Sensor location selection method, system, device and medium based on bridge monitoring

Abstract

The present invention belongs to the technical field of bridge monitoring. In order to solve the problem of how to determine the key positions of sensor deployment in a bridge structure for bridge reconstruction, a sensor position selection method, system, device and medium based on bridge monitoring are proposed, and an adaptive position selection mechanism is proposed. By generating a selection matrix of binary classification distribution for the positions of interest of the bridge, the key sensor positions are determined according to the on or off of the sensor at each position of interest, and the displacement response of the unselected positions of interest is reconstructed based on the measurement data at the key sensor positions through a neural network backbone. The sensor position selection and response reconstruction are consistently learned in a unified deep learning framework. Experiments have verified that the model proposed by the present invention is effective and accurate in reconstructing the displacement response of the bridge structure through adaptive sensor selection.

Description

Technical Field

[0001] The present invention belongs to the technical field related to bridge monitoring, and in particular relates to a sensor location selection method, system, equipment and medium based on bridge monitoring. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of urbanization and large-scale infrastructure, the integrity of bridges is critical to public safety and urban development. Traditional bridge monitoring methods usually rely on periodic inspections and manual analysis of structural behavior by experts, which is difficult to meet the development needs of modern infrastructure. With the development of sensing, communication and data analysis capabilities, structural health monitoring systems that can capture changes in bridge structural response online have been widely used in the field of bridge engineering. Bridge structural health monitoring, which makes automated condition assessment possible, has become an indispensable technology in bridge assessment and maintenance, which can achieve effective damage detection and condition assessment.

[0004] As the main component of the bridge structural health monitoring system, the sensors deployed on the bridge play a particularly important role in the entire system. Considering the large scale and complex structure of the bridge, it is impossible to install as many sensors as possible at each location of interest on the bridge. In practical applications, it is challenging to determine their installation locations due to the limited number of sensors. In order to obtain better monitoring performance, it is very important to select the key locations of sensors on the bridge based on the finite element analysis of the bridge structure.

[0005] In summary, how to select the key locations for sensor deployment on the bridge structure in order to accurately and effectively reconstruct the bridge response at other unmeasured locations based on the determined key locations of the sensors is a problem that needs to be solved at present. Summary of the invention

[0006] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a sensor position selection method, system, device and medium based on bridge monitoring, which effectively and accurately determine the key sensor positions and reconstruct the displacement response through an adaptive position selection mechanism and a neural network backbone.

[0007] To achieve the above object, a first aspect of the present invention provides a sensor location selection method based on bridge monitoring, the method comprising:

[0008] Based on the acquired locations of interest of the bridge to be deployed and the multivariate time series data measured at the locations of interest, the trained location selection and response reconstruction network model is used to obtain the key sensor locations of the bridge to be deployed;

[0009] The training process of the position selection and response reconstruction network model is specifically as follows:

[0010] An adaptive position selection mechanism is used to generate a selection matrix of binary classification distribution for the positions of interest of the bridge, and a candidate key sensor position for reconstruction is determined by turning on or off the sensor corresponding to each position of interest of the bridge in the selection matrix;

[0011] Based on the multivariate time series data measured at the determined candidate key sensor locations, the neural network backbone is used to reconstruct the bridge response at the unmeasured locations of the bridge. The location selection and response reconstruction network model is trained by minimizing the difference between the obtained bridge response and the true displacement response of the bridge.

[0012] A second aspect of the present invention provides a sensor location selection system based on bridge monitoring, the system comprising:

[0013] A sensor location determination module is configured to: obtain the key sensor locations of the bridge to be deployed based on the acquired locations of interest of the bridge to be deployed and the multivariate time series data measured at the locations of interest, and use the trained location selection and response reconstruction network model to obtain the key sensor locations of the bridge to be deployed;

[0014] The model training module is configured as follows: the training process of the position selection and response reconstruction network model is specifically as follows:

[0015] An adaptive position selection mechanism is used to generate a selection matrix of binary classification distribution for the positions of interest of the bridge, and a candidate key sensor position for reconstruction is determined by turning on or off the sensor corresponding to each position of interest of the bridge in the selection matrix;

[0016] Based on the multivariate time series data measured at the determined candidate key sensor locations, the neural network backbone is used to reconstruct the bridge response at the unmeasured locations of the bridge. The location selection and response reconstruction network model is trained by minimizing the difference between the obtained bridge response and the true displacement response of the bridge.

[0017] The third aspect of the present invention provides a computer device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, a sensor location selection method based on bridge monitoring is executed.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, a sensor location selection method based on bridge monitoring is executed.

[0019] One or more of the above technical solutions have the following beneficial effects:

[0020] In the present invention, an adaptive position selection mechanism is proposed. By generating a selection matrix of binary classification distribution for the positions of interest of the bridge, the key sensor positions are determined according to the on or off of the sensors at each position of interest. The displacement responses of unselected positions of interest are reconstructed based on the measured data at the key sensor positions through a neural network backbone. The sensor position selection and response reconstruction are consistently learned in a unified deep learning framework. Experiments have verified that the model proposed in the present invention is effective and accurate in reconstructing the displacement response of bridge structures through adaptive sensor selection.

[0021] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0023] Figure 1 It is the overall framework of the sensor placement and displacement reconstruction method in the first embodiment of the present invention;

[0024] Figure 2 The finite element model of the bridge in the first embodiment of the present invention;

[0025] FIG3 (a) is a schematic diagram of a true displacement response in Embodiment 1 of the present invention;

[0026] FIG3( b ) is a schematic diagram of reconstructed displacement response in Embodiment 1 of the present invention;

[0027] FIG3 (c) is a schematic diagram of absolute error in Embodiment 1 of the present invention;

[0028] Figure 4 This is a comparison diagram of absolute errors of the positions of interest in the first embodiment of the present invention. DETAILED DESCRIPTION

[0029] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0030] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0031] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0032] Embodiment 1

[0033] This embodiment mainly uses sensor measurement data to reconstruct bridge displacements at unmeasured locations of interest. Increasing the number of sensors on the bridge can capture more structural information. However, considering the complex structure and scale of the bridge, minimizing the number of sensors and selecting key sensor locations for measurement are crucial in bridge structural health monitoring. In order to achieve efficient bridge structural health monitoring with a limited number of sensors, this embodiment proposes an adaptive sensor selection and reconstruction neural network to determine key sensing locations and reconstruct them at all locations of interest. In particular, a novel adaptive position selection mechanism and a unified network model and its learning process are proposed. In the experiment, a long-span bridge was taken as an example to verify the proposed model. By comparing different sensor layout schemes, the experimental results show that the proposed model is effective and accurate in reconstructing the displacement response of the bridge structure through adaptive sensor selection.

[0034] This embodiment discloses a sensor location selection method based on bridge detection, including:

[0035] Based on the acquired locations of interest of the bridge to be deployed and the multivariate time series data measured at the locations of interest, the trained location selection and response reconstruction network model is used to obtain the key sensor locations of the bridge to be deployed;

[0036] The training process of the position selection and response reconstruction network model is specifically as follows:

[0037] An adaptive position selection mechanism is used to generate a selection matrix of binary classification distribution for the positions of interest of the bridge, and a candidate key sensor position for reconstruction is determined by turning on or off the sensor corresponding to each position of interest of the bridge in the selection matrix;

[0038] Based on the multivariate time series data measured at the determined candidate key sensor locations, the neural network backbone is used to reconstruct the bridge response at the unmeasured locations of the bridge. The location selection and response reconstruction network model is trained by minimizing the difference between the obtained bridge response and the true displacement response of the bridge.

[0039] This embodiment proposes a position selection and response reconstruction network model, a new hybrid deep neural network model, which adaptively selects sensor locations for measurement and reconstructs bridge responses at other unmeasured locations. The most important locations where sensors are placed are defined as key sensing locations, and the unmeasured locations to be reconstructed are defined as locations of interest. The position selection and response reconstruction network model integrates a hybrid neural network backbone and a new adaptive position selection mechanism. Specifically, in the neural network backbone, a one-dimensional convolutional neural network processes time series data and is able to identify local temporal features; the multi-head attention mechanism enables the model to better capture the hierarchical structure, periodic patterns, and local features in time series data; and the gated recurrent unit is used to capture temporal dependencies in time series data. In addition, the designed adaptive position selection mechanism module is proposed for selecting key sensing locations with a limited number of sensors. Then, the model uses the data measured by the sensors at the key sensing locations to reconstruct the structural displacement response at the locations of interest.

[0040] The selection and response reconstruction network model proposed in this embodiment learns the key locations of sensor placement and reconstructs the bridge displacement response at unselected locations of interest. Therefore, the proposed model aims to:

[0041] 1) Sensor location selection by determining key sensing locations;

[0042] 2) Reconstruct the displacement responses at other locations of interest that were not selected.

[0043] Figure 1 It is the overall framework of the model used for sensor position selection and displacement reconstruction in this embodiment. Figure 1 As shown, the key sensing positions are selected as sensor placement positions. The data samples collected by the sensors installed at the selected positions are input into the proposed selection and response reconstruction network model as input data, and the displacement responses of the unselected positions where no sensors are installed can be reconstructed by the proposed selection and response reconstruction network model.

[0044] To achieve the research objectives, the key sensing locations of sensors are first determined based on the locations of interest. The proposed deep neural network implemented based on the designed adaptive location selection mechanism is used to learn the key sensing locations for bridge displacement reconstruction.

[0045] set up represents the number of all locations of interest for sensor placement and response monitoring; Represents time series data measured at all locations of interest; the designed adaptive location selection mechanism generates key sensing locations for placing sensors.

[0046] Specifically, given any sensors, the selected key sensing locations are composed of Specify where to place these sensors.

[0047] The sensor placement selection process based on the adaptive location selection mechanism in the neural network backbone is as follows:

[0048] (1)

[0049] Among them, ALSM is an adaptive location selection mechanism. Indicates the selected The index set of locations. In order from largest to smallest, Representing a collection The number of elements.

[0050] The input data for the selected I key sensing locations consist of multivariate time series data. Indicates that at time step t Captured time series. For example, historical and current displacement data are recorded as , which can be further expressed as:

[0051] (2)

[0052] Where T represents the time window size.

[0053] The bridge response at the unmeasured locations is reconstructed using the sampled data at the selected locations I. The unmeasured locations are represented by the set J Represented by the proposed neural network backbone F , unmeasured position J The reconstruction response at Generates the following:

[0054] (3)

[0055] in, Indicates the unselected locations of interest on the bridge locations. N + N′ = M, N′ Indicates the number of unselected locations of interest. W Represents the neural network parameters that need to be trained.

[0056] To learn the proposed neural network backbone, we minimize the reconstruction response With real response The difference between them is used to train the model, and the formula is as follows:

[0057] (4)

[0058] Using a unified end-to-end framework, the adaptive sensor selection mechanism and the neural network backbone for response reconstruction are trained. During testing, the bridge response at unmeasured locations can be reconstructed using data measured at key sensing locations. The reconstruction is done using the learned neural network and the trained network parameters. The formula is as follows:

[0059] (5)

[0060] The selection and response reconstruction network model proposed in this embodiment is described in detail below:

[0061] The selection and response reconstruction network model proposed in this embodiment includes a neural network backbone, which is used to capture features in structural multivariate time series data. In addition, an adaptive position selection mechanism is designed and integrated into the neural network backbone to determine the key sensing positions for sensor placement; the adaptive position selection mechanism is used to generate a selection matrix of binary classification distribution for the positions of interest of the bridge, and the key sensing positions are selected by the neural network backbone.

[0062] In this embodiment, the neural network backbone structure used in the sensor key position selection part and the displacement reconstruction part is the same.

[0063] In the neural network backbone, multivariate time series measurement data is first input into the convolutional neural network layer to extract local spatiotemporal features. The weight sharing effect of the convolutional layer effectively reduces the model parameters and improves the training efficiency of the network. For a time window of size T, it is divided into T' segments of equal time length. Each segment is processed by a one-dimensional convolution operation, defined as follows;

[0064] (6)

[0065] in, is the feature extracted from the τth time block, τ = 1, 2, …, T′. represents the convolution kernel weight matrix, Represents the time series data of the measurements at the location of interest for the τth time bin.

[0066] Therefore, a positional self-attention mechanism is embedded in the neural network backbone to make up for the limitations of global feature extraction of convolutional neural networks. The one-dimensional convolution output is projected into the feature domain and used as the input of the position self-attention mechanism, which adaptively assigns different weights to input channels oriented at different positions. , key vector Sum value vector express:

[0067] (7)

[0068] in, , and are learnable projection parameters for the query, key, and value matrices. These matrices project their respective vectors into the corresponding subspaces.

[0069] The position self-attention mechanism is as follows:

[0070] (8)

[0071] The superscript T indicates transposition. The purpose of scaling is to prevent the dot product from being too large, which can lead to gradient explosion.

[0072] The multi-head attention mechanism can utilize attention features in parallel and focus on the latent features of information at different positions to facilitate feature extraction.

[0073] Let H represent the number of attention heads, and the position information self-attention output of multiple heads can be expressed as: (9)

[0074] The formula for each head is as follows:

[0075] (10)

[0076] in, , is the output matrix projection, Refers to τ The features extracted from each time block are used as the input of position self-attention.

[0077] Similarity information is learned in different subspaces to fully capture the attention features. The input is divided into H data subspaces, each head focuses on different levels of hidden information to capture features, integrates diverse features, and dynamically updates the weight matrix to obtain the final output.

[0078] Finally, the gated recurrent unit is used to process sequential dynamic patterns in a time window. In the gated recurrent unit, the gate is updated Determines how much information about the previous state to retain. Reset Gate Determines the extent to which previous state information is ignored.

[0079] The formula and update of the gated recurrent unit are as follows:

[0080]

[0081]

[0082]

[0083] (11)

[0084] in, , and is the weight matrix of the parameters, , and is the corresponding bias vector. is the hidden state at the previous time step, is the input of the current time block, is the current memory cell, is the output of the current time block.

[0085] In the sensor key position selection part, based on the same neural network backbone structure as above, representative features in the input time series can be extracted. In order to effectively utilize these features and determine the key sensing positions for sensor placement, this embodiment proposes a novel adaptive position selection mechanism module.

[0086] Considering all M locations of interest on the bridge, an adaptive location selection mechanism is designed to adaptively select input channels among the locations of interest to find out the most critical locations.

[0087] The adaptive position selection mechanism is defined as the selection matrix S ∈ R that generates a binary classification distribution M×2 , defined as follows:

[0088] (12)

[0089] (13)

[0090] in, The normalized exponential function transforms the output to a distribution in [0, 1]. Represents a latent feature matrix Maps to a function of size M×2. In particular, Each element Corresponding to m Select output of each position. Represents input Implicit correlations in and Represents the corresponding weights in the convolutional neural network process, represents the time series data measured at all locations of interest, It is the matrix dot product, which means the elements of the corresponding positions of the matrices are multiplied. The weights in formula (13) are automatically broadcast to the same matrix shape as the time series data. The two matrices are multiplied at the corresponding positions. The superscript T indicates transpose.

[0091] Specifically, ，m = 1, 2, , M represents a binary categorical distribution, where represents the distribution of the open state, represents the distribution of closed states, and Therefore, the output of the adaptive position selection mechanism is defined as the selection switch , whose value is 0 or 1, is represented by:

[0092] (14)

[0093] Among them, 0 represents the off state and 1 represents the on state. The selected or ignored sensing channel corresponding to each input position is determined.

[0094] During the training process, the minimum Value Maximum value corresponding position m* The sensor channel will be designated as off.

[0095] (15)

[0096] in, Represents the parameter value that makes the function value reach the minimum or minimum.

[0097] pass The output of indicates the on and off status of each channel to determine the key positions for reconstruction and thus determine the selected key sensing positions.

[0098] You can select in the following ways:

[0099] (16)

[0100] in, I Represents the set of selected key sensing locations.

[0101] By determining the key sensor locations, the neural network backbone can be used to J The reconstruction result is as follows:

[0102] (17)

[0103] Among them, F represents the neural network backbone. is a gated recurrent unit at a key sensing location in t-T+ 1:t Output.

[0104] Based on the above formula, the target location for sensor placement is determined by the output of the adaptive location selection mechanism.

[0105] During the training process, sensor location selection is mainly based on the adaptive selection of input sensing channels based on the switch output of each training cycle to determine the key sensing locations. This can be based on the designed adaptive location selection mechanism to utilize input information, making it more efficient and interpretable in terms of sensor location selection. The argmin (independent variable minimum) in formula (9) is discrete and non-differentiable, so it cannot be directly operated by gradient descent backpropagation during training. To solve this problem, Gumbel-Softmax is modified to estimate the gradient of the discrete independent variable minimum function. During the backpropagation process, the adaptive location selection mechanism is calculated as follows:

[0106]

[0107] (18)

[0108] Here, G is a random variable sampled from the Gumbel distribution G∼ (0, 1).

[0109] It means that the gumbel distribution and the delta weighting parameter are added to the original softmax. By applying the Softmax function after sampling the Gumbel distribution, the probability distribution of discrete variables can be generated. δ is the weighting parameter. The weighted Gumbel-Softmax is used to ensure the consistency and differentiability of the adaptive position selection mechanism in back propagation.

[0110] δ is a weighting parameter, calculated as follows:

[0111] (19)

[0112] Here, MLP(·) represents a multilayer perceptron. By adding 1 to the denominator, the range of δ is limited to between 0 and 1. represents the time series data measured at all locations of interest. Therefore, weighted Gumbel-Softmax is used to ensure the consistency and differentiability of the adaptive location selection mechanism in back-propagation.

[0113] The neural network backbone proposed in this embodiment includes a neural network backbone for capturing features in structured multivariate time series data.

[0114] The adaptive position selection mechanism module proposed in this embodiment is integrated in the neural network backbone to select the key positions for sensor placement. The module calculates the selection matrix of the key sensor positions by feedforward. During the training process, the key sensor positions for sensor placement are updated after each training cycle using a gradient-based optimization method until the predetermined number of sensors is reached. The goal of the adaptive position selection mechanism module is to find the optimal sensor placement scheme under a limited number of sensors to maximize the accuracy of response reconstruction.

[0115] In this embodiment, the neural network backbone is used to capture the representative features in the multivariate time series data, corresponding to Formula (6)-Formula (11). The adaptive position selection mechanism is integrated to select the key input channel to determine the key sensor position, corresponding to Formula (12)-Formula (16). Therefore, the state of the output channel of the adaptive position selection mechanism is determined by the features captured by the neural network.

[0116] Training and testing process: The neural network backbone is used to capture representative features in multivariate time series data, while the adaptive location selection mechanism is integrated to select key input channels to determine key sensing locations. During the training process, the initial value of the selection matrix for all locations of interest is set to 1, i.e., the open state. At this point, all input sensing channels are considered and input to the neural network backbone. Afterwards, the model is iteratively trained over multiple cycles, and every P cycles, the model that leads to the minimum Location to turn it off In this way, the least important locations corresponding to the selected outputs are excluded from the input data. Finally, the bridge response at the unmeasured locations of interest is reconstructed using the measured data at the selected sensor locations using the adaptive location selection mechanism.

[0117] The reconstruction results are as follows:

[0118] (20)

[0119] in, W represents the network parameters to be trained. In order to learn the proposed neural network model, the root mean square error is selected as the loss function during the training process. The loss function evaluates the difference between the reconstructed response and the true displacement response. t , the reconstruction response is , the true displacement response is expressed as The loss function is calculated as follows:

[0120] (twenty one)

[0121] The trained model can generate online the reconstructed responses at unmeasured locations given the measured data at key sensing locations:

[0122] (twenty two)

[0123] This example uses a long span bridge as a case study. The total length of the bridge is nearly 1,000 meters. The main span is 57 meters long, and the other typical spans are 37 meters long. The superstructure of the bridge is supported by ten prefabricated I-beams, on which a 15.5 cm thick concrete bridge deck is laid. Accelerometers installed on the actual bridge are used to collect structural response data. In this case study, finite element modeling software is used to Figure 2 It is a finite element model of the linear elastic structure of the bridge.

[0124] Based on the established finite element model, time series analysis was performed to generate a large amount of bridge response data. In order to ensure that the finite element model and the response data generated by it are accurate, the finite element model was calibrated and updated with actual recorded data from the 2014 earthquake. The locations of interest for the case study correspond to the locations of the 26 piers (from P1 to P26) on the bridge deck and the mid-span of the bridge (between P13 and P14). The locations of interest are used to select key sensing locations for sensor placement, and there are a total of 27 locations of interest.

[0125] The studied bridge is located in an active region where large earthquakes frequently occur. In the case study, the bridge was excited and tested by actual earthquake ground motions in the region. A set of 90 ground motion records were selected from the database. Table 1 provides the dataset configuration in the experiment. The second column shows the ground motion records in the training set, validation set, and test set. The third column shows the number of samples in each sub-dataset. The last column shows the dimensions of the dataset, including the number of data samples and the number of perception channels.

[0126] Table 1: Dataset configuration

[0127]

[0128] In the experiment, two quantitative indicators are used to evaluate the model performance, including the mean absolute error and RMS error , defined as follows:

[0129]

[0130]

[0131] in, represents the reconstructed displacement value, Indicates the true displacement value. Table 2 shows the hyperparameters of the network.

[0132] Table 2: Network hyperparameters

[0133]

[0134] In order to verify the performance of adaptive sensor location selection, the proposed model is evaluated by setting different numbers of sensors. The output frequency P is set to 5, that is, one input channel is selected to be closed every 5 cycles. No closed channels are input and trained in the model until the number of remaining input channels reaches N, realizing adaptive selection of key positions. Table 3 shows the key sensor locations on the bridge selected under different numbers of sensors.

[0135] Table 3: Selection of key bridge locations

[0136]

[0137] Table 4 shows the reconstruction results of placing different numbers of sensors. The smaller the values ​​of these two evaluation indicators, the better the reconstruction performance. The experimental results show that the more sensors there are, the smaller the reconstruction error. In addition, when 5 sensors are set to determine the key position, the performance of the evaluation indicators is close to that of placing 6 and 7 sensors. Therefore, compared with other sensor number settings, the selection of 5 sensors can achieve relatively effective performance experimental results with an efficient sensor setting (fewer sensors).

[0138] Table 4: Selection of key bridge locations

[0139]

[0140] In order to verify the performance of response reconstruction, the proposed model is compared with four benchmark models for processing time series data, including the cross-channel modeling (XCM) algorithm, a time series analysis algorithm Xceptiontime based on the Xception model architecture, a time series analysis algorithm Inceptiontime based on the Inception model architecture, and a sequence modeling algorithm (Sequencer), where Xception and Inception are two deep convolutional neural network architectures. Table 5 shows the experimental results. Considering the two evaluation indicators, the individual and average performance of the proposed model achieved the lowest error results in the comparison with other models, demonstrating the accuracy and reliability of the proposed method model in bridge displacement reconstruction.

[0141] Table 5: Comparison results of different models

[0142]

[0143] For demonstration purposes, Figures 3(a)-3(c) describe the comparison between the reconstructed displacement of the model and the target displacement. Due to space limitations, only the P15 position of a certain test data is taken as an example. In Figure 3, the straight line represents the true displacement response, and the dashed line represents the absolute error value between the reconstructed displacement and the true displacement. The dotted line represents the reconstructed displacement response. Through Figures 3(a), 3(b), and 3(c), the true displacement and reconstructed displacement of the model over a period of time, as well as the absolute error between the two, can be clearly and intuitively seen. The value of the absolute error is relatively small, indicating that the model can reliably infer the trend of the time series, capture the response details of the structure over time, and the reconstructed displacement has the effect of high accuracy and excellent stability.

[0144] Figure 4 It is a comparison chart of the absolute errors of all the positions of interest in a certain test set. The x-axis represents the time dimension, the y-axis represents the output channel of the position of interest, and the z-axis represents the absolute error. From the figure, the overall spatial distribution of the reconstruction errors at different positions can be intuitively observed, which helps to display the error distribution and error peak. The displacement errors on both sides of the bridge are small, while the displacement errors in the middle of the bridge are large. This is because the vibration amplitude in the middle of the bridge is greater than the vibration amplitude on both sides. Overall, the absolute error value of the model studied in this embodiment is low, and the experimental results show the superior performance of the proposed sensor placement scheme in response reconstruction.

[0145] This embodiment proposes a novel deep learning framework for adaptive sensor placement selection and response reconstruction. A novel deep neural network model is designed to select key sensing locations and reconstruct structural responses at unmeasured locations. In order to capture and reconstruct time series signals, the neural network backbone of the proposed model includes convolutional neural networks, multi-head self-attention mechanisms, and recurrent neural networks. In addition, the proposed adaptive location selection mechanism is integrated into the neural network backbone to determine the key locations for sensor placement. The sensor quantity selection and performance superiority of the proposed method are verified through actual bridge case studies.

[0146] Embodiment 2

[0147] The purpose of this embodiment is to provide a sensor location selection system based on bridge monitoring, the system comprising:

[0148] A sensor location determination module is configured to: obtain the key sensor locations of the bridge to be deployed based on the acquired locations of interest of the bridge to be deployed and the multivariate time series data measured at the locations of interest, and use the trained location selection and response reconstruction network model to obtain the key sensor locations of the bridge to be deployed;

[0149] The model training module is configured as follows: the training process of the position selection and response reconstruction network model is specifically as follows:

[0150] An adaptive position selection mechanism is used to generate a selection matrix of binary classification distribution for the positions of interest of the bridge, and a candidate key sensor position for reconstruction is determined by turning on or off the sensor corresponding to each position of interest of the bridge in the selection matrix;

[0151] Based on the multivariate time series data measured at the determined candidate key sensor locations, the neural network backbone is used to reconstruct the bridge response at the unmeasured locations of the bridge. The location selection and response reconstruction network model is trained by minimizing the difference between the obtained bridge response and the true displacement response of the bridge.

[0152] Embodiment 3

[0153] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0154] Embodiment 4

[0155] The purpose of this embodiment is to provide a computer-readable storage medium.

[0156] A computer-readable storage medium stores a computer program, which executes the steps of the above method when executed by a processor.

[0157] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0158] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0159] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A sensor location selection method based on bridge monitoring, characterized in that: The method comprises: Based on the acquired locations of interest of the bridge to be deployed and the multivariate time series data measured at the locations of interest, the trained location selection and response reconstruction network model is used to obtain the key sensor locations of the bridge to be deployed; The training process of the position selection and response reconstruction network model is specifically as follows: The adaptive location selection mechanism is used to generate a selection matrix of binary classification distribution for the locations of interest of the bridge, specifically: ， ， in, The normalized exponential function transforms the output into a distribution of [0, 1]; Represents a latent feature matrix Mapping to a function of size M×2; , each element Corresponding to m The selected output of each position; Represents input Implicit correlations in and Represents the corresponding weights in the convolutional neural network process, represents time series data measured at all locations of interest; Represents matrix dot product; Determine candidate key sensor positions for reconstruction by turning on or off sensors corresponding to each position of interest of the bridge in the selection matrix; Based on the multivariate time series data measured at the determined candidate key sensor positions, the neural network backbone is used to reconstruct the bridge response at the unmeasured position of the bridge, and the difference between the obtained bridge response and the actual displacement response of the bridge is minimized, and the position selection and response reconstruction network model is trained. In the training of the position selection and response reconstruction network model, in the back propagation process, the calculation of the adaptive position selection mechanism is specifically as follows: ， in, is a switch matrix that determines the selected or ignored sensing channels corresponding to each input position; ALSM stands for Adaptive Location Selection Mechanism; It means that the Gumbel distribution and δ weighting parameter are added to the normalized exponential function; G is a random variable sampled from the Gumbel distribution G ~(0, 1), δ is the weighting parameter, Represents a latent feature matrix Mapped to a function of size M×2, where M represents the number of locations of interest, is the normalized exponential function, Represented as time series data measured at all locations of interest.

2. A sensor location selection method based on bridge monitoring as claimed in claim 1, characterized in that: Based on the multivariate time series data measured at the identified candidate key sensor locations, the neural network backbone is used to reconstruct the bridge response at the unmeasured locations of the bridge, specifically: Convolutional neural networks are used to extract features from measured multivariate time series data to obtain local spatiotemporal features; Based on the obtained local spatiotemporal features, a multi-head attention mechanism is used to focus on the potential features of information at different locations; The gated recurrent unit is used to capture the temporal dependencies in the time series data on the output features of the multi-head attention mechanism, and the bridge response at the unmeasured position of the bridge is obtained.

3. A sensor location selection method based on bridge monitoring as claimed in claim 1, characterized in that: During the training process, the sensor location selection adaptively selects the input sensing channels based on the switch outputs of each training cycle until a predetermined number of sensors is reached.

4. The sensor location selection method based on bridge monitoring according to claim 1, characterized in that: During the training process of the position selection and response reconstruction network model, the selection output of each position is represented as a binary classification distribution, and a switch matrix is ​​constructed for the selected or ignored sensing channels corresponding to each input position. The output of the switch matrix indicates the on and off states of each channel to determine the key sensor positions for reconstruction and the selected key sensor positions.

5. A sensor location selection system based on bridge monitoring, characterized in that: The system comprises: A sensor location determination module is configured to: obtain the key sensor locations of the bridge to be deployed based on the acquired locations of interest of the bridge to be deployed and the multivariate time series data measured at the locations of interest, and use the trained location selection and response reconstruction network model to obtain the key sensor locations of the bridge to be deployed; The model training module is configured as follows: the training process of the position selection and response reconstruction network model is specifically as follows: The adaptive location selection mechanism is used to generate a selection matrix of binary classification distribution for the locations of interest of the bridge, specifically: ， ， in, The normalized exponential function transforms the output into a distribution of [0, 1]; Represents a latent feature matrix Mapping to a function of size M×2; , each element Corresponding to m The selected output of each position; Represents input Implicit correlations in and Represents the corresponding weights in the convolutional neural network process, represents time series data measured at all locations of interest; Represents matrix dot product; Determine candidate key sensor positions for reconstruction by turning on or off sensors corresponding to each position of interest of the bridge in the selection matrix; Based on the multivariate time series data measured at the determined candidate key sensor positions, the neural network backbone is used to reconstruct the bridge response at the unmeasured position of the bridge, and the difference between the obtained bridge response and the actual displacement response of the bridge is minimized, and the position selection and response reconstruction network model is trained. In the training of the position selection and response reconstruction network model, in the back propagation process, the calculation of the adaptive position selection mechanism is specifically as follows: ， in, is a switch matrix that determines the selected or ignored sensing channels corresponding to each input position; ALSM stands for Adaptive Location Selection Mechanism; It means that the Gumbel distribution and δ weighting parameter are added to the normalized exponential function; G is a random variable sampled from the Gumbel distribution G ~(0, 1), δ is the weighting parameter, Represents a latent feature matrix Mapped to a function of size M×2, where M represents the number of locations of interest, is the normalized exponential function, Represented as time series data measured at all locations of interest.

6. A sensor location selection system based on bridge monitoring as claimed in claim 5, characterized in that: In the model training module, it specifically includes: representing the selected output of each position as a binary classification distribution, constructing a switch matrix of the selected or ignored sensing channels corresponding to each input position, indicating the on and off states of each channel through the output of the switch matrix, so as to determine the key positions of sensors used for reconstruction and determine the selected key sensor positions.

7. A computer device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, a sensor location selection method based on bridge monitoring as described in any one of claims 1 to 4 is performed.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the sensor location selection method based on bridge monitoring according to any one of claims 1 to 4 is executed.

Images