels. By iteratively optimizing the loss function, the blank neural network model gradually learns the mapping relationship between sparse inputs and global outputs. It can be understood that the purpose of the above pre-training process is to allow the blank neural network model to learn universal physical laws in a virtual environment, thereby laying the foundation for subsequent physical correction based on real data.

[0067] The aforementioned supervision labels can also be complete global dynamic response outputs, used to adjust the model's learning of universal physical laws.

[0068] The aforementioned virtual model can refer to a neural network model obtained through pre-training. This virtual model is already able to predict the corresponding global dynamic response field based on the aforementioned sparse sensing data under virtual working conditions, and has a certain predictive ability and generalization. However, the aforementioned virtual model reflects the physical laws contained in the finite element simulation data, and there is still a simulation-reality gap that differs from the actual structure. It needs to be further adjusted to a realistic model through physical correction in order to make up for the difference between it and the real structure in the global response.

[0069] In one possible embodiment, the above-mentioned global dynamic response prediction platform uses the sparse input in the simulation dataset as the input to the model and the corresponding global dynamic response output as the supervision label. It uses a set loss function to iteratively train the model. During this training process, the model gradually learns the complex nonlinear mapping relationship between the sparse input and the global output until it converges and has generalization ability, and finally forms a virtual model that can characterize the dynamic features of the structure.

[0070] 104. Perform physical correction processing on the virtual model to obtain the real model.

[0071] In this embodiment of the invention, the virtual model can be corrected using real-collected sparse sensor data. Specifically, by introducing a small amount of real data into the output layer and correction layer of the virtual model for fine-tuning training, the model can gradually eliminate the difference between finite element simulation and actual structure, bridging the "simulation-reality gap". This avoids the large-scale data requirement required to train a neural network from scratch, and the model correction can be completed using only limited measured data.

[0072] The aforementioned realistic model can refer to a neural network model formed after physical correction. While maintaining the universal physical laws learned in the virtual model, this realistic model also incorporates the actual response characteristics of the structure, enabling it to accurately predict the full-domain dynamic response field of the structure under real-world operating conditions. It can be understood that this realistic model can output predictive results consistent with the actual dynamic behavior of the structure, even with real-time input of sparse sensor data, providing a reliable basis for structural health monitoring and risk warning.

[0073] In one possible embodiment, the aforementioned structure-wide dynamic response prediction platform collects a small amount of real sparse sensing data on the structure and uses it as input data to adjust and train the output and correction layers of the virtual model. By iteratively updating some network parameters, the model can compensate for the differences between the virtual model and the real structure. After training convergence, a realistic model that reflects the true physical characteristics of the target structure is obtained. This model can output a global dynamic response field that conforms to the actual situation when receiving real-time sparse sensing data.

[0074] 105. Process the sensing data of the current structure using a real-world model to determine the global dynamic response field data of the current structure.

[0075] In this embodiment of the invention, the above-mentioned global dynamic response field data can refer to the dynamic response results of the structure in the entire space under the current state, including but not limited to the temporal distribution of displacement field, strain field, stress field or acceleration field, etc. It should be noted that the above-mentioned global dynamic response field data is different from sparse sensing data that can only reflect local monitoring points. The above-mentioned global dynamic response field data can show the stress and deformation characteristics of the structure as a whole, thereby comprehensively reflecting the dynamic behavior of the structure under specific working conditions.

[0076] For example, during the operation of a long-span bridge, the dynamic response field data generated by a real-world model can demonstrate the following dynamic behavioral characteristics of the structure under specific working conditions:

[0077] Stress characteristics: When heavy vehicles pass through the mid-span of the bridge, the displacement and stress in the mid-span area increase significantly, while the top of the bridge tower and the anchorage area show the corresponding stress transmission path. This global distribution can reflect the overall transmission of vehicle load in the structure.

[0078] Deformation characteristics: Under strong wind conditions, the global dynamic response field data can depict the overall vibration mode of the bridge along the span direction, such as lateral swaying or vertical vibration, and show the deformation trend of the bridge deck and towers at different time points, thus reflecting the global dynamic behavior of the structure under wind load.

[0079] With this data, managers can not only observe the response of individual sensor points, but also obtain the stress distribution and dynamic deformation process of the entire bridge under specific working conditions, thereby making a more comprehensive judgment on potential risk areas.

[0080] In one possible embodiment, the above-mentioned structure-wide dynamic response prediction platform inputs the sparse sensor data collected in real time into the reality model. The model performs inference calculations on the input data based on the corrected network parameters to generate full-domain dynamic response field data covering the entire structure, reflecting the dynamic response distribution of the structure at different locations and times, for subsequent risk identification and early warning operations.

[0081] 106. Based on the dynamic response field data of the entire domain, predict the risk deformation of the current structure and execute the corresponding risk warning operation.

[0082] In this embodiment of the invention, the aforementioned structural global dynamic response prediction platform analyzes and judges the abnormal deformation trends that may occur in the structure over a period of time or under specific working conditions based on global dynamic response field data. Specifically, it can model the evolution of displacement, strain, or stress at different monitoring points of the structure and, in conjunction with preset risk thresholds, identify deformation quantities or growth rates that may lead to structural safety hazards. For example, when the deflection in a certain mid-span region of a bridge increases rapidly over time and approaches the design limit, a potential risk in that region will be predicted.

[0083] The aforementioned risk warning operation refers to the response measures automatically generated and issued after a potential risk deformation is predicted. This operation may include generating visual warning information (such as highlighting the risk area on a 3D structural model), triggering audible and visual alarm devices to alert on-site personnel, or transmitting warning commands to the structural health management system via a data interface, thereby achieving automated and coordinated control.

[0084] In one possible embodiment, the aforementioned structural global dynamic response prediction platform performs temporal and spatial feature analysis on the global dynamic response field data, extracts deformation trends at different locations, and compares them with preset risk reference values ​​to identify areas that may exceed limits. After identifying potential risk deformation areas, it generates risk warning information and executes risk warning operations through audible and visual alarms, visual prompts, or linkage with the management system, so that maintenance personnel can take timely measures.

[0085] In this embodiment of the invention, a real-world dataset of the structure in its current state is acquired; based on a preset finite element model, multiple sparse sensor data are processed to output a simulation dataset, which includes multiple pairs of sparse input-global output data; based on the simulation dataset, a blank neural network model is pre-trained to obtain a virtual model; the virtual model is physically corrected to obtain a real-world model; the sensor data of the current structure are processed using the real-world model to determine the global dynamic response field data of the current structure; based on the global dynamic response field data, the risk deformation of the current structure is predicted, and corresponding risk warning operations are executed. Through the above method steps, global dynamic response prediction of large and complex structures can be achieved with only a small amount of sensor data, reducing sensor deployment costs and improving the accuracy and precision of structural risk monitoring.

[0086] Optionally, in the step of obtaining the real dataset of the structure in the current state, a preset sensor array can be deployed at the vulnerable locations in the structure; the sparse sensing data of the vulnerable locations of the structure in the current state can be collected by the preset sensor array to obtain the real dataset of the structure in the current state.

[0087] In this embodiment of the invention, the aforementioned preset sensor array may include, but is not limited to, accelerometers, strain sensors, displacement sensors, and environmental sensors. For example, it may be an accelerometer that captures high-frequency vibrations, a strain gauge that monitors local stress, a GNSS displacement sensor that acquires macroscopic displacement, and sensors that collect environmental parameters such as wind speed and temperature, etc., which are sensor devices capable of acquiring the physical state of the structure. Specifically, the sensors in the aforementioned preset sensor array can be arranged according to structural mechanics analysis and optimal sensor layout methods, thereby maximizing the amount of information with a limited number of sensors.

[0088] The aforementioned vulnerable locations can refer to the critical areas most prone to damage or anomalies, as determined by the structural stress characteristics and long-term operational experience. For example, in large bridges, typical vulnerable locations include the mid-span area, the connection between the beam end and the support, and the top of the bridge tower. These locations often bear large dynamic loads or environmental loads and are identified as potential weak points in finite element simulation analysis and engineering experience. Therefore, by deploying sensors at these locations, it is possible to capture the potential risk response of the structure earlier and more directly, thereby improving the accuracy of reconstructing the global dynamic response field based on sparse data. In other words, the sparse sensing data corresponding to the aforementioned vulnerable locations can be used to maximize the information reflecting the overall state of the structure.

[0089] In one possible embodiment, the above-mentioned structural global dynamic response prediction platform collects sparse sensing data of the above-mentioned vulnerable locations under the current working conditions through the above-mentioned preset sensor array to form a real dataset. It can be understood that since the sensor deployment locations correspond to the potential weak points of the structure, the obtained data can more directly reflect the key response characteristics of the structure.

[0090] Optionally, the step of processing multiple sparse sensor data based on a preset finite element model and outputting a simulation dataset further includes: determining the virtual load case data required for the current simulation; loading the virtual load case data into the preset finite element model and processing the multiple sparse sensor data using a finite element algorithm to determine multiple corresponding finite element response outputs; pairing the multiple sparse sensor data with the corresponding multiple finite element response outputs to obtain multiple sparse input-global output data pairs; and constructing a simulation dataset based on the multiple sparse input-global output data pairs.

[0091] In this embodiment of the invention, the aforementioned virtual load condition data can refer to a combination of loads and boundary conditions artificially set in finite element simulation, including various load parameters such as traffic load, wind load, ground motion, and temperature change. It can be understood that the aforementioned virtual load condition data is applied to the finite element model in a virtual manner to simulate the dynamic response of the structure under different environments and actions.

[0092] The above finite element response output can be the structural response result obtained by solving the above finite element algorithm, including but not limited to time series or spatial distribution data such as nodal displacement, strain, stress, and acceleration.

[0093] In this embodiment, a correspondence can be established between the sparse sensing data and the finite element global response results obtained under the same working conditions. Specifically, a small number of sparse inputs can be matched one-to-one with the output results covering the entire structure to form training samples that meet the requirements of supervised learning.

[0094] In one possible embodiment, the above-mentioned structural global dynamic response prediction platform loads the virtual load condition data into a preset finite element model, and extracts the response output at the position corresponding to the sparse sensing data using the finite element algorithm to obtain multiple finite element response results. The sparse sensing data is paired with the corresponding finite element response output to form multiple sparse input-global output data pairs, and further, a simulation dataset is constructed based on these data pairs to provide training samples for the pre-training of the neural network model.

[0095] Optionally, in the step of pre-training the blank neural network model based on the simulation dataset to obtain the virtual model, the method further includes inputting sparse sensing data from the simulation dataset into the blank neural network model to obtain the training global output; using the finite element response output corresponding to the sparse sensing data as a supervision label, comparing and calculating the training global output with the finite element response output according to the time-frequency domain comparison regularization fidelity loss function to obtain the calculation result; and iteratively training the blank neural network model with the optimization objective of minimizing the calculation result, thereby obtaining the virtual model after the iterative training is completed.

[0096] In this embodiment of the invention, the above-mentioned training global output can refer to the global dynamic response prediction result covering the entire structure generated by the blank neural network model through network inference after receiving sparse sensing data as input during the pre-training process.

[0097] The aforementioned supervisory label can refer to the global dynamic response output generated by finite element simulation, which serves as a reference true value in neural network training.

[0098] The aforementioned time-frequency domain contrast regularization fidelity loss function can refer to the simultaneous introduction of contrast constraints in the time and frequency domains into the loss function. It combines dynamic time warping (soft-DTW) to ensure the temporal morphological fidelity of the predicted sequence, combines frequency domain feature contrast to enhance the robustness of the model in the spectral space, and maintains physical rationality through power spectral density consistency constraints. Specifically, the aforementioned time-frequency domain contrast regularization fidelity loss function can also be constructed based on the temporal morphological fidelity data, frequency domain robustness feature data, and multi-scale physical consistency data corresponding to the sparse sensing data of the current structure.

[0099] In this embodiment, the model can be guided to perform deeper, physically intuitive learning by combining three complementary sub-loss terms. The specific overall structure is as follows:

[0100] ;

[0101] in, To preserve fidelity, For frequency domain contrast loss, For multi-scale PSD consistency loss, and It is used as a hyperparameter to balance the weights of the various loss terms.

[0102] More specifically, the aforementioned fidelity loss It can be used to pursue the accuracy of structure in terms of morphology and time sequence. Differentiable dynamic time warping (soft-DTW) is used as the core metric. Unlike Euclidean distance, which is extremely sensitive to time shifts, DTW can find the optimal nonlinear alignment path between two time series, thereby measuring their morphological similarity. Specifically, the fidelity loss mentioned above can be constructed using the following formula:

[0103]

[0104]

[0105]

[0106] Among them, morphological loss Timing loss calculated directly by soft-DTW This penalizes excessive distortion of the alignment path. and These correspond to the actual sequence and the predicted sequence, respectively. The optimal alignment path was found by the DTW algorithm. It is a path Index pairs in For the true sequence Time index, For the predicted sequence Zhongyu Aligned time index, It is a hyperparameter used to balance the importance of form and timing.

[0107] The above frequency domain contrast loss This can be used to learn robust physical representations, employing a "narrowing of similar and pushing away of dissimilar" approach to learn robust representations invariant to data augmentation. The frequency domain representation is used to address noise and small time shifts. The above frequency domain contrast loss is constructed through the following steps:

[0108] 1. First, Transform to the frequency domain using real-number Fast Fourier Transform (RFFT). ;

[0109] 2. Construct positive and negative samples:

[0110] In this context, a positive sample can be represented as two slightly different augmented samples formed by applying two independent data augmentation operations (e.g., adding Gaussian noise, applying a random time-domain mask, or performing small-scale amplitude scaling and jittering) to the same original input within a batch. These two augmented samples are then fed into the model to obtain their frequency domain representations. and This forms a positive sample pair;

[0111] Negative samples can be represented as predictions from different original inputs in a batch, with their frequency domain representation... All of them are considered negative samples.

[0112] 3. Using the NT-Xent (Normalized Temperature-scaled Cross-Entropy) loss function, the similarity of positive sample pairs in the frequency domain representation space is maximized while minimizing the similarity with negative samples. This enables the model to learn stable frequency domain features that can distinguish different physical events, greatly enhancing the model's robustness to real-world noise. Specifically:

[0113]

[0114] in, Represents batch size. It is cosine similarity. This is a temperature hyperparameter used to adjust the sharpness of the similarity distribution; a smaller value is preferable. This will widen the difference in similarity scores between different samples, making the model focus more on distinguishing the negative samples that are difficult to differentiate. , and These represent the frequency domain representation vectors of the samples after the RFFT transform. and They come from the same original input and form positive sample pairs, while This means that the batch contains... Non-homologous negative samples.

[0115] The above-mentioned multi-scale power spectral density uniformity loss It can be used to ensure the physical consistency of multiple scales. By ensuring that the PSD of the signal remains consistent with the PSD of the real signal at both the original resolution and the downsampled low resolution, it can not only capture high-frequency vibration details (modal characteristics) but also accurately reproduce long-term macroscopic trends (thermal effects, etc.). Specifically, the above-mentioned multi-scale power spectral density consistency loss can be constructed using the following formula. :

[0116]

[0117] in,

[0118]

[0119] For downsampling operations, the sequence length is halved using average pooling. These are hyperparameters used to balance the importance of different scales. and Corresponding sequences and The Middle The values ​​at each point in time.

[0120] By combining the three losses mentioned above to construct the final time-frequency domain contrast-regularized fidelity loss function, the physical induction bias problem of ill-posed structures under specific operating conditions can be solved based on their synergistic effect. Through the organic synergy of the three losses, a set of methods for solving the "physical induction bias" of structures under specific operating conditions is provided. Specifically, through... Ensure the solution has a consistent form across the known data points; through Utilizing the frequency domain information of these sparse points provides a global physical constraint for the reconstruction of the entire unknown region, forcing the intrinsic physical properties of the solution to be robust and consistent; through This ensures that the energy distribution of the solution in the frequency domain conforms to physical laws at both macroscopic and microscopic scales. The combination of these three factors greatly compresses the possible solution space and jointly guides the model to learn a physically consistent and reasonable solution that has a high generalization ability to real-world data.

[0121] In this embodiment, the difference between the training global output and the supervision label can be calculated in both the time and frequency domains by using a loss function, thereby reflecting the gap between the model prediction and the target true value.

[0122] The above calculation results can refer to the error or difference values ​​obtained after comparison calculations. They are usually expressed in numerical form and serve as the basis for training optimization. It can be understood that the smaller the above calculation results are, the closer the model's predicted output is to the true global response.

[0123] The parameters of a neural network can be continuously updated using the calculation results in multiple training cycles. In each iteration, the network parameters are adjusted based on error feedback, thereby gradually improving the model performance and achieving the goal of iterative training.

[0124] In one possible embodiment, the above-mentioned structure-wide dynamic response prediction platform inputs sparse sensing data from the simulation dataset into a blank neural network model to generate a training global output. The finite element response output corresponding to the sparse sensing data is used as a supervision label. Based on the time-frequency domain contrast regularization fidelity loss function, the training global output is compared and calculated with the supervision label. The optimization objective is to minimize the calculation result. Iterative training is performed on the blank neural network model. After training convergence, a virtual model that can represent the mapping relationship between sparse input and global output is obtained.

[0125] Optionally, the step of performing physical correction processing on the virtual model to obtain the real model also includes acquiring sparse sensing data of the current structure; adjusting and training the output layer and correction layer of the virtual model based on the sparse sensing data of the current structure to determine the correction parameters corresponding to the output layer and correction layer of the virtual model; and performing correction training on the virtual model based on the correction parameters to obtain the real model after the training converges.

[0126] In this embodiment of the invention, the above-mentioned output layer can refer to the layer in the above-mentioned neural network structure used to generate the final prediction result. It is understood that the above-mentioned output layer is used to transform the intermediate feature mapping into global dynamic response field data. Therefore, adjusting the above-mentioned output layer can directly correct the final prediction accuracy of the model, making its output closer to the physical performance of the real structure.

[0127] The aforementioned correction layer can refer to the layer set in the virtual model for subsequent physical correction. It can generally be set before the output layer. It can be understood that by updating a small number of parameters, the difference between the virtual model and the actual structure can be compensated, thereby realizing the transition from simulation to reality. It can also reduce the need for large-scale parameter updates, thereby improving training efficiency and reducing dependence on real data.

[0128] In this embodiment, the above-mentioned structure-wide dynamic response prediction platform can use real-collected sparse sensing data as input and use supervised learning to update only the parameters of the output layer and the correction layer, thereby achieving the purpose of adjusting the training. It should be noted that, unlike full model training, the adjustment training only optimizes a limited number of parameters, so the training efficiency is faster and it can avoid destroying the universal physical laws that the virtual model has learned.

[0129] The aforementioned correction parameters can refer to the updated model parameters such as weights and biases in the output layer and correction layer during the training process. The correction values ​​of the corresponding layers can be determined through continuous iterative optimization to correct the prediction bias of the virtual model, so that the model output can be closer to the physical characteristics corresponding to the real sensing data.

[0130] In one possible embodiment, the above-mentioned structure-wide dynamic response prediction platform keeps the main parameters of the virtual model unchanged and optimizes only the output layer and the correction layer, so that the model gradually eliminates the difference between virtual simulation and real data.

[0131] In another possible embodiment, the above-mentioned global dynamic response prediction platform acquires sparse sensing data of the current structure and inputs it into a virtual model. Based on this real-world data, the output layer and correction layer of the virtual model are adjusted and trained to determine the correction parameters related to the layer. The virtual model is then corrected and trained based on the correction parameters so that its prediction results gradually approach the dynamic response of the real structure. After the training converges, a real model is obtained. The real model can be used to output data reflecting the physical characteristics of the current structure, as well as subsequent global dynamic response prediction and risk warning.

[0132] Optionally, in the step of processing the sensing data of the current structure through the real-world model to determine the global dynamic response field data of the current structure, the method may further include inputting the sparse sensing data of the current real-time structure into the real-world model for nonlinear mapping inference to determine the global dynamic response field data corresponding to the current real-time structure.

[0133] In this embodiment of the invention, after receiving sparse sensing data from the global dynamic response prediction platform, the aforementioned real-world model uses nonlinear activation functions, convolution operations, and spatiotemporal feature extraction mechanisms in the model network to map the sparse sensing data into a global dynamic response field. In this process, the real-world model no longer relies on simple linear interpolation, but instead uses learned complex nonlinear functional relationships to spatially expand and temporally reconstruct the finite node information. Through this reasoning method, the real-world model can bridge the gap between sparse observations and the global response, ensuring that the prediction results conform to physical laws and possess high temporal fidelity and frequency domain robustness. For example, when a bridge has only a limited number of sensors installed at mid-span and tower top, nonlinear mapping inference can reconstruct the global vibration modes and displacement distribution of the entire bridge at different time points.

[0134] The aforementioned global dynamic response field data can also be used to reflect the dynamic response distribution of the current structure at different locations and times.

[0135] In one possible embodiment, the above-mentioned structure global dynamic response prediction platform inputs the sparse sensing data of the current real-time structure into a physically calibrated reality model. The model performs nonlinear mapping inference based on the parameters obtained from its internal training, generating global dynamic response field data covering the entire structure from the finite sparse input.

[0136] Optionally, the steps of predicting the risk deformation of the current structure based on the global dynamic response field data and executing the corresponding risk warning operation may further include performing temporal and spatial feature calculations on the global dynamic response field data to determine the deformation trend of the current structure at different locations; comparing the deformation trend with the preset risk trend to identify potential deformation areas; generating corresponding risk warning information based on the potential deformation areas and triggering the execution of the corresponding risk warning operation.

[0137] In this embodiment of the invention, the above-mentioned structural global dynamic response prediction platform performs temporal and spatial feature calculations on the global dynamic response field data, extracts the deformation trend of the structure at different monitoring locations over time, compares the obtained deformation trend with the preset risk trend, and identifies the location as a potential deformation area when the deformation or evolution rate at a certain location exceeds the risk reference range. Based on the potential deformation area, corresponding risk warning information is generated, and the risk warning operation is triggered through audible and visual alarms, visual interface prompts, or linkage with the management system, so that maintenance personnel can obtain structural risk information in a timely manner.

[0138] For example, during the operation of a long-span suspension bridge, the structural global dynamic response prediction platform, based on global dynamic response field data generated from a real-world model, performs temporal and spatial analysis of the bridge's stress and deformation characteristics under traffic loads. It discovers that the displacement trend in the mid-span region of the bridge increases rapidly within a short period, and compared to the preset risk trend, the growth rate is approaching the safety threshold. By comparing the results, the platform identifies the mid-span region as a potential deformation area and immediately generates a risk warning. After triggering the warning, the platform sends the risk alert to the operation and maintenance management terminal and simultaneously highlights the area in the bridge monitoring center's visualization interface, prompting maintenance personnel to take load reduction or traffic control measures, thereby effectively preventing potential structural safety accidents.

[0139] like Figure 2 As shown, this embodiment of the invention also provides a structure-wide dynamic response prediction device 200, which includes:

[0140] The first acquisition module 201 is used to acquire the real dataset of the structure in the current state, wherein the real dataset contains multiple sparse sensing data corresponding to the current structure.

[0141] The first processing module 202 is used to process the multiple sparse sensing data based on a preset finite element model and output a simulation dataset, wherein the simulation dataset includes multiple pairs of sparse input-global output data pairs.

[0142] The first training module 203 is used to pre-train the blank neural network model based on the simulation dataset to obtain a virtual model;

[0143] The second processing module 204 is used to perform physical correction processing on the virtual model to obtain the real model;

[0144] The first determining module 205 is used to process the sensing data of the current structure through a real-world model to determine the global dynamic response field data of the current structure.

[0145] The first execution module 206 is used to predict the risk deformation of the current structure based on the global dynamic response field data, and to execute the corresponding risk warning operation.

[0146] Optionally, the first acquisition module 201 mentioned above includes:

[0147] The first deployment submodule is used to deploy the preset sensor array in the vulnerable locations of the structure. The preset sensor array includes an acceleration sensor, a strain sensor, a displacement sensor, and an environmental sensor.

[0148] The first acquisition submodule is used to collect sparse sensing data of vulnerable locations of the structure in the current state through the preset sensor array to obtain the actual dataset of the structure in the current state. The sparse sensing data corresponding to the vulnerable locations is used to maximize the information reflecting the overall state of the structure.

[0149] Optionally, the first processing module 202 mentioned above includes:

[0150] The first determination submodule is used to determine the virtual load condition data required for the current simulation;

[0151] The second determining submodule is used to load the virtual load condition data into the preset finite element model, and process the multiple sparse sensing data through the finite element algorithm to determine multiple corresponding finite element response outputs.

[0152] The second acquisition submodule is used to pair multiple sparse sensing data with multiple corresponding finite element response outputs to obtain multiple sparse input-global output data pairs.

[0153] The first construction submodule is used to construct a simulation dataset based on multiple sparse input-global output data pairs.

[0154] Optionally, the first training module 203 mentioned above includes:

[0155] The third acquisition submodule is used to input sparse sensing data from the simulation dataset into the blank neural network model to obtain the training global output.

[0156] The second construction submodule is used to take the finite element response output corresponding to the sparse sensing data as a supervision label, and compare the training global output with the finite element response output according to the time-frequency domain contrast regularization fidelity loss function to obtain the calculation result. The time-frequency domain contrast regularization fidelity loss function is constructed based on the time-domain morphological fidelity data, frequency-domain robustness feature data and multi-scale physical consistency data corresponding to the sparse sensing data of the current structure.

[0157] The first training submodule is used to iteratively train the blank neural network model with the optimization objective of minimizing the calculation result, and obtain a virtual model after the iterative training is completed.

[0158] Optionally, the second processing module 204 mentioned above includes:

[0159] The fourth acquisition submodule is used to acquire sparse sensing data of the current structure;

[0160] The second training submodule is used to adjust and train the output layer and correction layer of the virtual model based on the sparse sensing data of the current structure, and to determine the correction parameters corresponding to the output layer and correction layer of the virtual model.

[0161] The third training submodule is used to perform corrective training on the virtual model based on the corrected parameters, and obtain the real model after the training converges. The real model is used to output physical characteristic data reflecting the current structure.

[0162] Optionally, the first determining module 205 mentioned above includes:

[0163] The third determining submodule is used to input the sparse sensing data of the current real-time structure into the reality model for nonlinear mapping inference, and determine the global dynamic response field data corresponding to the current real-time structure. The global dynamic response field data is used to reflect the dynamic response distribution of the current structure at different locations and at different times.

[0164] Optionally, the first execution module 206 mentioned above includes:

[0165] The fourth determination submodule is used to perform temporal and spatial feature calculations on the global dynamic response field data to determine the deformation trend of the current structure at different locations;

[0166] The identification submodule is used to compare the deformation trend with the preset risk trend to identify potential deformation areas;

[0167] The operation submodule is used to generate corresponding risk warning information based on the potential deformation area and trigger the execution of the corresponding risk warning operation.

[0168] like Figure 3 As shown, this embodiment of the invention also provides an electronic device 300, including a processor, which can execute any of the above-described structural global dynamic response prediction methods.

[0169] Specifically, it includes processor 301 and memory 302, as well as a computer program for a global dynamic response prediction method of execution structure stored in memory 302 and capable of running on processor 301, wherein:

[0170] The processor 301 executes the calculator program for the structure-wide dynamic response prediction method stored in memory 302, and performs the following steps:

[0171] Obtain the real-world dataset of the structure in the current state, wherein the real-world dataset contains multiple sparse sensor data corresponding to the current structure;

[0172] Based on a preset finite element model, the multiple sparse sensing data are processed to output a simulation dataset, which includes multiple pairs of sparse input-global output data pairs.

[0173] Based on the simulation dataset, the blank neural network model is pre-trained to obtain a virtual model;

[0174] The virtual model is then subjected to physical correction processing to obtain the real model;

[0175] By processing the sensing data of the current structure using a realistic model, the full-domain dynamic response field data of the current structure can be determined.

[0176] Based on the global dynamic response field data, the risk deformation of the current structure is predicted, and the corresponding risk warning operation is executed.

[0177] Optionally, processor 301 executes the process of obtaining the real-world dataset of the structure in the current state, including:

[0178] The preset sensor array is deployed at vulnerable locations in the structure, and the preset sensor array includes an acceleration sensor, a strain sensor, a displacement sensor, and an environmental sensor.

[0179] The sparse sensing data of the vulnerable locations of the structure in the current state are collected by the preset sensor array to obtain the actual dataset of the structure in the current state. The sparse sensing data corresponding to the vulnerable locations are used to maximize the information reflecting the overall state of the structure.

[0180] Optionally, the processor 301 executes the processing of the multiple sparse sensing data based on the preset finite element model, and outputs a simulation dataset, including:

[0181] Determine the virtual load case data required for the current simulation;

[0182] The virtual load condition data is loaded into the preset finite element model, and the multiple sparse sensor data are processed by the finite element algorithm to determine multiple corresponding finite element response outputs.

[0183] Multiple sparse sensing data are paired with multiple corresponding finite element response outputs to obtain multiple sparse input-global output data pairs;

[0184] A simulation dataset is constructed based on multiple sparse input-global output data pairs.

[0185] Optionally, processor 301 performs pre-training on the blank neural network model based on the simulation dataset to obtain a virtual model, including:

[0186] The sparse sensing data in the simulation dataset is input into the blank neural network model to obtain the training global output;

[0187] The finite element response output corresponding to the sparse sensing data is used as a supervision label. The training global output is compared with the finite element response output according to the time-frequency domain contrast regularization fidelity loss function to obtain the calculation result. The time-frequency domain contrast regularization fidelity loss function is constructed based on the time-domain morphological fidelity data, frequency-domain robustness feature data and multi-scale physical consistency data corresponding to the sparse sensing data of the current structure.

[0188] With minimizing the computational result as the optimization objective, the blank neural network model is iteratively trained, and a virtual model is obtained after the iterative training is completed.

[0189] Optionally, the processor 301 performs the physical correction process on the virtual model to obtain the real model, including:

[0190] Acquire sparse sensing data of the current structure;

[0191] Based on the sparse sensing data of the current structure, the output layer and correction layer of the virtual model are adjusted and trained to determine the correction parameters corresponding to the output layer and correction layer of the virtual model.

[0192] Based on the correction parameters, the virtual model is corrected and trained. After the training converges, a real model is obtained. The real model is used to output physical characteristic data reflecting the current structure.

[0193] Optionally, the processor 301 performs the processing of the sensing data of the current structure through the reality model to determine the global dynamic response field data of the current structure, including:

[0194] The sparse sensing data of the current real-time structure is input into the reality model for nonlinear mapping inference to determine the global dynamic response field data corresponding to the current real-time structure. The global dynamic response field data is used to reflect the dynamic response distribution of the current structure at different locations and at different times.

[0195] Optionally, the processor 301 executes the prediction of risk deformation of the current structure based on the global dynamic response field data, and performs corresponding risk warning operations, including:

[0196] The temporal and spatial characteristics of the global dynamic response field data are calculated to determine the deformation trend of the current structure at different locations.

[0197] The deformation trend is compared with the preset risk trend to identify potential deformation areas;

[0198] Based on the potential deformation area, corresponding risk warning information is generated, and corresponding risk warning operations are triggered.

[0199] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the structure-wide dynamic response prediction method or the application-side structure-wide dynamic response prediction method provided in this invention, and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0200] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be done by a computer program instructing related hardware, and can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0201] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for predicting the dynamic response of a structure across its entire domain, characterized in that, include: Obtain the real-world dataset of the structure in the current state, wherein the real-world dataset contains multiple sparse sensor data corresponding to the current structure; Based on a preset finite element model, the multiple sparse sensing data are processed to output a simulation dataset, which includes multiple pairs of sparse input-global output data pairs. Based on the simulation dataset, the blank neural network model is pre-trained to obtain a virtual model; The virtual model is then subjected to physical correction processing to obtain the real model; By processing the sensing data of the current structure using a realistic model, the full-domain dynamic response field data of the current structure can be determined. Based on the global dynamic response field data, the risk deformation of the current structure is predicted, and the corresponding risk warning operation is executed. The process of pre-training a blank neural network model based on the simulation dataset to obtain a virtual model includes: The sparse sensing data in the simulation dataset is input into the blank neural network model to obtain the training global output; The finite element response output corresponding to the sparse sensing data is used as a supervision label. The training global output is compared with the finite element response output according to the time-frequency domain contrast regularization fidelity loss function to obtain the calculation result. The time-frequency domain contrast regularization fidelity loss function is constructed based on the time-domain morphological fidelity data, frequency-domain robustness feature data and multi-scale physical consistency data corresponding to the sparse sensing data of the current structure. With minimizing the computational result as the optimization objective, the blank neural network model is iteratively trained, and a virtual model is obtained after the iterative training is completed; The physical correction process performed on the virtual model to obtain the real-world model includes: Acquire sparse sensing data of the current structure; Based on the sparse sensing data of the current structure, the output layer and correction layer of the virtual model are adjusted and trained to determine the correction parameters corresponding to the output layer and correction layer of the virtual model. Based on the correction parameters, the virtual model is corrected and trained. After the training converges, a real model is obtained. The real model is used to output physical characteristic data reflecting the current structure.

2. The structural global dynamic response prediction method as described in claim 1, characterized in that, The method of obtaining the real-world dataset of the structure in the current state includes: A preset sensor array is deployed at vulnerable locations within the structure. The preset sensor array includes an accelerometer, a strain sensor, a displacement sensor, and an environmental sensor. The sparse sensing data of the vulnerable locations of the structure in the current state are collected by the preset sensor array to obtain the actual dataset of the structure in the current state. The sparse sensing data corresponding to the vulnerable locations are used to maximize the information reflecting the overall state of the structure.

3. The structural global dynamic response prediction method as described in claim 1, characterized in that, The process, based on a preset finite element model, processes the multiple sparse sensing data to output a simulation dataset, including: Determine the virtual load case data required for the current simulation; The virtual load condition data is loaded into the preset finite element model, and the multiple sparse sensor data are processed by the finite element algorithm to determine multiple corresponding finite element response outputs. Multiple sparse sensing data are paired with multiple corresponding finite element response outputs to obtain multiple sparse input-global output data pairs; A simulation dataset is constructed based on multiple sparse input-global output data pairs.

4. The structural global dynamic response prediction method as described in claim 1, characterized in that, The process of processing the sensing data of the current structure using a real-world model to determine the global dynamic response field data of the current structure includes: The sparse sensing data of the current real-time structure is input into the real-world model for nonlinear mapping inference to determine the global dynamic response field data corresponding to the current real-time structure. The global dynamic response field data is used to reflect the dynamic response distribution of the current structure at different locations and times.

5. The structural global dynamic response prediction method as described in claim 1, characterized in that, The process of predicting the risk deformation of the current structure based on the global dynamic response field data and executing corresponding risk warning operations includes: The temporal and spatial characteristics of the global dynamic response field data are calculated to determine the deformation trend of the current structure at different locations. The deformation trend is compared with the preset risk trend to identify potential deformation areas; Based on the potential deformation area, corresponding risk warning information is generated, and corresponding risk warning operations are triggered.

6. A structural global dynamic response prediction device, characterized in that, include: The first acquisition module is used to acquire the real dataset of the structure in the current state, wherein the real dataset contains multiple sparse sensing data corresponding to the current structure; The first processing module is used to process the multiple sparse sensing data based on a preset finite element model and output a simulation dataset, which includes multiple pairs of sparse input-global output data pairs. The first training module is used to pre-train the blank neural network model based on the simulation dataset to obtain a virtual model; The second processing module is used to perform physical correction processing on the virtual model to obtain the real model; The first determining module is used to process the sensing data of the current structure through a real-world model to determine the global dynamic response field data of the current structure. The first execution module is used to predict the risk deformation of the current structure based on the global dynamic response field data, and to execute the corresponding risk warning operation. The first training module is also used to input sparse sensing data from the simulation dataset into the blank neural network model to obtain the training global output; The finite element response output corresponding to the sparse sensing data is used as a supervision label. Based on the time-frequency domain contrast regularization fidelity loss function, the training global output is compared with the finite element response output to obtain the calculation result. The time-frequency domain contrast regularization fidelity loss function is constructed based on the time-domain morphological fidelity data, frequency-domain robustness feature data, and multi-scale physical consistency data corresponding to the sparse sensing data of the current structure. With minimizing the calculation result as the optimization objective, the blank neural network model is iteratively trained. After the iterative training is completed, a virtual model is obtained. The second processing module is also used to acquire sparse sensing data of the current structure; Based on the sparse sensing data of the current structure, the output layer and correction layer of the virtual model are adjusted and trained to determine the correction parameters corresponding to the output layer and correction layer of the virtual model; based on the correction parameters, the virtual model is corrected and trained, and a real model is obtained after the training converges. The real model is used to output physical characteristic data reflecting the current structure.

7. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps in the structure-wide dynamic response prediction method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps in the structure-wide dynamic response prediction method as described in any one of claims 1 to 5.

Images