Bridge vehicle load sensing method and system based on encoder-decoder

Abstract

The invention discloses a bridge vehicle load sensing method and system based on an encoder-decoder, and the method comprises the steps: building a variational inference model based on the encoder-decoder, constraining the shape of a hidden feature space in the middle of the encoder-decoder to be a discrete form of a bridge floor, an encoder part in the self-encoder is regarded as a variational inference model of a bridge deck load, and a decoder part is regarded as a constraint part of a training encoder network; building a finite element model and loading an actually measured vehicle load sequence to obtain a response data set of vehicle load and structure dynamic deformation, and taking the response data set as a supervised data set; structural deformation data of real operation of the bridge is collected as a fine tuning data set, and the result of the pre-training stage is docked for a subsequent fine tuning stage; carrying out pre-training and fine tuning on the network model; and inputting real-time bridge structure deformation to deduce the vehicle load in the characteristic space. The problem that real-time distribution of bridge deck vehicle loads cannot be obtained through a traditional method is solved, and the model can be rapidly deployed.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing and structural health monitoring, and in particular relates to a bridge vehicle load sensing method and system based on an encoder-decoder, which can be directly applied to technical fields such as smart infrastructure and bridge engineering. Background Art

[0002] Vehicle loads are the main type of loads borne by bridge structures. Accurately acquiring their data is of great significance for bridge structure health monitoring. However, due to the limitations of measurement methods in the past, real-time vehicle load information on the bridge deck is like being hidden in the fog, and it is extremely difficult to accurately capture. Traditional measurement methods are not only cumbersome to operate, but also often accompanied by drawbacks such as poor accuracy and discontinuous data collection, which are difficult to meet the stringent requirements of current refined bridge management and real-time safety warnings.

[0003] In recent years, the matching scheme of dynamic weighbridge data and computer vision technology has emerged. With the help of cameras to capture vehicle video streams, combined with information such as vehicle weight and image features at the dynamic weighbridge, the real-time position of the vehicle on the bridge is located. Although this method has made some achievements in some scenarios, it has shortcomings when facing a large number of small and medium-sized urban bridges that are not equipped with dynamic weighbridge equipment. Relying solely on video image data, it is limited by factors such as light changes, occlusion, image resolution, and complex and diverse vehicle models, making it extremely difficult to predict accurate vehicle load information, and the prediction error remains high, which greatly reduces the practicality of this method in small and medium-sized bridge scenarios.

[0004] At the same time, research in the field of bridge structural mechanics has revealed that the dynamic deformation caused by vehicle loads acting on bridges is like a "code book" that records key information about vehicles and structures, which contains rich load details and clues to the structural status. If the vehicle load can be accurately decoupled from the inherent information of the bridge structure from the structural deformation data, it will not only be possible to achieve high-precision vehicle load perception, but also to separate the real-time status information of the structure, opening up a new path for bridge health diagnosis. Therefore, it is urgent to develop a method that can accurately infer the vehicle load on the bridge deck in real time based only on response data such as structural deformation. This will not only fill the gap in existing technologies in the field of vehicle load monitoring on small and medium-sized bridges, but also meet the demands of the times for smart infrastructure construction to move towards intelligence and refinement. Summary of the invention

[0005] The present invention aims to solve the problem of real-time and accurate perception of vehicle load on the bridge deck of small and medium-sized bridges in cities. The real-time inference of vehicle load on the bridge deck is performed based on structural deformation data. A model is proposed which can realize the inference based only on response data such as structural deformation. The model can be widely used in the real-time and accurate perception of vehicle load on bridge decks without dynamic weighbridges, and overcomes the matching error based on computer vision methods.

[0006] In view of the above defects or improvement needs of the prior art, according to the first aspect of the present invention, the present invention proposes a bridge vehicle load sensing method based on an encoder-decoder, and the specific steps of the method include: S1. Establish a variational inference model based on encoder-decoder, constrain the shape of the hidden feature space between encoder and decoder to be the discrete form of the bridge deck, regard the encoder part of the autoencoder as the variational inference model of the bridge deck load, and regard the decoder part as the constraint part for training the encoder network; and introduce an additional detector in the encoder part of the network to identify the number of vehicles on the current bridge deck; S2. Build a finite element model, load the measured vehicle load sequence into the finite element model, and obtain the response data set of vehicle load and structural dynamic deformation as a supervised data set; and collect the structural deformation data of the actual operation of the bridge as a fine-tuning data set, connect it with the results of the pre-training stage, and use it for the subsequent fine-tuning stage; S3. Carry out pre-training and fine-tuning of the network model. The specific steps are as follows: (1) In the pre-training stage, in addition to the loss functions such as variational inference loss, reconstruction error loss, and parameter regularization term, a vehicle load supervised loss function is added to guide the encoder network to converge quickly; (2) After pre-training, the network parameters of the detector, encoder, and decoder are frozen for real structural deformation data, and an output layer is added between the encoder and hidden layer features. The vehicle load is inferred based on the encoder state. During the fine-tuning phase, only the network parameters of the output layer are trained. The loss function during the fine-tuning phase only includes the variational inference loss, reconstruction error loss, and parameter regularization term. S4. The real-time bridge structure deformation is input to infer the vehicle load in the feature space. The real-time and accurate perception results of the vehicle load are expressed in the form of activation values ​​on the discretized grid after top-k activation and visualized by a histogram.

[0007] Furthermore, the discretization form of the bridge deck in S1 is: the bridge deck is discretized according to the transverse direction and the longitudinal direction to obtain a characteristic space size M×N corresponding to the discretized grid shape of the bridge deck, where M is the number of discrete grids in the longitudinal direction and N is the number of discrete grids in the transverse direction.

[0008] Furthermore, the encoder input in S1 is the structural deformation response , the output is the mean and variance Real-time inference results of vehicle load distribution composed of matrices: , where the mean and variance The shape of the matrix is ​​M×N. Indicates that the input is The encoder network.

[0009] Furthermore, the detector described in S1 is specifically as follows: Considering the sparseness of vehicle load on the bridge deck, The output of is sparsely activated using the top-k activation method. An additional detector is introduced in the encoder part of the network to identify the number of vehicles on the current bridge deck. The largest k items are selected through a softmax activation function. k is the output result of the detector, and the input of the softmax activation function is the mean. , the non-zero items of the output can be expressed as: .

[0010] Furthermore, the input of the decoder in S1 is the vehicle load value inferred by the encoder, and the output is the structural deformation reconstruction data ,in For sampling of inferred vehicle loads.

[0011] Furthermore, the loss function of the variational inference model of the encoder-decoder in S1 includes the variational inference loss of the real-time vehicle load on the bridge, the reconstruction error loss, and the parameter regularization term; The loss function term for variational inference is:

[0012] in, and Represent the parameters of the encoder and decoder networks respectively, represents the lower bound of evidence, Represents the KL divergence between the inferred distribution and the prior Gaussian distribution; The loss of the decoder reconstruction error is:

[0013] That is, the original data of structural deformation input by the encoder and the deformed data reconstructed by the decoder The mean square error loss.

[0014] Furthermore, the supervised pre-training datasets in S2 include: vehicle load and structural deformation Equal response data set: Structural deformation data The dimension is C × T, where C represents the number of structural deformation response sensors and T represents the monitoring sequence length; Vehicle load Discretize the bridge deck and, based on the discretized grid, identify the locations of zero and non-zero vehicle weight items on the bridge deck grid.

[0015] Furthermore, the fine-tuning dataset in S2 uses the real response of the structure and the structural deformation data The dimension is consistent with the pre-training dataset, C×T, where C represents the number of structural deformation response sensors and T represents the length of the monitoring sequence.

[0016] According to a second aspect of the present invention, there is provided a bridge vehicle load sensing system based on an encoder-decoder, comprising: The variational inference model building unit of the encoder-decoder is used to build a variational inference model based on the encoder-decoder, constrain the shape of the hidden feature space between the encoder and the decoder to be a discrete form of the bridge deck, regard the encoder part in the autoencoder as a variational inference model of the bridge deck load, and regard the decoder part as a constraint part for training the encoder network; The data set construction unit for model training is used to build a finite element model, load the measured vehicle load sequence into the finite element model, and obtain the response data set of vehicle load and structural dynamic deformation as a supervised data set; and collect the structural deformation data of the actual operation of the bridge as a fine-tuning data set, which is connected to the results of the pre-training stage and used in the subsequent fine-tuning stage; The network model pre-training and fine-tuning unit is used to carry out the pre-training and fine-tuning phase of the network model. The specific steps are as follows: (1) In the pre-training stage, in addition to the loss functions such as variational inference loss, reconstruction error loss, and parameter regularization term, a vehicle load supervised loss function is added to guide the encoder network to converge quickly; (2) After pre-training, the network parameters of the detector, encoder, and decoder are frozen for real structural deformation data, and an output layer is added between the encoder and hidden layer features. The vehicle load is inferred based on the encoder state. During the fine-tuning phase, only the network parameters of the output layer are trained. The loss function during the fine-tuning phase only includes the variational inference loss, reconstruction error loss, and parameter regularization term. The vehicle load real-time identification unit is used to input the real-time bridge structure deformation to infer the vehicle load in the feature space. The real-time and accurate perception results of the vehicle load are expressed in the form of activation values ​​on a discretized grid after top-k activation and visualized by a histogram.

[0017] As a third aspect of the present invention, it also relates to a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor for a bridge vehicle load sensing method based on an encoder-decoder.

[0018] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art: 1. The variational inference method for accurate perception of vehicle loads on bridges based on the encoder-decoder architecture of the present invention constructs a unique bridge deck vehicle load inference model based on the encoder-decoder architecture and variational inference, and takes multi-source and multi-type real-time response monitoring data of bridge structures as input. It can realize inference based only on response data such as structural deformation, and overcomes the matching error of computer vision-based methods when applied to small and medium-sized urban bridges without dynamic weighbridges. In addition, the model can be widely used in scenarios of accurate real-time vehicle load perception of such bridges.

[0019] 2. The present invention adopts a variational inference method for accurate perception of vehicle loads on bridges based on an encoder-decoder architecture. By adopting the innovative training paradigm of "finite element data pre-training + real data fine-tuning", it uses a real-time vehicle load-structure dynamic deformation pre-training dataset generated based on bridge structure finite element model simulation data, as well as an acquired real structure dynamic deformation fine-tuning dataset. This allows the method to be quickly deployed to different bridge projects based on a small number of samples of real monitoring data, greatly reducing the dependence of model training on a large number of samples, shortening the deployment cycle, and improving practicality and adaptability.

[0020] 3. The present invention is based on the accurate perception variational inference method of bridge vehicle load based on the encoder-decoder architecture. Through carefully designed network architecture, the shape of the hidden feature space between the encoder and decoder is constrained to be the discrete form of the bridge deck, and a sparse activation mechanism is introduced. The variational inference result of the real-time vehicle load on the bridge is subject to the maximum evidence lower bound, reconstruction error and parameter regularization term. The loss function can accurately establish the correlation between the structural response and the bridge deck load, which solves the difficult problems that the traditional method cannot obtain the real-time distribution of the vehicle load on the bridge deck and the intelligent load identification model requires a lot of manual labeling. It provides key technical support for bridge structure health monitoring, smart infrastructure construction and refined operation and maintenance of bridge projects, and strongly promotes the industry to develop in the direction of intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flowchart of the variational inference method for accurate bridge vehicle load perception based on encoder-decoder architecture for better implementation; Figure 2 Schematic diagram of the variational inference network architecture for accurate bridge vehicle load perception based on the encoder-decoder architecture for better implementation; Figure 3 The bridge deck discretization grid and encoder output feature space shape and output format for the preferred implementation; Figure 4 Detailed architecture of the encoder module for the preferred implementation; Figure 5 Detailed architecture of the detector and sparse activation module for better implementation; Figure 6The detailed architecture of the decoder module for the preferred implementation; Figure 7 Finite element model for better implementation of vehicle load-cable force response simulation; Figure 8 Supervised pre-training datasets (structural response data) for better implementation of finite element simulations; Fig. 9 Supervised pre-training datasets (vehicle load data) for better implementation of finite element simulations; Fig.10 A dataset of real structural responses for unsupervised fine-tuning of better implementations (Structural Response Data); Fig.11 Real-time inference results of vehicle position for better implementation; Fig.12 Real-time inference results of vehicle load position and vehicle weight for better implementation; Fig.13 A schematic diagram of a system unit according to a preferred embodiment. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Example 1 Please refer to Figure 1 This embodiment provides a bridge vehicle load accurate perception variational inference method based on an encoder-decoder architecture, including: Please refer to Figure 2 The network architecture proposed in this embodiment 1 includes an encoder, a decoder and a detector. The encoder infers the vehicle load in real time based on the input structural deformation data, and the decoder reconstructs the structural deformation based on the inferred vehicle load; the feature space shape of the autoencoder is consistent with the discretized grid shape of the bridge deck, the vehicle load is sparsely distributed on the grid, and the number of vehicles on the bridge deck is predicted by the structural deformation data through the detector.

[0024] The training process of the proposed method consists of two stages: pre-training with finite element simulation data and fine-tuning with real data. The pre-training stage is a supervised data set, and the loss function includes vehicle load supervised loss, vehicle load variational inference loss, reconstruction error loss and parameter regularization loss; the fine-tuning stage is an unsupervised data set, and the loss function includes vehicle load variational inference loss, reconstruction error loss and parameter regularization loss.

[0025] The specific steps of the method in this embodiment are as follows: Step 1. Establish a variational inference model based on encoder-decoder, constrain the shape of the hidden feature space between encoder and decoder to be the discrete form of the bridge deck, regard the encoder part in the autoencoder as the variational inference model of the bridge deck load, and regard the decoder part as the constraint part for training the encoder network; Step 1.1: Establish a vehicle load variational inference encoder, which includes the following steps: Step 1.1.1: Based on the encoder-decoder architecture, the state value of the hidden layer feature space is used as the variational inference result of the real-time vehicle load on the bridge; Step 1.1.2: Consider the vehicle load as a mobile concentrated load. For small and medium-sized urban bridges, discretize the bridge deck according to the transverse and longitudinal directions, and obtain the characteristic space size M×N corresponding to the discretized grid shape of the bridge deck, where M is the number of discrete grids in the longitudinal direction and N is the number of discrete grids in the transverse direction (N can be set as the number of lanes). Figure 3 As shown; Step 1.1.3: Please refer to Figure 4 , based on the encoder and variational inference, the real-time vehicle load on the bridge deck is obtained. The input of the encoder is the structural deformation response , the output is the mean and variance Real-time inference results of vehicle load distribution composed of matrices: , where the mean and variance The shape of the matrix is ​​M×N. Indicates that the input is The encoder network of Step 1.1.4: Please refer to Figure 5 , considering the sparseness of vehicle load on the bridge deck, The output of is sparsely activated using the top-k activation method. An additional detector is introduced in the encoder part of the network to identify the number of vehicles on the current bridge deck. The largest k items are selected through a softmax activation function, where k is the output result of the detector and the input of the softmax activation function is the mean. , the non-zero items of the output can be expressed as: .

[0026] Step 1.2 establishes a structure response decoder, which specifically includes the following steps: Step 1.2.1: Please refer to Figure 6 The decoder input is the vehicle load value inferred by the encoder, and the output is the structural deformation reconstruction data ,in For sampling of inferred vehicle loads; Step 1.2.2: The loss function of the encoder-decoder network includes the variational inference loss of the real-time vehicle load on the bridge, the reconstruction error loss, and the parameter regularization term; Step 1.2.3: The loss function term of variational inference is

[0027] in, and Represent the parameters of the encoder and decoder networks respectively, represents the lower bound of evidence, Represents the KL divergence between the inferred distribution and the prior Gaussian distribution; Step 1.2.4: The loss of the decoder reconstruction error is: , that is, the original data of structural deformation input by the encoder and the deformed data reconstructed by the decoder The mean square error loss.

[0028] Step 2. Build a finite element model, load the measured vehicle load sequence into the finite element model, and obtain the response data set of vehicle load and structural dynamic deformation as a supervised data set; collect the structural deformation data of the actual operation of the bridge, and use it as a fine-tuning data set to connect the results of the pre-training stage for the subsequent fine-tuning stage, which specifically includes the following steps: Step 2.1: Please refer to Figure 7 , establish its finite element model according to the targeted urban small and medium bridge structure. Please refer to Figure 8 as well as Fig. 9 , load the measured vehicle load sequence onto the finite element model to obtain the vehicle load and structural deformation Equal response datasets are used as supervised pre-training datasets for pre-training of the encoder part; Step 2.2: Structural deformation data in finite element simulation pre-training data The dimension is C × T, where C represents the number of structural deformation response sensors and T represents the monitoring sequence length; Step 2.3: Vehicle load in finite element simulation pre-training data Perform data conversion according to step 1.1.2, and identify the locations of zero and non-zero vehicle weight items on the bridge deck grid based on the discretized grid; Step 2.4: When fine-tuning with real data, use the real response of the structure as the fine-tuning dataset, as shown in the attached Fig.10 As shown, the structural deformation data The dimension is consistent with the pre-training processing method, C×T, where C represents the number of structural deformation response sensors and T represents the length of the monitoring sequence.

[0029] Step 3. Carry out the pre-training and fine-tuning phase of the network model. The specific steps are as follows: Step 3.1: Considering the large scale of the network and the difficulty of converging from scratch training, during the pre-training of the encoder network, the pre-training loss function is set to include loss terms such as variational inference loss, reconstruction error loss, and parameter regularization term, and a vehicle load supervised loss function is added to guide the encoder network to converge quickly; Step 3.2: After pre-training, freeze the detector, encoder, and decoder network parameters for real structural deformation data, add an output layer between the encoder and hidden layer features, and perform vehicle load inference based on the encoder state. In the fine-tuning stage, only the output layer network parameters are trained; Step 3.3: The loss function in the fine-tuning phase only contains loss terms such as variational inference loss, reconstruction error loss, and parameter regularization term.

[0030] Step 4. Input the real-time bridge structure deformation to infer the vehicle load in the feature space. The real-time and accurate perception result of the vehicle load is expressed in the form of activation values ​​on the discretized grid after top-k activation and visualized by a histogram. The specific steps are as follows: Step 4.1: Freeze the trained and converged encoder-decoder variational inference model, and input the real-time bridge structure deformation to infer the vehicle load in the feature space; Step 4.2: Please refer to Fig.11 as well as Fig.12 ,The real-time and accurate perception results of vehicle load are expressed in the form of activation values ​​on the discretized grid after top-k activation, and are visualized by histogram; Step 4.3: The accuracy of the real-time and precise perception results of vehicle load can also be reflected by the deformation error of the reconstructed structure of the decoder.

[0031] Example 2 Please refer to Fig.13 , this embodiment 2 provides a bridge vehicle load accurate perception variational inference system based on an encoder-decoder architecture, including: The variational inference model building unit of the encoder-decoder is used to build a variational inference model based on the encoder-decoder, constrain the shape of the hidden feature space between the encoder and the decoder to be a discrete form of the bridge deck, regard the encoder part in the autoencoder as a variational inference model of the bridge deck load, and regard the decoder part as a constraint part for training the encoder network; The data set construction unit for model training is used to build a finite element model, load the measured vehicle load sequence into the finite element model, and obtain the response data set of vehicle load and structural dynamic deformation as a supervised data set; and collect the structural deformation data of the actual operation of the bridge as a fine-tuning data set, which is connected to the results of the pre-training stage and used in the subsequent fine-tuning stage; The network model pre-training and fine-tuning unit is used to carry out the pre-training and fine-tuning phase of the network model. The specific steps are as follows: (1) In the pre-training stage, in addition to the loss functions such as variational inference loss, reconstruction error loss, and parameter regularization term, a vehicle load supervised loss function is added to guide the encoder network to converge quickly; (2) After pre-training, the network parameters of the detector, encoder, and decoder are frozen for real structural deformation data, and an output layer is added between the encoder and hidden layer features. The vehicle load is inferred based on the encoder state. During the fine-tuning phase, only the network parameters of the output layer are trained. The loss function during the fine-tuning phase only includes the variational inference loss, reconstruction error loss, and parameter regularization term. The vehicle load real-time identification unit is used to input the real-time bridge structure deformation to infer the vehicle load in the feature space. The real-time and accurate perception results of the vehicle load are expressed in the form of activation values ​​on the discretized grid after topk activation and visualized by a histogram.

[0032] Example 3 This embodiment 3 also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, a variational inference method for accurate perception of bridge vehicle loads based on an encoder-decoder architecture can be implemented.

[0033] The computer-readable storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0034] For an introduction to the computer-readable storage medium provided in this application, please refer to the above method embodiment, and this application will not go into details here.

[0035] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A bridge vehicle load sensing method based on encoder-decoder, characterized in that: include: S1. Establish a variational inference model based on encoder-decoder, constrain the shape of the hidden feature space between encoder and decoder to be the discrete form of the bridge deck, regard the encoder part of the autoencoder as the variational inference model of the bridge deck load, and regard the decoder part as the constraint part for training the encoder network; and introduce an additional detector in the encoder part of the network to identify the number of vehicles on the current bridge deck; S2. Build a finite element model, load the measured vehicle load sequence into the finite element model, and obtain the response data set of vehicle load and structural dynamic deformation as a supervised data set; The structural deformation data of the actual operation of the bridge is collected and used as a fine-tuning dataset to connect with the results of the pre-training stage for the subsequent fine-tuning stage; S3. Carry out pre-training and fine-tuning of the network model. The specific steps are as follows: (1) In the pre-training stage, in addition to the loss functions such as variational inference loss, reconstruction error loss, and parameter regularization term, a vehicle load supervised loss function is added to guide the encoder network to converge quickly; (2) After pre-training, the network parameters of the detector, encoder, and decoder are frozen for real structural deformation data, and an output layer is added between the encoder and hidden layer features. The vehicle load is inferred based on the encoder state. During the fine-tuning phase, only the network parameters of the output layer are trained. The loss function during the fine-tuning phase only includes the variational inference loss, reconstruction error loss, and parameter regularization term. S4. The real-time bridge structure deformation is input to infer the vehicle load in the feature space. The real-time and accurate perception results of the vehicle load are expressed in the form of activation values ​​on the discretized grid after top-k activation and visualized by a histogram.

2. A bridge vehicle load sensing method based on encoder-decoder according to claim 1, characterized in that: The discretization form of the bridge deck in S1 is as follows: the bridge deck is discretized according to the transverse and longitudinal directions to obtain a characteristic space size M×N corresponding to the discretized grid shape of the bridge deck, where M is the number of discrete grids in the longitudinal direction and N is the number of discrete grids in the transverse direction.

3. A bridge vehicle load sensing method based on encoder-decoder according to claim 2, characterized in that: The encoder input in S1 is the structural deformation response , the output is the mean and variance Real-time inference results of vehicle load distribution composed of matrices: , where the mean and variance The shape of the matrix is ​​M×N. Indicates that the input is The encoder network.

4. A bridge vehicle load sensing method based on encoder-decoder according to claim 3, characterized in that: The detector described in S1 is as follows: Considering the sparseness of vehicle load on the bridge deck, The output of is sparsely activated using the top-k activation method. An additional detector is introduced in the encoder part of the network to identify the number of vehicles on the current bridge deck. The largest k items are selected through a softmax activation function. k is the output result of the detector, and the input of the softmax activation function is the mean. , the non-zero items of the output can be expressed as: .

5. A bridge vehicle load sensing method based on encoder-decoder according to claim 4, characterized in that: The input of the decoder in S1 is the vehicle load value inferred by the encoder, and the output is the structural deformation reconstruction data ,in For sampling of inferred vehicle loads.

6. A bridge vehicle load sensing method based on encoder-decoder according to claim 5, characterized in that: The loss function of the variational inference model of the encoder-decoder in S1 includes the variational inference loss of the real-time vehicle load on the bridge, the reconstruction error loss, and the parameter regularization term; The loss function term for variational inference is: ， in, and Represent the parameters of the encoder and decoder networks respectively, represents the lower bound of evidence, Represents the KL divergence between the inferred distribution and the prior Gaussian distribution; The loss of the decoder reconstruction error is: ， That is, the original data of structural deformation input by the encoder and the deformed data reconstructed by the decoder The mean square error loss.

7. A bridge vehicle load sensing method based on encoder-decoder according to claim 1, characterized in that: The supervised pre-training datasets in S2 include: vehicle load and structural deformation Equal response data set: Structural deformation data The dimension is C × T, where C represents the number of structural deformation response sensors and T represents the monitoring sequence length; Vehicle load Discretize the bridge deck and, based on the discretized grid, identify the locations of zero and non-zero vehicle weight items on the bridge deck grid.

8. A bridge vehicle load sensing method based on encoder-decoder according to claim 7, characterized in that: The fine-tuning dataset in S2 uses the real response of the structure and the structural deformation data The dimension is consistent with the pre-training dataset, C×T, where C represents the number of structural deformation response sensors and T represents the length of the monitoring sequence.

9. A bridge vehicle load sensing system based on encoder-decoder, characterized in that: include: The variational inference model building unit of the encoder-decoder is used to build a variational inference model based on the encoder-decoder, constrain the shape of the hidden feature space between the encoder and the decoder to be a discrete form of the bridge deck, regard the encoder part in the autoencoder as a variational inference model of the bridge deck load, and regard the decoder part as a constraint part for training the encoder network; The dataset construction unit for model training is used to build a finite element model, load the measured vehicle load sequence into the finite element model, and obtain the response dataset of vehicle load and structural dynamic deformation as a supervised dataset; The structural deformation data of the actual operation of the bridge is collected and used as a fine-tuning dataset to connect with the results of the pre-training stage for the subsequent fine-tuning stage; The network model pre-training and fine-tuning unit is used to carry out the pre-training and fine-tuning phase of the network model. The specific steps are as follows: (1) In the pre-training stage, in addition to the loss functions such as variational inference loss, reconstruction error loss, and parameter regularization term, a vehicle load supervised loss function is added to guide the encoder network to converge quickly; (2) After pre-training, the network parameters of the detector, encoder, and decoder are frozen for real structural deformation data, and an output layer is added between the encoder and hidden layer features. The vehicle load is inferred based on the encoder state. During the fine-tuning phase, only the network parameters of the output layer are trained. The loss function during the fine-tuning phase only includes the variational inference loss, reconstruction error loss, and parameter regularization term. The vehicle load real-time identification unit is used to input the real-time bridge structure deformation to infer the vehicle load in the feature space. The real-time and accurate perception results of the vehicle load are expressed in the form of activation values ​​on a discretized grid after top-k activation and visualized by a histogram.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the encoder-decoder based bridge vehicle load sensing method described in claims 1-8.

Images