Real-time Diagnosis Method for Urban Bridge Conditions Based on Health Monitoring Sequence Data and Variational Recurrent Neural Network

Abstract

The present invention proposes a real-time diagnosis method for the state of urban bridges based on health monitoring sequence data and variational recurrent neural networks. This method models the reconstruction process of structural response monitoring sequence data based on variational recurrent neural networks as an inference process of inferring vehicle loads from the structural response monitoring sequence data and a generation process of reconstructing the structural response monitoring sequence data based on the inferred vehicle loads, so that the generation process represents the surrogate model of the bridge structure. When the service state of the structure changes, the residual index between the response value predicted by the surrogate model and the measured value is used to diagnose the state of bridge structural components and the whole. This method establishes a bridge structure surrogate model by decoupling vehicle loads in the latent space, realizing the interpretability of the diagnosis process; by modeling the structural response monitoring sequence data, it realizes the real-time prediction and diagnosis of the structural response monitoring sequence data, solving the problems of lack of interpretability, low accuracy and low time accuracy in traditional structural diagnosis methods.

Description

Technical Field

[0001] The present invention belongs to the technical fields of signal processing and structural health monitoring, and particularly relates to a method for real-time diagnosis of the state of urban bridges based on health monitoring sequence data and a variational recurrent neural network. The method is based on the correlation modeling of structural response monitoring sequence data and the structural state evaluation based on the model prediction of the residual of the structural response sequence, and can be directly applied to technical fields such as intelligent infrastructure and bridge engineering. Background Art

[0002] Bridges are an important part of China's infrastructure construction and national economic development. Currently, the number and scale of bridges in China rank among the top in the world. With the rapid growth of traffic flow and the degradation of bridge performance, the peak period of highway bridge maintenance during the large-scale construction period is about to arrive intensively. The assessment of the structural state of in-service bridges has become the core issue for improving the safety and durability level of bridges in China.

[0003] Under the assumption of a healthy bridge structure in a linear time-invariant system (LTIs), there is a correlation corresponding to the structural state among the structural responses generated by the structure under the action of the same vehicle load field. Correlation modeling usually adopts a machine learning model and obtains its parameters through training in the structural response dataset in the healthy state. If the structure suffers losses and its service state changes, the correlation pattern of the structural response data related to the structural state will also change accordingly, resulting in a deviation between the prediction result of the machine learning model trained from the dataset in the healthy state and the measured result. Therefore, this deviation can be used to indicate the change in the correlation pattern of the structural response data, and further infer the health state of the bridge structure.

[0004] However, the above-mentioned correlation pattern of the structural response will also change due to the influence of the vehicle load field. Therefore, in practical applications, it is often necessary to meet the assumption condition of an unchanged vehicle load field. However, the spatio-temporal distribution characteristics of the vehicle load field make it difficult to guarantee the above assumption. Therefore, in practical applications, the correlation of the statistical values of the structural response within a certain time period is usually analyzed based on the periodic law of the vehicle load (such as the statistical laws are basically similar every day, every week, and every year). However, this correlation in the statistical sense cannot be exactly the same, and there will be errors in the structural state diagnosis; at the same time, the machine learning methods used for correlation analysis and structural state diagnosis have the nature of a black box, resulting in difficulty in guaranteeing the reliability of the method; in addition, the time accuracy of this method in state diagnosis is poor, and it is difficult to achieve real-time online structural state diagnosis. Summary of the Invention

[0005] The object of the present invention is to solve problems such as poor accuracy of structural state diagnosis and time accuracy in the statistical sense, and to realize real-time and reliable diagnosis of the structural state based on the structural response monitoring sequence data. A real-time diagnosis method for urban bridge state based on health monitoring sequence data and variational recurrent neural network is proposed.

[0006] The present invention is realized through the following technical solutions. The present invention proposes a real-time diagnosis method for urban bridge state based on health monitoring sequence data and variational recurrent neural network. The method includes the following steps:

[0007] Step 1, establish a data set: establish a simulation data set of vehicle load - structural response and a real data set containing only real structural responses;

[0008] Step 2, establish a model: establish an encoder - decoder decoupled representation learning model based on variational recurrent neural network;

[0009] Step 3, design physical constraint conditions: establish constraint conditions for variational recurrent neural network;

[0010] Step 4, pre-training: train the encoder - decoder decoupled representation learning model of variational recurrent neural network based on the simulation data set;

[0011] Step 5, fine-tuning: freeze the parameters of the encoder part of the variational recurrent neural network, and fine-tune the parameters of the encoder part of the variational recurrent neural network based on the real structural response monitoring data set;

[0012] Step 6, structural state diagnosis: freeze all parameters, based on the real structural response monitoring sequence data, gradually deduce the future response values, and infer the real-time change of the bridge structural state according to the residual between the deduced values and the real values.

[0013] Further, in Step 1, for the bridge structure to be analyzed, establish its finite element model, load the measured sequence data of vehicle load and the sampled sequence data based on the vehicle load distribution onto the finite element model, and obtain the structural response sequence data corresponding to the measuring points of the whole bridge. If there are C structural response sensors in the whole bridge, the size of the obtained structural response sequence data is The length of the vehicle load sequence data is T.

[0014] Further, in Step 1, the bridge deck is evenly discretized along the longitudinal and transverse directions of the bridge into discrete grids, where M is the number of longitudinal grids and N is the number of transverse grids; according to the vehicle load sequence data and the position of the vehicle on the bridge deck, formulate the vehicle load matrix corresponding to each moment, and the size of the matrix is For the grids and moments without vehicle action, zero values are used for supplementation.

[0015] Further, in Step 2, the vehicle load on the bridge deck is represented by a triple, namely time, location, and amplitude. When using a discrete grid format, the triple is simplified to (t, W t ), where W t ∈R N×M is the vehicle load matrix of the shape defined on the discrete grid, containing information on the location and load magnitude of the vehicle; in SHM, the structural response monitoring data X t ∈R C is regarded as the structural response at each time step t under the action of the vehicle load sampling W t ~P(W;t);

[0016] X t =G(W t )

[0017] where t = 1, 2,..., T, T is the length of the monitoring time, and C is the number of channels;

[0018] In the variational recurrent neural network, at each moment t, a variational auto - encoding module is established, which consists of four parts: the structural response monitoring data embedder encoder Enc, the vehicle load embedder decoder Dec; among them, the embedder maps the structural response monitoring data and the vehicle load data to the same dimension as the recurrent connection hidden space state; the role of the encoder is to infer the vehicle load distribution P(W;t) corresponding to the moment t from the structural response monitoring sequence data, and the role of the decoder is to generate the reconstructed structural response monitoring sequence data from the inferred load sampling . The goal of the variational auto - encoding module is to minimize the reconstruction error When using variational inference, the optimization goal is to minimize the evidence lower bound ELBO:

[0019]

[0020] where, D KL (P φ (W t |X t )||P(W,t)) is the KL divergence between the inferred vehicle load distribution and the true distribution;

[0021] The recurrent connection in the variational recurrent neural network is implemented by a gated neural unit GRU, and its state update method is:

[0022]

[0023] Then, the calculation method for inferring the load distribution based on the recurrent connection is:

[0024] where

[0025] Among them, f I (·) represents an inference module composed of an encoder Enc, and μ t ∈R N×M and σ t ∈R N×M are respectively the mean and covariance matrix of the inferred vehicle load distribution;

[0026] The calculation method of the generation process based on the inferred load is as follows:

[0027] Where

[0028] Among them, f G (·) represents a generation module composed of a decoder Dec, and μ x,t and σ x,t are respectively the distribution parameters of the reconstructed structural response monitoring sequence data. During the generation process, directly let the mean represent the reconstructed data

[0029] On this basis, the loss function of the variational recurrent neural network model is to minimize the evidence lower bound L VRNN of the entire sequence and the reconstruction error L recons :

[0030]

[0031] Furthermore, in step three, the constraint conditions include that the shape of the latent space is the same as the discretized grid, the vehicle load position maintains spatial sparsity on the bridge deck, and the vehicle load amplitude and action position are continuous in the time domain.

[0032] Furthermore, in step four, the structural response monitoring sequence data is normalized by the maximum value according to the channels, that is, for different types of monitoring data, the structural response values at the channel and time with the largest absolute value are respectively selected for maximum value normalization.

[0033] Furthermore, in step four, the input structural response monitoring sequence data is randomly masked according to the channels, that is, in each iteration step of training, a certain proportion of the channels of the structural response monitoring data are randomly selected for masking, and the normal pre-training process is carried out.

[0034] Furthermore, in step six, by analyzing the residual between the measured sequence data of the bridge structure response in the healthy state and the predicted sequence data of the variational recurrent neural network, the mean m and standard deviation σ of the reconstruction residual in the healthy state are calculated based on the control chart, and m±3σ is used as the upper and lower thresholds for real-time inference of the bridge structure state.

[0035] The present invention provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the real-time diagnosis method for the state of urban bridges based on health monitoring sequence data and variational recurrent neural network are implemented.

[0036] The present invention provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the real-time diagnosis method for the state of urban bridges based on health monitoring sequence data and variational recurrent neural network are implemented.

[0037] Advantages of the present invention:

[0038] 1. The present invention formalizes the damage identification method based on reconstruction into an inference and generation process based on the VRNN model and decoupled representation learning, enhancing the interpretability of the data-driven model. By aligning the learned features with vehicle loads in the latent space, the model provides understandable and reliable results, which are crucial for actual structural state diagnosis.

[0039] 2. The present invention directly trains based on the structural response monitoring sequence data instead of performing correlation modeling in the probabilistic sense, and thus can achieve real-time diagnosis of the structural state and improve the time accuracy of diagnosis.

[0040] 3. Adopt a two-stage training strategy of pre-training + fine-tuning, that is, pre-train based on the simulation data set and then fine-tune based on a small amount of real monitoring data, indicating that with the help of finite element simulation data simulation, the method can be quickly extended to other structures to achieve rapid migration of the method. Description of the Drawings

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0042] Figure 1 It is a flowchart of the real-time diagnosis method for the state of urban bridges based on health monitoring sequence data and variational recurrent neural network of the present invention.

[0043] Figure 2 It is a schematic diagram of the processing of vehicle load sequence data in the pre-training data set.

[0044] Figure 3 It is a schematic diagram of the variational recurrent neural network model architecture.

[0045] Figure 4Schematic diagram of an explicit filtering function (assigning higher values to the shaded part). Detailed implementation manners

[0046] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0047] The model proposed in the present invention considers the structural response under vehicle loads. By adding physical constraint conditions of vehicle loads in the latent space, a variational recurrent neural network is guided to infer the corresponding vehicle load sequence from the input structural response monitoring sequence data. Furthermore, the model is divided into two parts: inference and generation. By decoupling the vehicle loads, the influence of vehicle loads is separated in the generative model composed of the decoder, and only the components related to the structural service state are retained. Thus, the generator becomes a surrogate model of the bridge structure and can be directly used for reliable and real-time diagnosis of the structural state. The method described in the present invention is trained through a paradigm of pre-training with simulation data + fine-tuning with real data. Based on the finite element simulation data, it can be extended to the diagnosis of the structural states of different types of bridge structures with only a small amount of data. The flowchart of the real-time diagnosis method for urban bridge states based on health monitoring sequence data and variational recurrent neural network proposed in the present invention is as Figure 1 shown.

[0048] Specifically, the present invention proposes a real-time diagnosis method for urban bridge states based on health monitoring sequence data and variational recurrent neural network. The method establishes a correlation model for structural response monitoring sequence data through a variational recurrent neural network and an encoder-decoder architecture, and establishes an unsupervised model training method based on the data reconstruction method. The method includes the following steps:

[0049] Step 1: Establish a data set: Establish a simulation data set of vehicle loads - structural responses and a real data set containing only real structural responses;

[0050] In Step 1, for the bridge structure to be analyzed, establish its finite element model, load the measured sequence data of vehicle loads (including vehicle weight, lane, vehicle speed, time of getting on the bridge, etc.) and the sampled sequence data based on the vehicle load distribution into the finite element model to obtain the sequence data of structural responses (including types such as main deflection, bearing displacement, inclination angle, cable force, etc.) corresponding to the measuring points of the whole bridge. There are C structural response sensors in the whole bridge, so the size of the obtained structural response sequence data is The length of the vehicle load sequence data is T.

[0051] The bridge deck is evenly discretized along the longitudinal and transverse directions of the bridge into a discrete grid, where M is the number of grids in the longitudinal direction of the bridge, and N is the number of grids in the transverse direction of the bridge (generally equal to the number of lanes); according to the vehicle load sequence data and the position of the vehicle on the bridge deck, a vehicle load matrix corresponding to each moment is formulated, and the size of the matrix is For grids and moments without vehicle action, zero values are used for supplementation. The data processing is as Figure 2 shown. Collect the structural response monitoring sequence data in the initial stage of the bridge structure operation (which can be regarded as the healthy state), and establish a pre-training and fine-tuning data set.

[0052] Step 2: Establish a model: Establish an encoder-decoder decoupled representation learning model based on the variational recurrent neural network;

[0053] The decoupled representation learning method focuses on separating the potential factors of data change into different and interpretable components. By realizing the decoupled representation of data, the correlation machine learning model is expected to extract the interpretable factors in the data structure and effectively generalize across tasks, thus providing a reliable method for structural state diagnosis. Aiming at problems such as poor accuracy and time accuracy in the correlation analysis of structural responses in the statistical sense, the structural response monitoring sequence data can be directly used for correlation analysis, and the physical constraints satisfying the vehicle load characteristics are imposed on the latent space through decoupled representation learning to guide the decoupled representation learning process, so as to extract the vehicle load corresponding to the given structural response monitoring sequence in the latent space. Essentially, it represents the correlation analysis of the structural response monitoring sequence data and the structural state diagnosis as an inference process of inferring the vehicle load from the structural response monitoring sequence data, and a generation process of reconstructing the structural response monitoring sequence data based on the inferred vehicle load. At this time, the generation process represents the surrogate model of the bridge structure.

[0054] In Step 2, the vehicle load on the bridge deck is represented by a triple, namely time, position, and amplitude. When using the discrete grid form, the triple is simplified to (t, W t ), where W t ∈R N×M is the vehicle load matrix of the shape defined on the discrete grid, which contains the position and load magnitude information of the vehicle; in SHM, the structural response monitoring data X t ∈R C is regarded as the structural response at each time step t under the action of the vehicle load sampling W t ~P(W; t);

[0055] X t = G(W t )

[0056] where t = 1, 2,..., T, T is the length of the monitoring time, and C is the number of channels (sensor structural response);

[0057] In a variational recurrent neural network, at each time step \(t\), the architecture of the variational recurrent neural network model is as follows Figure 3 As shown, a variational autoencoder module (VAE module) is established, which consists of a structural response monitoring data embedder encoder Enc, a vehicle load embedder and decoder Dec; among them, the embedder maps the structural response monitoring data and vehicle load data to the same dimension as the recurrent connection hidden space state; the role of the encoder is to infer the vehicle load distribution \(P(W;t)\) corresponding to time step \(t\) from the structural response monitoring sequence data, and the role of the decoder is to generate the reconstructed structural response monitoring sequence data from the sampled inferred load The goal of the variational autoencoder module is to minimize the reconstruction error When using variational inference, the optimization goal is to minimize the evidence lower bound ELBO:

[0058]

[0059] where \(D KL (P φ (W t |X t ))||P(W,t)) is the KL divergence between the inferred vehicle load distribution and the true distribution;

[0060] The recurrent connection in the variational recurrent neural network is implemented by a gated neural unit GRU, and its state update method is:

[0061]

[0062] Then the calculation method for inferring the load distribution based on the recurrent connection is:

[0063] where

[0064] where \(f I (·)\) represents the inference module composed of the encoder Enc, \(\mu t \in R N×M and \(\sigma t \in R N×M are respectively the mean and covariance matrix of the inferred vehicle load distribution;

[0065] The calculation method for the generation process based on the inferred load is:

[0066] where

[0067] where \(f G (·)\) represents the generation module composed of the decoder Dec, \(\mux,t and σ x,t are the distribution parameters of the reconstructed structural response monitoring sequence data respectively. During the generation process, the mean value is directly used to represent the reconstructed data

[0068] On this basis, the loss function of the variational recurrent neural network model is to minimize the lower bound of evidence L of the entire sequence VRNN and the reconstruction error L recons :

[0069]

[0070] Step 3. Design physical constraint conditions: establish the constraint conditions of the variational recurrent neural network;

[0071] In Step 3, the constraint conditions include that the shape of the latent space is the same as the discretized grid, the vehicle load position maintains spatial sparsity on the bridge deck, and the vehicle load amplitude and action position are continuous in the time domain.

[0072] When the percentage of zero elements in the matrix W t exceeds 50%, it is called spatial sparsity. Due to the limitation of the safety distance, the spatial sparsity of the vehicle load action position is a common characteristic of vehicles on the bridge deck. As a regularization method, the sparsity constraint is beneficial to reducing the optimization space and improving the optimization efficiency. The present invention introduces spatial sparsity in the way of top-k activation. Specifically, three sub-modules are designed in the decoder Enc, which are respectively used to output the mean matrix μ t ∈R N×M , the variance matrix σ t ∈R N×M and the gate module. The gate module is realized by activating the SoftMax function of the largest k items in the mean matrix, which is called top-k activation. Therefore, the sampling result has a sparse format, and the values of most discrete grids are zero.

[0073] For the time continuity of the load amplitude and loading position, implicit and explicit constraints are considered respectively. For the continuity constraint of the load amplitude, the MSE loss penalty between adjacent hidden states is introduced

[0074]

[0075] For the continuity constraint of the load action position, it is realized by adding an explicit filter, that is, for the grid in the forward direction corresponding to the vehicle loading position at the previous moment, a higher filtering value is given, as Figure 4 shown.

[0076] Step 4. Pretraining: train the encoder-decoder decoupled representation learning model of the variational recurrent neural network based on the simulation data set;

[0077] To improve the generalization performance and convergence speed of the model, the structural response monitoring sequence data is normalized by the maximum value for each channel, that is, for different types of monitoring data, the structural response values at the channel and moment with the largest absolute value are respectively selected for maximum value normalization.

[0078] To improve the robustness of the model under conditions such as data loss, the input structural response monitoring sequence data is randomly masked for each channel, that is, in each iteration step of training, a certain proportion of channels of the structural response monitoring data are randomly selected for masking, and the normal pre-training process is carried out.

[0079] Considering loss functions such as the evidence lower bound, reconstruction error, and time continuity constraint, vehicle load data is used as evidence in the latent space to calculate the evidence lower bound function, and the variational recurrent neural network model is pre-trained based on an unsupervised framework.

[0080] Step Five: Fine-tuning: Freeze the parameters of the encoder part of the variational recurrent neural network, and based on the real structural response monitoring data set, fine-tune the parameters of the encoder part of the variational recurrent neural network;

[0081] In Steps Four and Five, the network model is trained using the pre-training - fine-tuning strategy. The specific method is to pre-train based on the finite element simulation vehicle and response sequence data, and fine-tune based on the real response sequence data, so that this method can be quickly migrated to other bridge structures based on the structural finite element simulation data and a small amount of real response data, improving the application scope and generalization ability of the method.

[0082] Step Six: Structural state diagnosis: Freeze all parameters, based on the real structural response monitoring sequence data, gradually deduce the future response values, and according to the residuals between the deduced values and the real values, infer the change of the bridge structure state in real time.

[0083] In Step Six, by analyzing the residuals between the measured sequence data of the bridge structure response in the healthy state and the predicted sequence data of the variational recurrent neural network, the mean m and standard deviation σ of the reconstruction residuals in the healthy state are calculated based on the control chart, and m ± 3σ are used as the upper and lower thresholds for real-time inference of the bridge structure state.

[0084] The method described in the present invention uses an index based on prediction residuals for structural state diagnosis. Essentially, the reconstruction process of the structural response monitoring sequence data based on the variational recurrent neural network is modeled as an inference process of inferring vehicle loads from the structural response monitoring sequence data, and a generation process of reconstructing the structural response monitoring sequence data based on the inferred vehicle loads, so that the generation process represents the surrogate model of the bridge structure. When the service state of the structure changes, the residual index between the response value predicted based on the surrogate model and the measured value is used to diagnose the state of bridge structural components and the whole. The structural diagnosis method proposed by the present invention realizes the interpretability of the diagnosis process by decoupling vehicle loads in the latent space; by modeling the structural response monitoring sequence data, it realizes the real-time prediction and real-time diagnosis of the structural response monitoring sequence data, and solves the problems of lack of interpretability, low accuracy and low time accuracy in traditional structural diagnosis methods.

[0085] The present invention proposes a real-time diagnosis method for urban bridge states based on health monitoring sequence data and variational recurrent neural networks. The method includes variational recurrent neural network modeling suitable for correlation modeling of structural response monitoring sequence data, a latent space physical constraint method based on prior information of vehicle loads, a pre-training - fine-tuning network training architecture based on finite element simulation data and real data, a structural state rapid evaluation method based on prediction residuals of structural response monitoring sequences, etc. The method described in the present invention separates the potential factors of data changes into different and interpretable components based on the decoupled representation learning rule, and then extracts the interpretable factors in the structural response monitoring sequence data and effectively generalizes across tasks. Essentially, the correlation analysis of the structural response monitoring sequence data and the structural state diagnosis are represented as an inference process of inferring vehicle loads from the structural response monitoring sequence data, and a generation process of reconstructing the structural response monitoring sequence data based on the inferred vehicle loads. At this time, the generation process represents the surrogate model of the bridge structure. It solves the problems such as poor accuracy and time accuracy of structural state diagnosis in the existing methods in the statistical sense, and realizes real-time and reliable diagnosis of the structural state based on the structural response monitoring sequence data.

[0086] The present invention proposes an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the real-time diagnosis method for urban bridge states based on health monitoring sequence data and variational recurrent neural networks are realized.

[0087] The present invention proposes a computer-readable storage medium for storing computer instructions, and when the computer instructions are executed by a processor, the steps of the real-time diagnosis method for urban bridge states based on health monitoring sequence data and variational recurrent neural networks are realized.

[0088] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DRRAM). It should be noted that the memory for the methods described in the present invention is intended to include, but not be limited to, these and any other suitable types of memory.

[0089] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that includes one or more integrated available media. The available media may be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as high-density digital video discs (DVDs)), or semiconductor media (such as solid state discs (SSDs)), etc.

[0090] In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor or the instructions in the form of software. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed by the hardware processor, or executed by a combination of the hardware and software modules in the processor. The software module may be located in a mature storage medium in the art such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, register, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0091] It should be noted that the processor in the embodiments of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, the steps of the above method embodiments can be completed by the integrated logic circuit in the hardware of the processor or instructions in software form. The above-mentioned processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by the hardware decoding processor, or executed and completed by a combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0092] The above has introduced in detail the real-time diagnosis method for the state of urban bridges based on health monitoring sequence data and variational recurrent neural networks proposed by the present invention. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A real-time diagnosis method for the state of urban bridges based on health monitoring sequence data and variational recurrent neural networks, characterized in that The method includes the following steps: Step 1, establish a data set: establish a simulation data set of vehicle load - structural response and a real data set containing only real structural responses; Step 2, establish a model: establish an encoder - decoder decoupled representation learning model based on a variational recurrent neural network; Step 3, design physical constraint conditions: establish constraint conditions for the variational recurrent neural network; in Step 3, the constraint conditions include that the shape of the latent space is the same as the discretized grid, the vehicle load position maintains spatial sparsity on the bridge deck, and the vehicle load amplitude and acting position are continuous in the time domain, and the ways of implicitly and explicitly constraining the time continuity of the load amplitude and position; Step 4, pre - training: train the encoder - decoder decoupled representation learning model of the variational recurrent neural network based on the simulation data set; Step 5, fine - tuning: freeze the parameters of the encoder part of the variational recurrent neural network, and fine - tune the parameters of the encoder part of the variational recurrent neural network based on the real structural response monitoring data set; Step 6, structural state diagnosis: freeze all parameters, based on the real structural response monitoring sequence data, gradually deduce the future response values, and in real - time infer the change of the bridge structure state according to the residual between the deduced value and the real value.

2. The method according to claim 1, wherein In Step 1, for the bridge structure to be analyzed, establish its finite element model, load the measured sequence data of vehicle loads and the sampled sequence data based on the vehicle load distribution into the finite element model, and obtain the structural response sequence data corresponding to the measuring points of the whole bridge. If there are C structural response sensors in the whole bridge, the size of the obtained structural response sequence data is C×T, and the length of the vehicle load sequence data is T.

3. The method according to claim 2, wherein In Step 1, evenly discretize the bridge deck into M×N discrete grids along the longitudinal and transverse directions of the bridge, where M is the number of longitudinal grids and N is the number of transverse grids; according to the vehicle load sequence data and the position of the vehicle on the bridge deck, formulate the corresponding vehicle load matrix at each moment, the matrix size is T×M×N, and for the grids and moments without vehicle action, fill them with zero values.

4. The method according to claim 1, wherein In step two, the vehicle load on the bridge deck is represented by a triple, namely time, position, and amplitude. When using a discrete grid format, the triple is simplified to (t, W t ), where W t ∈R N×M is the vehicle load matrix of the shape defined on the discrete grid, which contains the position and load magnitude information of the vehicle; in SHM, the structural response monitoring data X t ∈R C is regarded as the structural response at each time step t under the action of the vehicle load samples W t ~P(W, t). X t = G(W t ) Where t = 1, 2,..., T, T is the length of the monitoring time, and C is the number of channels; In the variational recurrent neural network, at each time step t, a variational auto - encoder module is established, which consists of a structural response monitoring data embedder encoder Enc, a vehicle load embedder decoder Dec; among them, the embedder maps the structural response monitoring data and vehicle load data to the same dimension as the recurrently connected hidden space state; the role of the encoder is to infer the vehicle load distribution P(W;t) corresponding to time step t from the structural response monitoring sequence data, and the role of the decoder is to generate the reconstructed structural response monitoring sequence data from the sampled inferred load The goal of the variational auto - encoder module is to minimize the reconstruction error When using variational inference, the optimization goal is to minimize the evidence lower bound ELBO: Among them, D KL (P φ (W t |X t )||P(W,t)) is the KL divergence between the inferred vehicle load distribution and the true distribution; The recurrent connection in the variational recurrent neural network is implemented by a gated neural unit GRU, and its state update method is: Then the inference calculation method based on the recurrent connection of the load distribution is: wherein Among them, f I (·) represents an inference module composed of the encoder Enc, and μ t ∈R N×M and σ t ∈R N×M are respectively the mean and covariance matrix of the inferred vehicle load distribution, M is the number of grids in the longitudinal direction of the bridge, and N is the number of grids in the transverse direction of the bridge; The calculation method based on the generation process of the inferred load is: Among them Among them, f G (·) represents the generation module composed of the decoder Dec, μ x,t and σ x,t are respectively the distribution parameters of the reconstructed structure response monitoring sequence data. During the generation process, directly let the mean represent the reconstructed data On this basis, the loss function of the variational recurrent neural network model is to minimize the lower bound of evidence L of the entire sequence VRNN and the reconstruction error L recons :

5. The method according to claim 1, wherein In Step 4, normalize the maximum value of the structural response monitoring sequence data according to the channels, that is, for different types of monitoring data, respectively select the structural response value of the channel and moment with the largest absolute value for maximum value normalization.

6. The method according to claim 1, wherein In Step 4, perform random masking on the input structural response monitoring sequence data according to the channels, that is, in each iteration step of training, randomly select a certain proportion of the channels of the structural response monitoring data for masking, and perform the normal pre - training process.

7. The method according to claim 1, wherein In Step 6, by analyzing the residual between the measured sequence data of the bridge structure response in the healthy state and the predicted sequence data of the variational recurrent neural network, calculate the mean m and standard deviation σ of the reconstruction residual in the healthy state based on the control chart, and use m±3σ as the upper and lower thresholds for real - time inference of the bridge structure state.

8. An electronic device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1-7 are implemented.

9. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, the steps of the method according to any one of claims 1-7 are implemented.

Images