A method and apparatus for health monitoring data prediction for bridge structures

Abstract

The application discloses a health monitoring data prediction method and device for a bridge structure, and comprises the following steps: receiving historical health monitoring data of a bridge structure from a server and preprocessing the historical health monitoring data to obtain to-be-processed data; constructing a bridge monitoring data tensor according to the to-be-processed data; constructing a weight tensor according to a time sequence state updating network and the bridge monitoring data tensor, then performing low-rank decomposition on the weight tensor to obtain a core tensor set and a factor matrix set, and then obtaining a reconstructed low-rank parameter according to the core tensor set and the factor matrix set; updating each state vector in the time sequence state updating network based on the bridge monitoring data tensor and the reconstructed low-rank parameter to obtain an updated state vector; iteratively optimizing a constructed training loss function by using the updated state vector, and outputting an optimized model parameter; and obtaining a prediction result by combining the bridge monitoring data tensor, the reconstructed low-rank parameter and the optimized model parameter.

Description

Technical Field

[0001] This invention relates to the field of bridge data processing technology, and in particular to a method and apparatus for predicting health monitoring data of bridge structures. Background Technology

[0002] As a crucial component of transportation infrastructure, the structural health of bridges directly impacts the safety and reliability of the transportation system. With the development of structural health monitoring technology, numerous sensors have been deployed at key locations on bridges to continuously collect monitoring data such as displacement, strain, acceleration, temperature, and vibration frequency, forming high-dimensional, multimodal, and spatiotemporally coupled monitoring data. This data provides an important foundation for bridge structural condition assessment, damage identification, and life prediction.

[0003] However, in practical applications, factors such as sensor failure, communication interruption, power supply anomalies, and environmental interference often lead to missing monitoring data, resulting in high-dimensional incomplete features. Existing methods mostly employ "data completion + prediction" or deep learning methods based on flattened matrices / vectors, but under conditions of high missing rates and high-dimensional multimodal data, they are prone to problems such as error accumulation, spatial structure destruction, and excessively large parameter scales, thus significantly reducing prediction accuracy and stability.

[0004] Therefore, there is an urgent need for a method and device that can directly perform end-to-end prediction on high-dimensional incomplete monitoring data while maintaining the original multi-dimensional structural features, so as to achieve high-precision multi-step time series prediction. Summary of the Invention

[0005] To address the technical problem of low accuracy in predicting bridge structure data in existing technologies, this invention proposes a method and apparatus for predicting health monitoring data of bridge structures.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for predicting health monitoring data for bridge structures, specifically including the following steps:

[0008] S1: Obtain historical data of bridge structural health monitoring from the server, preprocess the historical monitoring data to obtain bridge monitoring data to be processed;

[0009] S2: Based on the bridge monitoring data to be processed, construct a bridge monitoring data tensor according to the time dimension, sampling mode dimension, and sensor dimension;

[0010] S3: Construct a weight tensor based on the time-series state update network and bridge monitoring data tensor, then perform low-rank decomposition on the weight tensor to obtain the core tensor set and factor matrix set, and then obtain the reconstructed low-rank parameters based on the core tensor set and factor matrix set.

[0011] S4: Based on the reconstructed low-rank parameters, the bridge monitoring data tensor is used to perform missing perception tensor state update calculation, and the network state tensor is iteratively updated through update gate, reset gate and candidate state gate.

[0012] S5: Iteratively optimize the constructed training loss function using the updated state vector, and output the optimized model parameters;

[0013] S6: The prediction results are obtained by combining the bridge monitoring data tensor, the reconstructed low-rank parameters, and the optimized model parameters.

[0014] In summary, by adopting the above technical solution, the present invention has at least the following beneficial effects compared with the prior art:

[0015] This invention organizes the monitoring data collected by multiple sensors in a bridge structural health monitoring system into a high-order incomplete tensor, and combines low-rank decomposition to generate the core tensor and factor matrix in the prediction model. At the same time, it performs joint constraint training through missing perception temporal state update and tensor distance optimization. This can effectively extract the temporal evolution features, sampling modal features and sensor spatial correlation features from the bridge monitoring data, thereby improving the accuracy and stability of predicting future bridge monitoring data.

[0016] This method and device can be widely applied in engineering fields such as response prediction of key bridge measuring points, structural state evolution analysis, abnormal risk early warning, intelligent operation and maintenance decision-making, and bridge structural safety assessment. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a health monitoring data prediction method for bridge structures according to an exemplary embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a health monitoring data prediction device for bridge structures according to an exemplary embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of a storage module according to an exemplary embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of a parameter reconstruction module according to an exemplary embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram of a model parameter optimization module according to an exemplary embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to embodiments and specific implementation methods. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0023] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0024] like Figure 1 As shown, the present invention provides a method for predicting health monitoring data of bridge structures, specifically including the following steps:

[0025] S1: Receive historical health monitoring data and prediction instructions for the bridge structure from the server, and preprocess them to obtain the data to be processed.

[0026] In this embodiment, the historical health monitoring data of the bridge structure comes from different sensors. To facilitate subsequent data processing, the monitoring data from different sensors needs to be preprocessed.

[0027] The preprocessing includes time synchronization, outlier detection, missing value labeling, and data standardization. The monitoring data includes displacement, velocity, acceleration, strain, stress, temperature, deflection, vibration frequency, and other monitoring data characterizing the bridge structure. By unifying the sampling time reference, asynchronous monitoring data from different sensors are time-aligned, and a continuous monitoring sequence is constructed according to a preset sampling period. This establishes the correspondence between sampling time, monitoring modes, and sensor measurement points, providing a unified data input for subsequent high-dimensional tensor construction.

[0028] In this embodiment, the historical health monitoring data of the bridge structure, the data to be processed, the bridge monitoring data tensor, the prediction results of the monitoring data corresponding to multiple future time steps of the bridge, and the intermediate state data generated during the prediction process are all stored in the data storage module for traceability.

[0029] S2: Construct a bridge monitoring data tensor based on the data to be processed.

[0030] S2-1: Construct indexes of different dimensions based on the data to be processed, including time dimension index, sampling mode dimension index and sensor dimension index.

[0031] Among them, the time dimension is used to represent the evolution of bridge monitoring data over the number of sampling days, the sampling mode dimension is used to represent the data patterns formed at different sampling times under the same number of sampling days, and the sensor dimension is used to represent the monitoring data collected by different sensors in the bridge monitoring system.

[0032] By establishing a dimensional index, the data organization structure at different times, sampling modes, and sensor nodes is described, providing a foundation for subsequent tensor construction.

[0033] S2-2: Generate bridge monitoring data tensors based on dimensional indexes, with the size of each time-series data window being [size missing]. Let represent the historical bridge monitoring data corresponding to P days, I sampling modes, and J sensors; simultaneously, construct the known data set Λ and the missing data set. , where Λ≫ This represents a high-dimensional incomplete tensor where the vast majority of data is missing. The constructed tensor is used to represent the bridge monitoring information collected by the corresponding sensor at a specific time and under a specific sampling mode, thereby maintaining the multidimensional correlation characteristics of bridge monitoring data in the time, mode and sensor dimensions, and providing a tensor data foundation for subsequent low-rank decomposition of weighted tensors and time series prediction.

[0034] S3: Construct a weight tensor based on the time-series state update network and bridge monitoring data tensor, then perform low-rank decomposition on the weight tensor to obtain the core tensor set and factor matrix set, and then obtain the reconstructed low-rank parameters based on the core tensor set and factor matrix set.

[0035] S3-1: Extract the state tensor from the time-series state update network involved in the bridge monitoring data prediction process, and combine the bridge monitoring data tensor and the state tensor to obtain the weight tensor. The weight tensor includes the input weight tensor, the historical state weight tensor, and the output layer weight tensor, while maintaining its original high-order tensor structure.

[0036] In this embodiment, the temporal state update network employs the Missing-Aware Tucker Gated Recurrent Unit Network (MATR) model to extract temporal features and update the state of the bridge monitoring data tensor. The Missing-Aware Tucker Gated Recurrent Unit includes updating the state tensor. Reset state tensor Candidate hidden state tensor and hidden state tensor Each state tensor maintains the multidimensional tensor structure of the bridge monitoring data.

[0037] Among them, the state tensor is updated. Used to control the fusion ratio of historical hidden state information and current input information; reset the state tensor. Used to control the degree to which historical state information is retained during candidate state calculation; candidate hidden state tensor Used to fuse current monitoring data features and historical state information to generate new candidate states; hidden state tensor Used to save the timing characteristics information of the current time step.

[0038] Unlike traditional recurrent neural networks, the state update process in this embodiment is calculated based on a high-dimensional bridge monitoring data tensor, and only valid observation data is used for state update by constraining the missing location set, thereby achieving prediction of incomplete bridge health monitoring time series data.

[0039] In this embodiment, the bridge monitoring data tensor for each time window is: , of which The input tensor at each time step is represented as: The state tensor of a time-series state update network includes the update state tensor. Reset state tensor Candidate state tensor and hidden state tensor , The aforementioned state tensor is generated by the temporal state update network based on the input tensor at the current time step. Hidden state tensor of the previous time step The weight parameters after low-rank decomposition are obtained through iterative calculation using the gated state update formula.

[0040] Furthermore, based on the bridge monitoring data tensor The dimensional structure is used to perform high-order tensor parameterization on the input weight tensor, and the updated state input weight tensor is learned through the network training process. Reset the state input weight tensor and candidate state input weight tensor , ;

[0041] Furthermore, the historical state (the first) The hidden state tensor H obtained by the state update network at each time step is... (t−1) The historical state tensor (used to store the historical time-series characteristics of bridge monitoring data) is combined with each historical state tensor (according to the gating state update formula) to obtain the influence of historical states on the update gate, reset gate, and candidate states. The historical state tensor includes... ;

[0042] Furthermore, the output layer weight tensor is: Output layer weight tensor The tensor is initialized according to the required dimensions of the predicted output tensor and optimized through backpropagation of the loss function during the training phase. It is used to map the hidden state tensor to the predicted tensor of future bridge monitoring data.

[0043] The weight tensors are all represented in the form of higher-order tensors to maintain the original multidimensional structural features of the bridge monitoring data.

[0044] The input weight tensor, the historical state weight tensor, and the output layer weight tensor are then used to construct a weight tensor: .

[0045] S3-2: Perform Tucker low-rank decomposition on the extracted weight tensors (input weight tensor, historical state weight tensor, and output layer weight tensor) to generate the corresponding core tensor and factor matrix.

[0046] In this embodiment, a low-rank decomposition is performed on each weight tensor to generate the corresponding core tensor and factor matrix. The following formula is the single-element form of the low-rank decomposition of the weight tensor:

[0047]

[0048] in, , , This represents the input weight tensor. , , It is W Z W R , ∈ I×J×I×J A single element in; , , Represents the historical state weight tensor. , , It's U Z U R , ∈ I×J×I×J A single element in; This represents the output layer weight tensor. It is W Y ∈ L×I×J×I×J A single element in; ∈ W 1Z ∈ W 2Z ∈ W 3Z ∈ W4Z ∈ They represent the weight tensor W respectively. Z The core matrix and three factor matrices after decomposition; , , , and They represent W 1Z W 2Z W 3Z and W 4Z A single element.

[0049] For each weight tensor in the bridge monitoring data prediction network, the same low-rank decomposition method is used to obtain the corresponding core tensor and factor matrix parameters.

[0050] S3-3: Reconstruct the weight parameters in the time-series state update network based on the core tensor and factor matrix to obtain the reconstructed low-rank parameters, which are used in the bridge monitoring data prediction process.

[0051] For the bridge monitoring data tensor of each time series window: And prediction tensors: P represents the length of the timing window. This indicates the length of the prediction time window. A low-rank weight parameter set is constructed by combining the core tensor and the factor matrix to replace the original high-dimensional weight tensor for state updates and prediction calculations, thereby reducing the parameter size, training overhead, and improving the model's generalization ability.

[0052] Furthermore, by utilizing core tensors and factor matrices to complete network parameter mapping, while maintaining the structural information of bridge monitoring data in terms of time, modality, and sensor dimensions, the number of original weight parameters is reduced, thereby lowering model storage space and computational complexity.

[0053] The above tensor operations are used to realize multi-dimensional interactive calculations between input tensors, state tensors, and low-rank weight parameters, thereby providing a parameter basis for the subsequent missing-aware temporal state update module.

[0054] S4: Based on the bridge monitoring data tensor and the reconstructed low-rank parameters, the state vectors in the time-series state update network are updated to obtain the updated state vectors, so as to avoid error accumulation caused by missing data.

[0055] S4-1: Identify the current time step in the bridge monitoring data tensor known data set And mark the current time step tensor Which elements can be used for updating?

[0056] S4-2: The known data set Λ based on the current time step t Current time step input tensor X (t) The hidden state tensor H from the previous time step (t-1) The reconstructed low-rank parameters are used to update the state vectors in the temporal state update network.

[0057] In this embodiment, for the bridge monitoring data tensor in the th... Set of valid observation data locations at each time step The location index of any observed data is represented as:

[0058]

[0059] in, This represents the sampling modality dimension index in the bridge monitoring data tensor. Indicates the sensor dimension index; Indicates the first The set of valid observation locations in each time step where no data loss occurred.

[0060] Using only the known data set Λ at the current time step t The elements in the calculation are all driven by real observation data for updating the state, resetting the state, candidate state, and hidden state. Among them, the updated state element is used to independently control the fusion ratio between the historical state of the current position and the current candidate state; the reset state element is used to independently control the degree of retention of the historical state information of the current position; the candidate state element is used to generate a new state representation of the current position based on the current position observation data and its neighborhood effective observation data; and the hidden state element is used to perform position-by-position weighted fusion based on the hidden state element of the previous time step and the current candidate state element.

[0061] In this embodiment, the formula for calculating the state vector update is as follows:

[0062]

[0063]

[0064]

[0065]

[0066] in, The update state tensor Z represents the state after time step t. (t) ∈ I×J A single element; Indicates the activation function; X represents (t) ∈ I×J single element in Tensor Two-dimensional location index in and These correspond to the sampling mode and sensor number, respectively. Represents the time step in the bridge monitoring data tensor The known data set; It is W Z ∈ I×J×I×J single element in ∈I×J×I×J; It's U Z ∈ I×J×I×J A single element in; Indicates the first The hidden state tensor of the data at time step (i1, j1) -1; It updates the gate bias parameter B. Z ∈ I×J A single element in.

[0067] The update reset state tensor R represents the state after the update at time step t. (t) ∈ I×J A single element; W R ∈ I×J×I×J single element in U R ∈ I×J×I×J single element in Indicates resetting the gate bias parameter B R ∈ I×J A single element in.

[0068] Represents the candidate state tensor updated at time step t. (t) ∈ I×J A single element; Indicates the activation function; express ∈ I×J×I×J single element in express ∈ I×J×I×J A single element in; Bias parameters representing candidate hidden states ∈ I×J A single element in.

[0069] H represents the hidden state tensor after time step t. (t) ∈ I×J A single element; Represents the hidden state tensor after the update at time step t-1. (t) ∈ I×J A single element.

[0070] S4-3: Based on the hidden state tensor Through the output layer weight tensor The mapping yields the bridge monitoring data prediction tensor:

[0071]

[0072] in, It is a prediction tensor ∈ L×I×J single element, ( )∈L×I×J; yes ∈ L×I×J×I×J single element, ( )∈L×I×J×I×J; It is the output layer bias tensor B Y ∈ L×I×J single element, ( )∈L×I×J.

[0073] S5: Iteratively optimize the constructed training loss function using the updated state vector, and output the optimized model parameters.

[0074] S5-1: Three-dimensional prediction tensor output by the bridge monitoring data prediction module and the corresponding real monitoring data tensor The tensor is expanded in a fixed order according to the prediction time step, sampling mode, and sensor dimension, transforming the three-dimensional tensor into one-dimensional prediction and true vectors, and establishing a one-to-one correspondence between tensor elements and vector elements. A tensor metric matrix is ​​constructed based on the structural relationships between elements in the bridge monitoring data tensor to describe the degree of correlation between different tensor elements.

[0075] Specifically, the prediction result tensor and the actual monitoring tensor output by the bridge monitoring data prediction module both maintain a three-dimensional structure composed of time dimension, modal dimension and sensor dimension.

[0076] To facilitate the subsequent construction of the tensor distance metric model, the prediction tensor and the bridge monitoring data tensor are first expanded into vector form in a preset order (the monitoring data tensor is expanded in index order, that is, arranged in time dimension, sampling modality dimension, and sensor dimension, thereby converting the tensor data into a one-dimensional data vector).

[0077] For prediction tensor and real monitoring tensor The tensor is expanded in a fixed order according to the prediction time step, sampling mode, and sensor, transforming the three-dimensional tensor into a one-dimensional vector. For any element in the tensor... Mapped to the first element in the vector through an index mapping relationship This establishes a one-to-one correspondence between the elements of a three-dimensional tensor and the elements of a vector:

[0078]

[0079] Meanwhile, a corresponding tensor metric matrix is ​​established based on the positional relationship of tensor elements in the original tensor, so that elements with a high degree of correlation in the tensor have a greater weight contribution, while elements with a low degree of correlation have a smaller weight contribution.

[0080] The formula for constructing the tensor metric matrix is ​​as follows:

[0081]

[0082]

[0083] in, Represents the tensor metric matrix The Middle Line number The elements of the column are used to describe the structural correlation between corresponding elements in the predicted tensor and the actual monitored tensor; among them, Represents the first tensor after the prediction tensor expansion. Element index, The tensor expansion of the actual monitoring data represents the first... Element index; This represents the scaling parameter of the Gaussian kernel function; and They represent the first and second tensors in the prediction tensor, respectively. The element and the first element in the tensor of the actual monitoring data The position coordinates of each element in the original bridge monitoring data tensor; , These represent the time dimension indices corresponding to the m-th predicted element and the n-th true element, respectively. , These represent the sampling modality dimension indices corresponding to the m-th predicted element and the n-th true element, respectively. These represent the sensor dimension indices corresponding to the m-th predicted element and the n-th true element, respectively.

[0084] The tensor metric matrix can reflect the overall structural characteristics of bridge monitoring data in the time dimension, modal dimension, and sensor dimension, and embeds the spatial correlation information in the original tensor into the subsequent error calculation process, thereby avoiding the problem that traditional Euclidean distance only considers the corresponding position error and ignores the structural relationship.

[0085] S5-2: Construct the loss function based on the tensor metric matrix.

[0086] Specifically, this unit uses tensor distance as the primary optimization objective. During model training, the tensor distance between the predicted results and the actual monitoring data is gradually reduced by continuously adjusting the core tensor parameters and factor matrix parameters in the weight tensor low-rank decomposition module, as well as the gating parameters in the missing data-aware temporal state update module. Through this optimization process, the model not only improves the prediction accuracy of future monitoring data but also maintains the original structural characteristics of bridge monitoring data in both time and spatial dimensions, thereby enhancing the model's stability and generalization ability in scenarios with high missing data rates. Furthermore, when the loss function reaches the preset convergence condition, the final optimized model parameters are output and used for predicting future bridge monitoring data to achieve bridge structural state evolution trend analysis, anomaly early warning, and operation and maintenance decision support.

[0087] The specific formula for calculating the loss function is as follows:

[0088]

[0089]

[0090]

[0091]

[0092] in, Represents the training loss function; Let represent the training loss function at time step t; L represents the prediction time window length. Represents the tensor distance function; This represents the prediction result at time step t; This represents the actual result at time step t; , Represent tensors respectively and The expanded vector; G represents the tensor metric matrix; Represents the prediction vector The Middle One element, Represents the real data vector The Middle One element, Represents the prediction vector The Middle One element, Represents the real data vector The nth element.

[0093] Therefore, element-wise loss The calculation formula can be obtained as follows:

[0094]

[0095] in, Indicates the first The prediction tensor corresponding to the prediction time step is the first... The loss function value corresponding to each element is used to measure the structured error between the model-predicted tensor and the actual monitored tensor.

[0096] S5-3: Iteratively optimize the loss function based on the updated state vector.

[0097] In this embodiment, iterative optimization using temporal backpropagation is performed. Since each time step shares the weight tensor, core tensor, factor matrix, and bias tensor during the missing-aware temporal state update process, the gradient of each time step is backpropagated along the temporal direction to earlier time steps, and the gradients of the same parameter at different time steps are accumulated.

[0098]

[0099] in, Represents the loss function Update the gate weight core tensor after low-rank decomposition Mid-lane Element The partial derivatives; The first tensor of the prediction tensor The loss function value corresponding to each element; Indicates updating the gate weight tensor The core tensor obtained after low-rank decomposition median coordinate Elements at the location; These represent the updated gate weight core tensor, respectively. Single-element index; Indicates the first In the tensor of the prediction results at the nth time step, the nth The prediction step, the first The modality, the first Predicted values ​​corresponding to each sensor location; This represents the position in the hidden state tensor after time step t. The corresponding hidden element; This represents the learning rate.

[0100] The parameters to be optimized include the core tensor obtained from the low-rank decomposition of the weight tensor, the factor matrix, the gated weight parameters in the missing-aware temporal state update process, the output layer weight parameters, and the corresponding bias parameters. Gradient descent is used to update the parameters.

[0101]

[0102] in, This represents the learning rate.

[0103] S5-4: Determine whether the iterative optimization has reached the termination condition. If not, continue training. If yes, output the optimized model parameters.

[0104] In this embodiment, the termination conditions include: the number of training iterations reaches the preset maximum number of iterations, or the difference between the current round loss function value and the previous round loss function value is less than the preset convergence threshold; if the termination conditions are met, training is stopped, and the optimized core tensor, factor matrix, gating weight parameters, output layer weight parameters, and bias parameters are output.

[0105] Through the aforementioned temporal backpropagation process, the tensor distance loss is propagated along the time direction to each historical time step, and the shared parameters are cumulatively updated, enabling the trained prediction model to maintain stable temporal feature extraction and multi-step prediction capabilities under the condition of high-dimensional incomplete bridge monitoring data.

[0106] S6: Using the model parameters optimized in step S5, predict the new bridge monitoring data tensor and output the predicted results of bridge health monitoring data within the future time window.

[0107] Specifically, the tensor of the bridge monitoring data to be predicted is input into the trained missing perception tensor state update network, and the updated hidden state tensor is used as the basis for the update. Through the output layer weight tensor Mapping yields the prediction tensor , , indicating the future Each time step Each monitoring mode and Predicted bridge monitoring data corresponding to each sensor.

[0108] Finally, the prediction results are stored for bridge structural condition analysis or subsequent early warning.

[0109] In this embodiment, the generated bridge monitoring data prediction results for multiple future time steps are used for bridge key measuring point response prediction, structural state evolution analysis, abnormal risk warning, and operation and maintenance decision support.

[0110] like Figure 2 As shown, based on the above-mentioned method for predicting health monitoring data for bridge structures, the present invention also provides a device for predicting health monitoring data for bridge structures, including a data receiving module 1, a storage module 2, a tensor construction module 3, a parameter reconstruction module 4, a state vector update module 5, a model parameter optimization module 6, and an output module 7.

[0111] The output of the data receiving module 1 is connected to the first input of the storage module 2. The output of the storage module 2 is connected to the input of the tensor construction module 3. The output of the tensor construction module 3 is connected to the input of the parameter reconstruction module 4. The output of the parameter reconstruction module 4 is connected to the input of the state vector update module 5. The output of the state vector update module 5 is connected to the input of the model parameter optimization module 6. The output of the model parameter optimization module 6 is connected to the input of the output module 7. The output of the output module 7 is connected to the second input of the storage module 2.

[0112] Among them, the data receiving module 1 is used to receive historical health monitoring data and prediction instructions of the bridge structure from the server;

[0113] Storage module 2 is used to store historical health monitoring data sent by data receiving module 1 and prediction results sent by output module 7;

[0114] Tensor construction module 3 is used to construct bridge monitoring data tensors based on historical health monitoring data;

[0115] The parameter reconstruction module 4 is used to update the network and bridge monitoring data tensors according to the time series status to construct a weight tensor, then perform low-rank decomposition on the weight tensor to obtain the core tensor set and the factor matrix set, and then obtain the reconstructed low-rank parameters according to the core tensor set and the factor matrix set.

[0116] The state vector update module 5 is used to update the state vector based on the bridge monitoring data tensor and the reconstructed low-rank parameters to obtain the updated state vector.

[0117] Model parameter optimization module 6 is used to iteratively optimize the constructed training loss function using the updated state vector and output the optimized model parameters;

[0118] Output module 7 is used to output prediction results based on bridge monitoring data tensors, reconstructed low-rank parameters, and optimized model parameters, and send them to storage module 2 for storage.

[0119] In this embodiment, the data receiving module 1 includes a data receiving unit and a preprocessing unit. The data receiving unit is used to receive historical health monitoring data of the bridge structure; the preprocessing unit is used to preprocess the historical health monitoring data to obtain data to be processed.

[0120] In this embodiment, as Figure 3 As shown, storage module 2 includes a first storage unit 21 and a second storage unit 22. The first storage unit 21 is used to store historical health monitoring data and data to be processed; the second storage unit 22 is used to store the prediction results output by output module 7.

[0121] In this embodiment, as Figure 4 As shown, the parameter reconstruction module 4 includes a weight tensor construction unit 41, a low-rank decomposition unit 42, and a reconstruction unit 43.

[0122] Among them, the weight tensor construction unit 41 is used to extract the state tensor from the time-series state update network and combine the bridge monitoring data tensor and the state tensor to obtain the weight tensor. The weight tensor includes the input weight tensor, the historical state weight tensor and the output layer weight tensor.

[0123] Low-rank decomposition unit 42 is used to perform low-rank decomposition on the weight tensor to generate a core tensor set and a factor matrix set.

[0124] Reconstruction unit 43 is used to obtain the reconstructed low-rank parameters based on the core tensor set and the factor matrix set.

[0125] In this embodiment, as Figure 5 As shown, the model parameter optimization module 6 includes a tensor metric matrix construction unit 61, a loss function construction unit 62, and an iterative optimization unit 63.

[0126] Among them, the tensor metric matrix construction unit 61 is used to expand the bridge monitoring data tensor into a vector form according to a preset order and establish the corresponding tensor metric matrix.

[0127] Loss function construction unit 62 is used to construct a loss function based on the tensor metric matrix;

[0128] The iterative optimization unit 63 is used to iteratively optimize the loss function based on the updated state vector and output the optimized model parameters.

[0129] In this embodiment, a storage medium is also provided, which stores a computer program for executing the above-described method for predicting health monitoring data for bridge structures.

[0130] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of the present invention.

Claims

1. A method for predicting health monitoring data of bridge structures, characterized in that, Specifically, the following steps are included: S1: Obtain historical bridge structural health monitoring data from the server, preprocess the historical monitoring data to obtain bridge monitoring data to be processed; S2: Based on the bridge monitoring data to be processed, construct a bridge monitoring data tensor according to the time dimension, sampling mode dimension, and sensor dimension; S3: Construct a weight tensor based on the time-series state update network and bridge monitoring data tensor, then perform low-rank decomposition on the weight tensor to obtain the core tensor set and factor matrix set, and then obtain the reconstructed low-rank parameters based on the core tensor set and factor matrix set. S4: Based on the reconstructed low-rank parameters, the bridge monitoring data tensor is used to perform missing perception tensor state update calculation, and the network state tensor is iteratively updated through update gate, reset gate and candidate state gate. S5: Iteratively optimize the constructed training loss function using the updated state vector, and output the optimized model parameters; S6: The prediction results are obtained by combining the bridge monitoring data tensor, the reconstructed low-rank parameters, and the optimized model parameters.

2. The method for predicting health monitoring data for bridge structures as described in claim 1, characterized in that, In S1, the preprocessing includes time synchronization processing, outlier detection processing, missing value marking processing, and data standardization processing.

3. The method for predicting health monitoring data for bridge structures as described in claim 1, characterized in that, In S2, the size of the bridge monitoring data tensor for each time window is , This represents the historical monitoring data of the bridge corresponding to the p-th day, the i-th sampling mode, and the j-th sensor.

4. The method for predicting health monitoring data for bridge structures as described in claim 1, characterized in that, S3 includes: S3-1: Extract the state tensor from the temporal state update network, and combine the bridge monitoring data tensor and the state tensor to obtain the weight tensors. The weight tensors include the input weight tensor, the historical state weight tensor, and the output layer weight tensor. S3-2: Perform low-rank decomposition on each weight tensor to generate the corresponding core tensor and factor matrix. The following formula is the single-element form of the low-rank decomposition of the weight tensor: in, , , This represents the input weight tensor. , , It is W Z W R , ∈ I×J×I×J The single element in ( )∈I×J×I×J; , , Represents the historical state weight tensor. , , It's U Z U R , ∈ I×J×I×J The single element in ( )∈I×J×I×J; This represents the output layer weight tensor. It is W Y ∈ L×I×J×I×J The single element in ( )∈L×I×J×I×J; ∈ W 1Z ∈ W 2Z ∈ W 3Z ∈ W 4Z ∈ They represent the weight tensor W respectively. Z The core matrix and three factor matrices after decomposition; , , , and They represent W 1Z W 2Z W 3Z and W 4Z The single element (r1, r2, r3, r4) ∈ , ( ,r1)∈ , ( r2)∈ , ( r3)∈ , ( r4)∈ ; S3-3: Reconstruct the temporal state update network weight parameters based on the core tensor and factor matrix to obtain the reconstructed low-rank parameters.

5. The method for predicting health monitoring data for bridge structures as described in claim 1, characterized in that, In step S4, the formula for calculating the state vector update is: in, Z represents the update state tensor after time step t. (t) ∈ I×J A single element; Indicates the activation function; X represents (t) ∈ I×J single element in Tensor Two-dimensional location index in and These correspond to the sampling mode and sensor number, respectively. Represents the time step in the bridge monitoring data tensor The known data set; It is W Z ∈ I×J×I×J single element in ∈I×J×I×J; It's U Z ∈ I×J×I×J A single element in; Indicates the first The hidden state tensor of the data at time step (i1, j1) -1; It updates the gate bias parameter B. Z ∈ I×J A single element in; The update reset state tensor R represents the state after the update at time step t. (t) ∈ I×J A single element; W R ∈ I×J×I×J single element in U R ∈ I×J×I×J single element in Indicates resetting the gate bias parameter B R ∈ I×J A single element in; Represents the candidate state tensor updated at time step t. (t) ∈ I×J A single element; Indicates the activation function; express ∈ I×J×I×J single element in express ∈ I×J×I×J A single element in; Bias parameters representing candidate hidden states ∈ I×J A single element in; H represents the hidden state tensor after time step t. (t) ∈ I×J A single element; Represents the hidden state tensor after the update at time step t-1. (t) ∈ I×J A single element.

6. The method for predicting health monitoring data for bridge structures as described in claim 1, characterized in that, S5 includes: S5-1: Vectorize the prediction tensor output by the bridge monitoring data prediction module and the corresponding bridge monitoring data tensor to construct a tensor metric matrix: in, Represents the tensor metric matrix The Middle Line 1 Column elements; This represents the scaling parameter of the Gaussian kernel function; and They represent the first and second tensors in the prediction tensor, respectively. The element and the first element in the tensor of the actual monitoring data The position coordinates of each element in the original bridge monitoring data tensor , These represent the time dimension indices corresponding to the m-th predicted element and the n-th true element, respectively. , These represent the sampling modality dimension indices corresponding to the m-th predicted element and the n-th true element, respectively. These represent the sensor dimension indices corresponding to the m-th predicted element and the n-th true element, respectively. S5-2: Constructing the loss function based on the tensor metric matrix: in, Represents the training loss function; Let represent the training loss function at time step t; L represents the prediction time window length. Represents the tensor distance function; This represents the prediction result at time step t; This represents the actual result at time step t; , Represent tensors respectively and The expanded vector; G represents the tensor metric matrix; Represents the prediction vector The Middle One element, Represents the real data vector The Middle One element, Represents the prediction vector The Middle One element, Represents the real data vector The nth element; S5-3: Iteratively optimize the loss function based on the updated state vector: in, Represents the loss function Update the gate weight core tensor after low-rank decomposition Mid-lane Element The partial derivatives; The first tensor of the prediction tensor The loss function value corresponding to each element; Indicates updating the gate weight tensor The core tensor obtained after low-rank decomposition median coordinate Elements at the location; These represent the updated gate weight core tensor, respectively. Single-element index; Indicates the first In the tensor of the prediction results at the nth time step, the nth The prediction step, the first The modality, the first Predicted values ​​corresponding to each sensor location; This represents the position in the hidden state tensor after time step t. The corresponding hidden element; Indicates the learning rate; S5-4: Determine whether the iterative optimization has reached the termination condition. If not, continue training; otherwise, output the optimized model parameters.

7. The method for predicting health monitoring data for bridge structures as described in claim 6, characterized in that, In S5-4, the termination conditions include: The training iteration rounds reach the preset maximum number of iterations; or the difference between the current round loss function value and the previous round loss function value is less than the preset convergence threshold.

8. A health monitoring data prediction device for bridge structures based on the method of any one of claims 1-7, characterized in that, include: The data receiving module is used to receive historical health monitoring data and prediction instructions for the bridge structure from the server; The storage module is used to store historical health monitoring data sent by the data receiving module and prediction results sent by the output module; The tensor construction module is used to construct bridge monitoring data tensors based on historical health monitoring data. The parameter reconstruction module is used to update the network and bridge monitoring data tensors according to the time series status to construct a weight tensor, then perform low-rank decomposition on the weight tensor to obtain the core tensor set and the factor matrix set, and then obtain the reconstructed low-rank parameters based on the core tensor set and the factor matrix set. The state vector update module is used to update the state vector based on the bridge monitoring data tensor and the reconstructed low-rank parameters to obtain the updated state vector. The model parameter optimization module is used to iteratively optimize the constructed training loss function using the updated state vector and output the optimized model parameters. The output module is used to output the prediction results based on the bridge monitoring data tensor, the reconstructed low-rank parameters, and the optimized model parameters, and send them to the storage module for storage.

9. The health monitoring data prediction device for bridge structures as described in claim 8, characterized in that, The parameter reconstruction module includes a weight tensor construction unit, a low-rank decomposition unit, and a reconstruction unit. The weight tensor construction unit is used to extract the state tensor from the temporal state update network and combine the bridge monitoring data tensor and the state tensor to obtain the weight tensor. The weight tensor includes the input weight tensor, the historical state weight tensor, and the output layer weight tensor. The low-rank decomposition unit is used to perform low-rank decomposition on the weight tensor to generate a set of core tensors and a set of factor matrices. The reconstruction unit is used to obtain the reconstructed low-rank parameters based on the core tensor set and the factor matrix set.

10. The health monitoring data prediction device for bridge structures as described in claim 8, characterized in that, The model parameter optimization module includes a tensor metric matrix construction unit, a loss function construction unit, and an iterative optimization unit; Among them, the tensor metric matrix construction unit is used to expand the bridge monitoring data tensor into vector form according to a preset order and establish the corresponding tensor metric matrix; Loss function construction unit, used to construct the loss function based on the tensor metric matrix; The iterative optimization unit is used to iteratively optimize the loss function based on the updated state vector and output the optimized model parameters.

Images