Bridge long-term deflection signal prediction method and system

Abstract

The invention provides a bridge long-term deflection signal prediction method and system. The method comprises the following steps: acquiring historical bridge deflection signal data; historical bridge deflection signal data is decomposed to obtain a periodic component, a random component and a structure dead load and damage component, and the components are respectively input into the improved Informer model to obtain corresponding prediction sequences; and superposing the prediction sequence corresponding to the periodic component, the prediction sequence corresponding to the random component and the prediction sequence corresponding to the structure dead load and damage component to obtain a bridge deflection prediction result. Historical bridge deflection signal data are accurately expressed through a periodic component, a random component, a structural dead load and a damage component, the problems of prediction precision reduction and long time consumption caused by excessive components are avoided, and the long-term deflection of a bridge is more accurately predicted by using the strong learning ability and sequence modeling ability of the improved Informer model.

Description

Technical Field

[0001] The present invention relates to the cross - field of bridge structure safety monitoring and computer science, and particularly to a method and system for predicting long - term deflection signals of bridges. Background Art

[0002] Bridges are one of the important infrastructure for modern transportation. However, due to the long - term influence of natural and human factors, such as structural aging caused by natural weathering and corrosion, overloading, natural disasters, poor maintenance, etc., bridges may suffer catastrophic consequences such as structural damage and collapse.

[0003] As a key indicator to measure the state of bridge structures, deflection can directly reflect the deformation of bridges under stress. When affected by various factors such as external loads, temperature difference changes, and material degradation, abnormal fluctuations in deflection often indicate potential problems in the structure or loads exceeding the design capacity. Therefore, real - time monitoring and prediction of deflection are particularly important, which can help in early diagnosis and taking corresponding maintenance measures to ensure the long - term stability and reliability of bridges.

[0004] In recent years, with the increasing maturity of sensing technology and the Internet of Things, bridge deflection data can be collected at high frequency and accurately. This provides a broad research platform for data - driven bridge deflection prediction methods, such as ARIMA, LSTM, Informer and other models. However, each of the above - mentioned models has its limitations when applied to bridge deflection prediction. The ARIMA model is somewhat insufficient in dealing with non - linear trends and complex seasonal data; although the LSTM model can capture long - term data dependencies, it has a large amount of calculation and may not be accurate enough in long - term time - series prediction; although the Informer model has improved the prediction accuracy to a certain extent, it has not fully explored the core characteristics of bridge deflection data, and the prediction accuracy still needs to be improved.

[0005] Based on this, seeking a high - accuracy and efficient prediction method for bridge deflection characteristics not only has theoretical value but also great practical application significance. Summary of the Invention

[0006] The present invention aims to solve the technical problems existing in the prior art and provides a method and system for predicting long - term deflection signals of bridges.

[0007] To achieve the above object of the present invention, according to the first aspect of the present invention, there is provided a method for predicting long-term deflection signals of a bridge, comprising: obtaining historical bridge deflection signal data, where the historical bridge deflection signal data includes historical deflection value sequences of more than one bridge cross-section; decomposing the historical bridge deflection signal data to obtain a periodic component, a random component, and a structural dead load and damage component; respectively inputting the periodic component, the random component, and the structural dead load and damage component into a pre-trained improved Informer model to obtain a prediction sequence corresponding to the periodic component, a prediction sequence corresponding to the random component, and a prediction sequence corresponding to the structural dead load and damage component; and superimposing the prediction sequence corresponding to the periodic component, the prediction sequence corresponding to the random component, and the prediction sequence corresponding to the structural dead load and damage component to obtain a bridge deflection prediction result.

[0008] To achieve the above object of the present invention, according to the second aspect of the present invention, there is provided a device for predicting long-term deflection signals of a bridge, which is used to implement the method for predicting long-term deflection signals of a bridge described in the first aspect of the present invention, comprising: a data acquisition module for obtaining historical bridge deflection signal data, where the historical bridge deflection signal data includes historical deflection value sequences of more than one bridge cross-section; a decomposition module for decomposing the historical bridge deflection signal data to obtain a periodic component, a random component, and a structural dead load and damage component; a prediction sequence acquisition module for respectively inputting the periodic component, the random component, and the structural dead load and damage component into a pre-trained improved Informer model to obtain a prediction sequence corresponding to the periodic component, a prediction sequence corresponding to the random component, and a prediction sequence corresponding to the structural dead load and damage component; and a superimposing and fusing module for superimposing the prediction sequence corresponding to the periodic component, the prediction sequence corresponding to the random component, and the prediction sequence corresponding to the structural dead load and damage component to obtain a bridge deflection prediction result.

[0009] To achieve the above object of the present invention, according to the third aspect of the present invention, there is provided a system for predicting long-term deflection signals of a bridge, comprising a data acquisition device and a processor; the data acquisition device collects bridge deflection data from a deflection sensor installed on the bridge and converts the collected bridge deflection data into historical bridge deflection signal data, and the processor obtains the historical bridge deflection signal data and obtains a bridge deflection prediction result according to the method for predicting long-term deflection signals of a bridge described in the first aspect of the present invention.

[0010] The present invention innovatively decomposes complex bridge deflection signal data into periodic components, random components, and structural dead load and damage components. The periodic components represent periodic signal data such as daily temperature difference effects and annual temperature difference effects, the random components represent random noises such as vehicle loads and wind loads, and the structural dead load and damage components are used to represent the information of the structural dead load action and damage of the bridge. The three components obtained by decomposition are respectively input into a pre-trained improved Informer model to obtain their respective corresponding prediction sequences, and then the three obtained prediction sequences are superimposed and fused to obtain the final bridge deflection prediction result. By means of the three components, not only the historical bridge deflection signal data is accurately expressed, but also the problems of cumulative prediction error and long time consumption caused by too many components reducing the prediction accuracy are avoided. Moreover, by utilizing the powerful learning ability and sequence modeling ability of the improved Informer model, the long-term deflection of the bridge can be predicted more accurately, enabling technicians to detect abnormal deflection situations in a timely manner and make corresponding treatments and warnings, which helps to take measures in advance to avoid potential structural problems and improve the operation safety and reliability of the bridge. Description of the Drawings

[0011] Figure 1 It is a schematic flow chart of the method for predicting the long-term deflection signal of a bridge in a preferred embodiment of the present invention;

[0012] Figure 2 It is an execution flow chart in a specific application scenario of the method for predicting the long-term deflection signal of the bridge of the present invention;

[0013] Figure 3 It is a schematic diagram of the network structure of the improved Informer model in a preferred embodiment of the present invention;

[0014] Figure 4 It is a comparison chart of experimental effects of the method for predicting the long-term deflection signal of the bridge of the present invention on different data sets;

[0015] Figure 5 It is a comparison chart of experiments when various prediction methods are used to predict the bridge deflection. Detailed Embodiments

[0016] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation of the present invention.

[0017] In the description of the present invention, it should be understood that the orientation or positional relationships indicated by the terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. are based on the orientation or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention.

[0018] In the description of the present invention, unless otherwise specified and defined, it should be noted that the terms "installed", "connected", "coupled" should be understood in a broad sense. For example, it may be a mechanical connection or an electrical connection, or it may be the communication inside two elements. It may be directly connected or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0019] The present invention discloses a method for predicting long-term deflection signals of bridges. In a preferred embodiment, the schematic flow diagram of the prediction method is as Figure 1 shown Figure 2 The implementation process of a specific application scenario is disclosed. The prediction method includes:

[0020] Step S1, obtaining historical bridge deflection signal data, where the historical bridge deflection signal data includes historical deflection value sequences of more than one bridge cross-section. The historical bridge deflection signal data is the historical data of the deflection values of a certain cross-section or multiple cross-sections of the bridge collected by deflection sensors installed on the bridge. The inventors of the present application have conducted a large number of experiments and found that the bridge deflection data can be regarded as the superposition of periodic information, random information, structural dead load effect information and damage information, where the periodic information mainly includes daily temperature difference effect and annual temperature difference effect, and the random information mainly includes vehicle load, wind load random noise, etc.

[0021] Step S2, decomposing the historical bridge deflection signal data to obtain a periodic component, a random component, and a structural dead load and damage component.

[0022] For the convenience of accurately decomposing the periodic component, random component, and structural dead load and damage component from the historical bridge deflection signal data, preferably, step S2 includes:

[0023] Step S20, using the Local Mean Decomposition method (LMD for short) to decompose the historical bridge deflection signal data to obtain a plurality of local mean function components and a residual function component; taking the residual function component as the structural dead load and damage component.

[0024] Step S21, determining the period bias score of each local mean function component.

[0025] Step S22: Arrange the local average function components in descending order of the periodic deviation score to obtain a sequence of local average function components.

[0026] Step S23: In the sequence of local average function components, determine the partitioning index, and superimpose and fuse the local average function components whose superimposing and fusing index is less than or equal to the partitioning index to obtain the periodic component, and superimpose and fuse the local average function components whose superimposing and fusing index is greater than the partitioning index to obtain the random component.

[0027] Step S3: Input the periodic component, the random component, and the structural dead load and damage component into the pre-trained improved Informer model respectively to obtain the prediction sequence corresponding to the periodic component, the prediction sequence corresponding to the random component, and the prediction sequence corresponding to the structural dead load and damage component.

[0028] Step S4: Superimpose the prediction sequence corresponding to the periodic component, the prediction sequence corresponding to the random component, and the prediction sequence corresponding to the structural dead load and damage component to obtain the bridge deflection prediction result.

[0029] In this embodiment, preferably, step S20 includes:

[0030] S201: Denote all local extreme points in the historical bridge deflection signal data D as n q (q = 1, 2,..., M), where M represents the number of extreme points. Let any two extreme points be n q and n q+1 , then there is: In the formula, m q is the average value of two adjacent extreme points, and a q is the envelope estimation value.

[0031] S202: Connect the discrete points of m q and a q into broken lines respectively, and use the moving average method to process m q and a q to obtain the local mean function m 11 (t) and the local envelope function a 11 (t), where t represents time.

[0032] S203: Separate the local mean function m 11 (t) from the historical bridge deflection signal data D to obtain h 11 (t) = D - m 11 (t), and demodulate h 11 (t) to obtain the modal function h 11(t) is the envelope function, representing the difference between the original signal and its local mean, and the mode function s 11 (t) represents a signal component with certain frequency characteristics;

[0033] S204: If s 11 (t) is a pure frequency modulation signal, then s 11 (t)'s local envelope function a 12 (t) = 1. If a 12 (t) ≠ 1, then iterate according to the following formula until s 1r (t) becomes a pure frequency modulation function, that is, s 1r (t)'s local envelope function a 1(r+1) (t) = 1, and the following formula is obtained:

[0034] The iteration termination condition is: a 1r (t) = 1 + ε, where the first parameter ε is a custom small value to accelerate convergence without affecting the accuracy.

[0035] S205: After the iteration is completed, all the envelope estimation functions generated during the iteration are multiplied to obtain the envelope signal a1(t), that is

[0036] S206: Multiply the envelope signal a1(t) and the pure frequency modulation signal s 1r (t) to obtain the first PF component PF1(t) in the original signal = a1(t)s 1r (t). The PF component is the local average function component. Among them, the amplitude of PF1(t) is a1(t), and the instantaneous frequency is

[0037] S207: Separate PF1(t) from the historical bridge deflection signal data D to obtain the residual signal RES = D - PF1.

[0038] S208: Repeat steps S201 to S207 a total of n times until the final residual signal RES is a monotonic function. The final residual signal RES is the residual function component. At this time, D is divided into n PF components (local average function components) and 1 residual component (residual function component), expressed as

[0039] In a preferred embodiment, in step S21, in order to accurately obtain the period deviation score, the process of determining the period deviation score of the i-th local average function component includes:

[0040] Step S211, determine the periodicity score P i , P i= max(|PF i (f)|), where PF i (f) is the fast Fourier transform result of the i-th local average function component PF i (t), and |PF i (f)| represents the amplitude of PF i (f). First, perform a fast Fourier transform on the i-th local average function component PF i (t), PF i (f) = FFT(PF i (t)), where PF i (f) is the representation of PF i (t) in the frequency domain. Then, calculate the amplitude spectrum of PF i (f) and find the amplitude peak, and take the amplitude peak as the periodic score, P i = max(|PF i (f)|).

[0041] Step S212, determine the live load score L i of the i-th local average function component, L i = ∫(PF i (t)) 2 dt. Use the instantaneous energy of PF i (t) to determine its correlation with the live load, that is, the live load score L i .

[0042] Step S213, take the entropy of the i-th local average function component as the random noise score N i of the i-th local average function component. The specific process is as follows:

[0043] First, divide it into Ω intervals according to the maximum and minimum values of the i-th local average function component PF i (t), where Ω is a positive integer, and the width of each interval is The interval index q * ∈[1, Ω], and the q * -th interval can be expressed as [min(PF i (t))+(q * -1)W, min(PF i (t))+q * W]. Then, calculate the probability that the signal value of the local average function component PF i (t) falls into the q * -th interval Then, use the entropy N i of the i-th local average function component PF i to estimate its randomness, and the calculation formula is Among them, is the number of signal values in each interval, M * is the total number of data points of the i-th local average function component PF i (i).

[0044] Step S214, the period bias score S of the i-th local average function component i is:

[0045] S i = ω1×P i + ω2×(1 - L i ) + ω3×(1 - N i ), where ω1 + ω2 + ω3 = 1, ω1 represents the first weight, ω2 represents the second weight, ω3 represents the third weight, and ω1, ω2, ω3 ∈ [0, 1].

[0046] In this embodiment, by (1 - L i ) and (1 - N i ), the live load score and the random noise score are inverted to ensure that the higher the period bias score, the higher the period bias score, the greater the possibility of the periodic component of the signal, and the smaller the randomness, improving the accuracy of the three-component decomposition.

[0047] In a preferred embodiment, step S23 specifically includes:

[0048] Step S231, set a score threshold; the score threshold is the threshold of the period bias score. Set the score threshold S according to the following formula τ : S τ = μ S + λ×σ S , where μ S represents the average value of the period bias scores of the local average function components, σ S is the standard deviation of the period bias scores of the local average function components, λ is the first coefficient. Due to the differences in various bridges and environmental conditions, the score threshold needs to be dynamically adjusted for different bridges or conditions. Therefore, λ can be adjusted according to the actual application. The value range of λ is preferably but not limited to 0.5 to 3. For example, λ = 1 means selecting the S S value that is higher than the average value μ S plus one standard deviation σ i as the periodic component.

[0049] Step S232, use the index of the local average function component with the period bias score closest to the score threshold in the local average function component sequence as the division index. Specifically, let the division index be j * , j * = argmin i |Si -S τ |, argmin i means to assign the index i with the minimum |S i -S τ | value to j * . By knowing the partition index, in the local average function component sequence, is periodic information, is randomness information.

[0050] Step S233, superimpose and fuse the local average function components with indexes less than or equal to the partition index to obtain the periodic component, and superimpose and fuse the local average function components with indexes greater than the partition index to obtain the randomness component. Superimpose and fuse the periodic information to obtain the periodic component through superimposed fusion randomness information to obtain the randomness component through superimposed fusion t represents time.

[0051] In a preferred embodiment, to improve the processing speed and prediction accuracy, the periodic component, the randomness component, and the structural dead load and damage component are respectively normalized before being input into the improved Informer model. For the periodic component randomness component and the structural dead load and damage component RES are respectively normalized by Z-score. The Z-score normalization calculation formula is: x represents the numerical value of the data point before normalization in the component, represents the mean value of the data point numerical values in the component, and σ represents the variance of the data point numerical values in the component.

[0052] In a preferred embodiment, such as Figure 3As shown in the figure, the improved Informer model includes: an input embedding layer that embeds local timestamps and global timestamps in the encoded input vector to obtain encoded input data; wherein, the encoded input vector is a normalized periodic component, random component, or structural dead load and damage component; an encoder that encodes the encoded input data to obtain a fused feature map, and the encoder includes two or more cascaded encoding modules, and the encoding module includes a cascaded multi-head probabilistic sparse self-attention layer and a distillation operation layer; an output embedding layer that is used to connect the start position marker sequence and the predicted padding sequence to obtain a decoded input vector, and embeds local timestamps and global timestamps in the decoded input vector to obtain decoded input data; a decoder that includes a masked multi-head probabilistic sparse self-attention layer and a cross-attention layer, the masked multi-head probabilistic sparse self-attention layer processes the decoded input data, and the cross-attention layer cross-encodes the fused feature map and the output data of the masked multi-head probabilistic sparse self-attention layer to obtain a cross-encoded vector; a one-dimensional convolutional layer that processes the cross-encoded vector to obtain a predicted sequence.

[0053] In this embodiment, the input embedding layer performs input unified transformation, and the specific method is: the encoded input data consists of a feature scalar, a local timestamp (PE), and a global timestamp (SE); the transformation formula is: where i′ ∈ {1,..., L x}, i′ represents the element index in the encoded input vector, and L x represents the number of elements in the encoded input vector, represents the i′-th element of the encoded input vector, and α is a factor that balances the size between the scalar mapping and the local / global embedding. The in the formula corresponding to the feature scalar is specifically operated to convert it into a 512-dimensional vector through one-dimensional convolution. The local timestamp (PE) adopts the PositionEmbedding in Transformer, and the calculation formula is:

[0054] where d model is the input feature dimension,

[0055] After the unified transformation, the encoded input data is input into the Encoder of the model, and probabilistic sparse self-attention calculation is performed in the attention module. Each key only focuses on u main queries, and the calculation formula is: where, is a sparse matrix with the same size as Q and only contains the Top-u queries under the coefficient metric M(q i′ , K); a sampling factor c is added, and the set number u = lnL Q; Randomly sample c * 1nL keys for each query, and calculate the sparsity score M(q i′ , K); The coefficient metric M(q i′ , K) is calculated as follows Select the N queries with the highest sparsity scores. N is defaulted to c * 1nL. Only calculate the dot product results of N queries and keys, and do not calculate the remaining L - N queries; After the probability sparsity self-attention calculation, there are redundant combinations of V values in the output. Therefore, a distillation operation layer is used to assign higher weights to the dominant features with main features and generate a focused self-attention feature map in the next layer; Specifically, it is implemented through two one-dimensional convolutional layers and a max pooling layer. The formula for the "distillation" process from the j-th layer to the (j + 1)-th layer is as follows After repeating the combination of "multi-head probability sparse self-attention mechanism + distillation" iteratively, a fused feature map is obtained as one of the inputs to the decoder Decoder.

[0056] The output embedding layer is used to connect the start position token sequence and the predicted padding sequence to obtain the decoded input vector, and embed the local timestamp and global timestamp in the decoded input vector to obtain the decoded input data

[0057] Set the output embedding layer to obtain the decoded input data of the decoder. The output embedding layer connects the start position token sequence and the predicted padding sequence to obtain the decoded input vector Among them, the start position token sequence is the start token sequence, and the predicted padding sequence is the sequence to be predicted, that is, the padding sequence of the one-dimensional convolution, Figure 3 a sequence composed of 5 zeros filled with 0 in. Embed the local timestamp and global timestamp in the decoded input vector to obtain the decoded input data and input it into the masked multi-head probability sparse self-attention layer of the decoder. Through the cross-attention layer, cross-encode the fused feature map and the input data to obtain the cross-encoded vector, so as to better obtain the mapping relationship between the input and output, and improve the prediction accuracy; Finally, through a one-dimensional convolutional layer, capture the local time-related patterns to obtain the accurate prediction sequence.

[0058] The training process of improving the Informer model includes:[[]]

[0059] Decompose the historical bridge deflection signal data into periodic components, random components, and structural dead load and damage components, divide the dataset for each component, and the first 70% is the training set PF n_train , 10% is the validation set PF n_val , and the last 20% is the test set PF n_test, retain sufficient deflection component data for verification and testing to ensure the reliability and generalization ability of the model. Take samples of batch-size from the training set PF n_train and train the network structure of the improved Informer model. During the training process, calculate the loss function. The formula for the loss function is:

[0060]

[0061] where m represents the number of training samples, represents the predicted sequence data, represents the true sequence data corresponding to the predicted sequence data, and i * represents the training sample index.

[0062] After training, use the validation set to verify the trained improved Informer model and use the test set to test it. Evaluate the performance of the trained improved Informer model through model evaluation metrics. Preferably, the model evaluation metrics are selected as MAE and MASE. The Mean Absolute Scaled Error (MASE) is an index for evaluating the prediction accuracy of time series, which can compare the prediction accuracies between datasets of different scales and different models, is not affected by units, and has stronger interpretability. Its calculation formula is as follows,

[0063]

[0064] Conduct experimental verification on the prediction effect of the long-term bridge deflection signal prediction method (hereinafter referred to as: LMD-Informer) disclosed in the present invention.

[0065] Experiment 1: The experimental data comes from the deflection data collected by the deflection sensor D1-1 of the actual bridge A in China, with a total of 90087 pieces of data. In this experiment, compare the results of the original dataset D (i.e., historical bridge deflection signal data), the dataset D H with a resampling time interval of 1 hour and the dataset D D with a resampling time interval of 1 day under the prediction lengths of 24, 48, and 96. Among them, the dataset D D has the lowest sampling frequency, and the dataset D H has the highest sampling frequency. The experimental model is the long-term bridge deflection signal prediction method LMD-Informer of the present invention, and MAE and MASE are used as evaluation metrics. The experimental results are shown in Table 1, Figure 4 are the predicted slice values of LMD-Informer on the deflection datasets D, D H , D D , Figure 4Among them, from top to bottom, the first curve is the real sequence curve, the second is the predicted sequence curve, and the third curve is the deviation curve between the real sequence and the predicted sequence.

[0066] Table 1 Original dataset D, resampled dataset D H , D D Comparison chart of prediction results using LMD-Informer

[0067]

[0068] It can be seen from the various evaluation indicators in Table 1 that the prediction accuracy decreases as the prediction length (predicted sequence length) increases, and the prediction accuracy of data with a higher sampling frequency is higher. In addition, although there are differences in the data length after resampling, the change in MASE is not significant, indicating that the LMD-Informer model is less affected by the size of the dataset.

[0069] Experiment 2: The data for this experiment comes from the deflection data collected by sensor D1-1 of a domestic real bridge A, with a total of 90,087 pieces of data. In this experiment, LMD-Informer (this application), Informer, EMD-Informer, and LMD-LSTM will be compared, and the deflection dataset is D H , the prediction lengths are 24, 48, and 96 respectively, and MAE and MASE are used as evaluation indicators. The experimental results are shown in Table 2, and the comparison results of the MASE error curves of each model are as Figure 5 shown.

[0070] Table 2 Comparison chart of results with prediction lengths of 24, 48, and 96

[0071]

[0072] It can be seen from the various evaluation indicators in Table 2 that the rankings of the MAE and MASE indicators are both LMD-Informer < EMD-Informer < Informer < LMD-LSTM, indicating that compared with the other three models, there are no large errors in the predicted values of the LMD-Informer model of this application, and it also shows that the model proposed in this application has certain advantages in terms of prediction stability.

[0073] The present invention also discloses a device for predicting long-term deflection signals of a bridge, which is used to implement the method for predicting long-term deflection signals of a bridge provided by the present invention, and includes: a data acquisition module that acquires historical bridge deflection signal data, where the historical bridge deflection signal data includes historical deflection value sequences of more than one bridge cross-section; a decomposition module that decomposes the historical bridge deflection signal data to obtain a periodic component, a random component, and a structural dead load and damage component; a prediction sequence acquisition module that respectively inputs the periodic component, the random component, and the structural dead load and damage component into a pre-trained improved Informer model to obtain a prediction sequence corresponding to the periodic component, a prediction sequence corresponding to the random component, and a prediction sequence corresponding to the structural dead load and damage component; and a superposition and fusion module that superposes the prediction sequence corresponding to the periodic component, the prediction sequence corresponding to the random component, and the prediction sequence corresponding to the structural dead load and damage component to obtain a bridge deflection prediction result.

[0074] The above device corresponds to each step of the method for predicting long-term deflection signals of a bridge, and will not be elaborated here.

[0075] The present invention also discloses a system for predicting long-term deflection signals of a bridge, which includes a data acquisition device and a processor; the data acquisition device acquires bridge deflection data from a deflection sensor installed on the bridge and converts the acquired bridge deflection data into historical bridge deflection signal data, and the processor acquires the historical bridge deflection signal data and obtains a bridge deflection prediction result according to the method for predicting long-term deflection signals of a bridge provided by the present invention. The data acquisition device is preferably but not limited to a data acquisition card or an Internet of Things processing device. The processor is preferably but not limited to an embedded system, a server, or a PC personal computer.

[0076] In the description of this specification, the descriptions with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0077] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and purposes of the present invention, and the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A long-term deflection signal prediction method for bridges, characterized in that, Including: Obtain historical bridge deflection signal data, where the historical bridge deflection signal data includes historical deflection value sequences of more than one bridge section; Decompose the historical bridge deflection signal data to obtain a periodic component, a random component, and a structural dead load and damage component; Input the periodic component, the random component, and the structural dead load and damage component into a pre-trained improved Informer model respectively to obtain a prediction sequence corresponding to the periodic component, a prediction sequence corresponding to the random component, and a prediction sequence corresponding to the structural dead load and damage component; Superimpose the prediction sequence corresponding to the periodic component, the prediction sequence corresponding to the random component, and the prediction sequence corresponding to the structural dead load and damage component to obtain the bridge deflection prediction result.

2. The long-term deflection signal prediction method for bridges according to claim 1, wherein The decomposing the historical bridge deflection signal data to obtain a periodic component, a random component, and a structural dead load and damage component includes: Use the local mean decomposition method to decompose the historical bridge deflection signal data to obtain multiple local mean function components and a residual function component; use the residual function component as the structural dead load and damage component; Determine the period bias score of each local mean function component; Arrange the local mean function components in descending order of the period bias score to obtain a local mean function component sequence; In the local mean function component sequence, determine the division index, and superimpose and fuse the local mean function components whose fusion index is less than or equal to the division index to obtain the periodic component, and superimpose and fuse the local mean function components whose fusion index is greater than the division index to obtain the random component.

3. The long-term deflection signal prediction method for bridges according to claim 2, characterized in that Determining the period bias score of the i-th local mean function component includes: Determine the periodicity score P of the i-th local average function component i , P i = max(|PF i (f)|), where PF i (f) is the fast Fourier transform result of the i-th local average function component PF i (t), and |PF i (f)| represents the amplitude of PF i (f); Determine the live load score L of the i-th local average function component i , L i = ∫(PF i (t)) 2 dt; Take the entropy of the i-th local average function component as the random noise score N of the i-th local average function component i ; The period bias score S of the i-th local average function component i is as follows: S i = ω1 × P i + ω2 × (1 - L i ) + ω3 × (1 - N i ), where ω1 + ω2 + ω3 = 1, ω1 represents the first weight, ω2 represents the second weight, ω3 represents the third weight, and ω1, ω2, ω3 ∈ [0, 1].

4. The long-term deflection signal prediction method for a bridge according to claim 2 or 3, characterized in that In the local mean function component sequence, determining the division index includes: Set a score threshold; Use the index of the local mean function component whose period bias score is closest to the score threshold in the local mean function component sequence as the division index.

5. The long-term deflection signal prediction method for a bridge according to claim 4, characterized in that, Set the score threshold S according to the following formula τ :[[]]END]] S τ = μ S + λ × σ S , where μ S represents the average of the periodic bias scores of the local average function components, σ S is the standard deviation of the periodic bias scores of the local average function components, and λ is the first coefficient.

6. The long-term deflection signal prediction method for a bridge according to claim 1 or 2 or 3 or 5, characterized in that The periodic component, the random component, and the structural dead load and damage component are respectively normalized before being input into the improved Informer model.

7. The long-term deflection signal prediction method for a bridge according to claim 6, wherein The improved Informer model includes: An input embedding layer that embeds local timestamps and global timestamps in the encoded input vector to obtain encoded input data; where the encoded input vector is the normalized periodic component or random component or structural dead load and damage component; An encoder that encodes the encoded input data to obtain a fused feature map. The encoder includes more than two cascaded encoding modules, and each encoding module includes a cascaded multi-head probabilistic sparse self-attention layer and a distillation operation layer; An output embedding layer that is used to connect the start position marker sequence and the prediction padding sequence to obtain a decoded input vector, and embeds local timestamps and global timestamps in the decoded input vector to obtain decoded input data; A decoder that includes a masked multi-head probabilistic sparse self-attention layer and a cross-attention layer. The masked multi-head probabilistic sparse self-attention layer processes the decoded input data, and the cross-attention layer cross-encodes the fused feature map and the output data of the masked multi-head probabilistic sparse self-attention layer to obtain a cross-encoded vector; A one-dimensional convolutional layer that processes the cross-encoded vector to obtain a prediction sequence.

8. The long-term deflection signal prediction method for a bridge according to claim 1 or 2 or 3 or 5 or 7, characterized in that, The calculation formula of the loss function adopted when training the improved Informer model is as follows: where m represents the number of training samples, represents the predicted sequence data, represents the true sequence data corresponding to the predicted sequence data, and i * represents the training sample index.

9. A long-term bridge deflection signal prediction device for implementing the long-term bridge deflection signal prediction method according to any one of claims 1-7, comprising: A data acquisition module that acquires historical bridge deflection signal data, where the historical bridge deflection signal data includes historical deflection value sequences of more than one bridge cross-section; A decomposition module that decomposes the historical bridge deflection signal data to obtain a periodic component, a random component, and a structural dead load and damage component; A prediction sequence acquisition module that inputs the periodic component, the random component, and the structural dead load and damage component into a pre-trained improved Informer model respectively to obtain a prediction sequence corresponding to the periodic component, a prediction sequence corresponding to the random component, and a prediction sequence corresponding to the structural dead load and damage component; A superposition and fusion module that superposes the prediction sequence corresponding to the periodic component, the prediction sequence corresponding to the random component, and the prediction sequence corresponding to the structural dead load and damage component to obtain a bridge deflection prediction result.

10. A long-term bridge deflection signal prediction system, characterized in that, It includes a data acquisition device and a processor; the data acquisition device acquires bridge deflection data from a deflection sensor installed on the bridge and converts the acquired bridge deflection data into historical bridge deflection signal data, and the processor acquires the historical bridge deflection signal data and obtains a bridge deflection prediction result according to the long-term bridge deflection signal prediction method according to any one of claims 1-7.

Images