A vehicle-bridge parameter synchronous inversion identification method based on a deep time sequence model

Abstract

The application discloses a vehicle-bridge parameter synchronous inversion identification method based on a deep time sequence model, which comprises the following steps: collecting bridge response data; constructing an MFNI model; training the MFNI model; extracting vehicle moving force characteristics by using the trained MFNI model, and obtaining dynamic force time courses of each axle of the vehicle on the bridge; correcting the output of the decoder by using a DMF-corr block; extracting wheelbase and vehicle speed information from the vehicle moving force characteristics output by the MFNI model through a time sequence feature derivation method; obtaining vehicle-induced response samples; constructing a PTAD model, capturing damage-sensitive features in the vehicle-induced response samples through the PTAD model; obtaining damage indicators through the damage-sensitive features, evaluating the bridge state according to the damage indicators, and quantifying the damage degree of the bridge. The application realizes accurate and efficient detection of bridge structure damage, provides a new technical means for structure health monitoring in the field of bridge engineering, and has important significance for promoting the intelligentization and refinement development of the field of bridge engineering.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of bridge structure health monitoring, and particularly relates to a vehicle-bridge parameter synchronous inversion identification method based on a deep time series model. BACKGROUND

[0002] In the field of bridge engineering, with the continuous increase of traffic flow and the extension of bridge service life, it is particularly important to conduct real-time and accurate health monitoring of bridge structures. Traditional monitoring methods often focus on static analysis of a single parameter, and it is difficult to comprehensively capture the dynamic response of bridge structures under vehicle loads and the complex interaction with vehicle parameters.

[0003] Traditional vehicle-bridge interaction analysis methods usually rely on prior knowledge of vehicle speed, axle number, axle distance and other parameters. The acquisition of these parameters often depends on complex vehicle detection systems or manual measurement, which is not only costly but also inefficient. In addition, existing moving load identification (MFI) techniques are mostly based on time domain, frequency domain or interpretative methods. These methods are affected to varying degrees by the dynamic characteristics of bridge structures, the diversity of vehicle types and environmental factors (such as road roughness), limiting their accuracy and robustness in practical applications. In recent years, with the rapid development of deep learning technology, especially its outstanding performance in time series analysis and pattern recognition, it has provided a new approach for vehicle-bridge parameter synchronous inversion identification. Deep neural networks have shown great potential in handling complex system dynamic response data due to their powerful non-linear mapping ability and automatic feature extraction ability. However, existing deep learning-based MFI methods have largely failed to effectively address the problem of high-precision identification in the absence of detailed vehicle parameters (such as vehicle speed and axle configuration).

[0004] In the field of bridge engineering, structural damage state detection is also one of the key technologies to ensure the safe operation of bridges. Traditional bridge damage detection methods rely on periodic manual detection or model-based damage identification techniques, which have limitations such as long detection period, high cost, susceptibility to human factors, and large model errors. Especially when the bridge structure is complex and the sensor deployment is sparse, the accuracy and efficiency of traditional methods often cannot meet the actual demand. In recent years, with the rapid development of deep learning technology, especially in the field of time series analysis, deep neural network-based models have shown great potential in structural health monitoring. In particular, the Transformer model has achieved great success in natural language processing and computer vision, and its strong sequence modeling capability provides a new approach to processing time series data. However, direct application of the Transformer model in the field of structural health monitoring still faces many challenges, such as high model complexity, large computational load, and insufficient ability to capture local features. To overcome the above challenges, researchers have proposed an improved model based on PatchTST (Patch-based Time Series Transformer), which introduces a patch mechanism to reduce the length of the input sequence while emphasizing local semantic information, thereby reducing computational load while ensuring model performance. This innovation provides a new solution for processing time series data and a new approach for bridge structural health monitoring. However, existing deep learning-based bridge damage detection methods mostly rely on a large amount of labeled data, which is often difficult to achieve in practical applications. SUMMARY

[0005] The purpose of the present application is to overcome the deficiencies in the prior art and provide a vehicle-bridge parameter synchronous inversion identification method based on a deep time series model, which realizes accurate and efficient detection of bridge structural damage, provides a new technical means for structural health monitoring in the field of bridge engineering, and has important significance for promoting the intelligent and refined development of the field of bridge engineering.

[0006] Technical solution: To achieve the above purpose, the present application provides a vehicle-bridge parameter synchronous inversion identification method based on a deep time series model, comprising the following steps:

[0007] S1: Collect bridge response data and perform preprocessing; data includes displacement, acceleration, etc., preprocessing includes denoising, outlier removal, etc., to ensure the quality of input data;

[0008] S2: Construct an MFNI model, including an encoder, a decoder, and a dynamic moving load correction DMF-corr block;

[0009] S3: Based on the pretreated bridge response data, the finite element model of the target bridge is corrected to generate simulation response samples for MFNI model training; by optimizing the mean absolute error loss function, the model can accurately identify the vehicle moving force from the bridge response;

[0010] Parameter sensitivity analysis is performed to evaluate the influence of different parameters on the model performance, further optimize the model structure and parameter settings, and improve the robustness and identification accuracy of the model;

[0011] S4: Use the trained MFNI model to process new bridge response data, extract vehicle moving force features, and obtain dynamic force time history of each vehicle axle on the bridge;

[0012] S5: Use the DMF-corr block to correct the output of the decoder to ensure the accuracy of the identification results; through the correction function, the model can more accurately judge the number of vehicle axles and correct the load time history of non-bridge action time steps;

[0013] S6: Extract the wheelbase and vehicle speed information from the vehicle moving force features output by the MFNI model through the time series feature derivation method;

[0014] S7: Integrate the MFNI model into the bridge health monitoring system to realize real-time monitoring and identification of vehicle moving force; this method can be used in bridge structure health monitoring, traffic flow analysis, overload detection and other fields, and improve the intelligent level of bridge management and maintenance;

[0015] Through the MFNI model, the vehicle-induced response signal in the bridge monitoring signal is collected and separated in real time, the vehicle-induced response signal is denoised, and the sliding window technology is used to divide the vehicle-induced response sample;

[0016] S8: Construct the PTAD model based on PatchTST and grouped convolution, and train the PTAD model using the vehicle-induced response sample;

[0017] S9: Capture the damage-sensitive features in the vehicle-induced response sample through the trained PTAD model;

[0018] S10: Obtain the damage index through the damage-sensitive features, and evaluate the bridge state according to the damage index to quantify the degree of bridge damage.

[0019] Further, the encoder in step S2 adopts a dilated causal convolution structure to capture local feature information and interaction dependency patterns in the input data; the decoder uses a recurrent neural network model with a stacked dilated structure to capture complex temporal dependency patterns; the DMF-corr block is used to correct and modify the decoding results to improve the identification accuracy.

[0020] Further, the load correction function of the DMF-corr block in step S2 is realized by adding a ReLU activation function after the output full connection layer, and the specific form is wherein z is the output of the neural network before correction.

[0021] Further, the generation of the simulation response sample in step S3 includes:

[0022] A1: Correct the finite element model to randomly generate road roughness data;

[0023] A2: Generate a vehicle model with some key to-be-identified parameters randomly, and use it to generate a large number of simulation training samples.

[0024] Further, the time series feature derivation method in step S6 is realized by using the corrected load time history results to solve the wheelbase and vehicle speed information, and specifically includes:

[0025] B1: Calculate the time step difference τ i acted on the bridge by different axles based on the cross-correlation function, and the formula is as follows:

[0026]

[0027] B2: Obtain the average action time Δt of each axle load on the bridge by using the first-order difference to obtain the on-bridge and off-bridge time steps of the axles;

[0028] B3: Substitute and solve the following linear inhomogeneous equation group to obtain the wheelbase a i and the vehicle speed estimate v:

[0029]

[0030] Further, the damage index in step S10 includes a long-term damage index DI long-term and a short-term damage index DI short-term , by calculating the reconstruction error of the test data in the PTAD model, obtaining the error vector θ composed of the average REI value of each sensor in the sliding window, and comparing it with the reference error vector θ0 in the healthy state, and calculating DI long-term and DI short-term ;

[0031] The calculation formula is as follows:

[0032]

[0033] Further, the bridge state is evaluated according to the damage index in step S10, and the damage degree of the bridge is quantified, specifically including: the DIs values obtained in the healthy state are respectively used to establish the short-term and long-term damage thresholds with a confidence level of 95%; according to the long-term damage index DIlong-term Comparing with the preset damage threshold, it is judged whether the bridge has potential damage; at the same time, the short-term damage index DI short-term , so as to evaluate the damage degree in real time;

[0034] According to the results of DI long-term and DI short-berm , the damage degree of the bridge is quantified, which provides a basis for subsequent maintenance decision.

[0035] Advantages: compared with the prior art, the present application has the following advantages:

[0036] 1. The present application fuses deep learning and time series analysis technology to construct an advanced vehicle-bridge parameter synchronous inversion identification framework. The model can fully utilize the rich information contained in the bridge vibration signal, combined with the dynamic characteristics in the vehicle driving process, to realize the synchronous and high-precision inversion of the key parameters such as vehicle moving load and bridge structure state. This innovation not only overcomes the limitations of traditional methods in dealing with complex dynamic systems, but also significantly improves the utilization efficiency of monitoring data and the accuracy of parameter identification.

[0037] 2. Aiming at developing a new method that can accurately identify and synchronously invert the key parameters (such as vehicle axle load, axle distance, vehicle speed and bridge dynamic response) in the vehicle-bridge interaction process without relying on specific vehicle information, it is of great significance to improve the intelligent level and practical application effect of bridge health monitoring. The present application method is just for this technical problem, through constructing a specially designed deep neural network architecture, realizing efficient and accurate identification of vehicle-bridge interaction process, providing strong support for safety evaluation and maintenance management of bridge structure.

[0038] 3. The present application provides a method for bridge damage detection using unlabeled data (i.e. unsupervised learning), which takes into account the important role of vehicle-bridge interaction (VBI) in bridge dynamic response, integrates VBI data into the damage detection model, realizes the synchronous inversion identification of vehicle-bridge parameters, and further improves the accuracy and real-time performance of damage detection.

[0039] 4. The present application aims to realize accurate and efficient detection of bridge structure damage, and provides new technical means for structure health monitoring in the field of bridge engineering, which has important significance for promoting the intelligent and fine development of bridge engineering field. The method framework of the present application has an automatic processing flow from the original signal to the final identification result, and further provides the synchronous change of vehicle and bridge parameters in real time, which provides strong theoretical method support for safety evaluation, damage warning and maintenance decision of bridge. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The schematic diagram of the overall process of the method of the present application is shown in the figure;

[0041] Figure 2 The deep learning network structure involved in the bridge dynamic weighing system of the present application is shown in the figure;

[0042] Figure 3 The deep learning network structure involved in the bridge abnormal state diagnosis of the present application is shown in the figure;

[0043] Figure 4 The preliminary verification effect diagram of the bridge dynamic weighing system is shown in the figure;

[0044] Figure 5 The preliminary verification effect diagram of the long-term abnormal state diagnosis is shown in the figure;

[0045] Figure 6 The preliminary verification effect diagram of the short-term abnormal state diagnosis is shown in the figure. DETAILED DESCRIPTION

[0046] The present application will be further illustrated below in combination with the drawings and specific embodiments, and it should be understood that these embodiments are only used to illustrate the present application and not used to limit the scope of the present application, and after reading the present application, the modifications of various equivalent forms of the present application by those skilled in the art all fall within the scope defined by the appended claims of the present application.

[0047] The present application provides a vehicle-bridge parameter synchronous inversion identification method based on a deep time sequence model, as shown in the figure, which sequentially includes five stages of data acquisition, bridge dynamic weighing network training, load weighing network deployment, vehicle-induced response sample acquisition, and bridge abnormal state diagnosis, and the method of the present application will be described below in combination with Figure 1 Figures 1-3 , including the following steps:

[0048] 1) Data acquisition stage

[0049] Collect bridge response data and perform preprocessing; the data includes displacement, acceleration, etc., and the preprocessing includes, for example, denoising, outlier rejection, etc., to ensure the quality of the input data; in this embodiment, db8 wavelet is used for 4-layer decomposition to achieve the purpose of signal denoising.

[0050] 2) Bridge dynamic weighing network training stage

[0051] 2-1) Construct the MFNI model, including the encoder, the decoder, and the dynamic moving load correction DMF-corr block;

[0052] As shown in Figure 2 ​As shown, the encoder adopts an expanded causal convolution structure to capture local feature information and interaction dependency patterns in the input data; the decoder uses a recurrent neural network model with a stacked expansion structure to capture complex temporal dependency patterns; the DMF-corr block is used to correct and modify the decoding results to improve recognition accuracy; based on Figure 2 The encoder-decoder agent model with multiple time resolutions is shown, thereby realizing accurate mapping of multi-source structural responses and vehicle load inputs;

[0053] 2-2) Based on the pre-processed bridge response data, the finite element model of the target bridge is corrected to generate simulation response samples for MFNI model training, and the average absolute error loss function is optimized through the gradient back propagation algorithm, so that the model can accurately identify the vehicle moving force from the bridge response;

[0054] The generation of simulation response samples includes:

[0055] A1: Correct the finite element model to randomly generate road roughness data;

[0056] A2: Generate a vehicle model with some key parameters to be identified randomly, and use it to generate a large number of simulation training samples.

[0057] Perform parameter sensitivity analysis to evaluate the test performance of the MFNI model under different vehicle parameter conditions, further optimize the model structure and parameter settings, and improve the robustness and recognition accuracy of the model in actual application scenarios;

[0058] 2-3) Use the trained MFNI model to process the real-time collected bridge response data, extract the vehicle moving force features, and obtain the dynamic force time history of each axle on the bridge;

[0059] 2-4) Use the DMF-corr block to correct the output of the decoder to ensure the accuracy of the identification results; through the correction function, the model can more accurately determine the number of vehicle axles and modify the load time history of the non-bridge action time step;

[0060] The load correction function of the DMF-corr block is realized by adding a ReLU activation function after the output fully connected layer, and its specific form is Where z is the neural network output before correction.

[0061] 3) Load weighing network deployment stage

[0062] The wheelbase and vehicle speed information is extracted from the vehicle moving force features output by the MFNI model through a time series feature derivation method;

[0063] The time series feature derivation method solves the wheelbase and vehicle speed information by using the corrected load time history results, which specifically includes:

[0064] B1: Calculate the time step difference value τ of different axles acting on the bridge based on the cross-correlation function i , as follows:

[0065]

[0066] B2: Obtain the up-bridge and down-bridge time steps of the axles by using the first-order difference, so as to obtain the average action time Δt of each axle load on the bridge;

[0067] B3: Substitute and solve the following linear inhomogeneous equation group to obtain the wheelbase a i and the vehicle speed estimate v:

[0068]

[0069] 4) Vehicle-induced response sample acquisition stage

[0070] Integrate the MFNI model into the bridge health monitoring system to realize real-time monitoring and identification of vehicle moving force; the method can be used in multiple fields such as bridge structure health monitoring, traffic flow analysis, overload detection, etc., to improve the intelligent level of bridge management and maintenance;

[0071] Real-time acquisition and separation of vehicle-induced response signals in bridge monitoring signals through the MFNI model, denoising of the vehicle-induced response signals, and division of vehicle-induced response samples by using the sliding window technology.

[0072] 5) Bridge abnormal state diagnosis stage

[0073] Referring to Figure 3 , specifically comprising the following steps:

[0074] 5-1) As shown in (a) and (b) in Figure 3 , a PTAD model is constructed based on PatchTST and grouped convolution, and the PTAD model is trained by using the vehicle-induced response samples, specifically by optimizing the mean square error loss function through gradient back propagation;

[0075] 5-2) Hidden bridge damage sensitive features in the input vehicle-induced response samples are captured by the trained PTAD model;

[0076] 5-3) Damage indicators are obtained by the damage sensitive features, the bridge state is evaluated according to the damage indicators, and the bridge damage degree is quantified;

[0077] The damage indicators include a long-term damage indicator DI long-term and a short-term damage indicator DI short-termBy calculating the reconstruction error of the test data in the PTAD model, the error vector θ, composed of the average REI values ​​of each sensor within the sliding window, is obtained and compared with the reference error vector θ0 under healthy conditions to calculate the DI. long-term and DI short-term ;

[0078] The calculation formula is as follows:

[0079]

[0080] The bridge condition is assessed based on damage indicators, and the degree of bridge damage is quantified. Specifically, this includes: establishing short-term and long-term damage thresholds with a 95% confidence level based on the DIs values ​​obtained under healthy conditions; and using the long-term damage indicator DIs... long-term The bridge is compared with a preset damage threshold to determine if there is potential damage; at the same time, the short-term damage index DI is analyzed periodically. short-term To assess the extent of damage in real time;

[0081] According to DI long-term and DI short-term The calculation results of the indicators quantify the degree of bridge damage, providing a basis for subsequent maintenance decisions.

[0082] To verify the effectiveness and efficacy of the method of the present invention, this embodiment conducts multiple experiments, and the specific data and analysis are as follows:

[0083] Figure 4 Preliminary verification effect diagram of the bridge dynamic weighing system, from Figure 4 As can be seen from the results of load weighing of the input bridge response using the MFNI model, the static load reference value and the load prediction value almost coincide, indicating that the proposed MFNI method can be trained using simulation data to complete the vehicle load weighing task.

[0084] Figure 5 Preliminary validation results for diagnosing long-term abnormal conditions are shown in the diagram. Figure 5 This shows that the DI extracted by the PTAD model long-term The indicator can effectively detect bridge damage by inputting vehicle-induced response samples into the trained PTAD model to extract sensitive features, thereby completing anomaly detection and quantification.

[0085] Figure 6 Preliminary validation results for short-term abnormal state diagnosis are shown in the figure. Figure 6 This shows that the DI extracted by the PTAD model short-term The indicator can effectively detect bridge damage by inputting vehicle-induced response samples into the trained PTAD model to extract sensitive features, thereby completing anomaly detection and quantification.

Claims

1. A vehicle-bridge parameter synchronous inversion identification method based on a deep time series model, characterized in that, The method comprises the following steps: S1: collecting bridge response data and preprocessing; S2: constructing an MFNI model, including an encoder, a decoder, and a dynamic moving load correction DMF-corr block; S3: correcting the finite element model of the target bridge based on the preprocessed bridge response data, generating simulation response samples for MFNI model training; S4: processing real-time collected bridge response data using the trained MFNI model to extract vehicle moving force features and obtain dynamic force time histories of each axle of the vehicle on the bridge; S5: correcting the output of the decoder using the DMF-corr block; S6: extracting wheelbase and vehicle speed information from the vehicle moving force features output by the MFNI model through a time series feature derivation method; S7: real-time collecting and separating vehicle-induced response signals in the bridge monitoring signals through the MFNI model, denoising the vehicle-induced response signals, and dividing the vehicle-induced response samples using a sliding window technique; S8: constructing a PTAD model based on PatchTST and grouped convolution, and training the PTAD model using the vehicle-induced response samples; S9: capturing damage-sensitive features in the vehicle-induced response samples through the trained PTAD model; S10: obtaining damage indicators through the damage-sensitive features, evaluating the bridge state according to the damage indicators, and quantifying the damage degree of the bridge; In step S2, the encoder adopts a dilated causal convolution structure to capture local feature information and interaction dependency patterns in the input data; the decoder uses a recurrent neural network model with a stacked dilated structure to capture complex time-dependent patterns; and the DMF-corr block is used to correct and modify the decoding results to improve recognition accuracy. The load correction function of the DMF-corr block in step S2 is implemented by adding a ReLU activation function after the output full connection layer, and the specific form is wherein, is the output of the neural network before correction. In step S6, the time series feature derivation method solves the wheelbase and vehicle speed information by using the corrected load time history results, specifically including: B1 : Calculate the time step difference value of different axles acting on the bridge based on the cross-correlation function The formula is as follows: ； B2: Obtain the upper and lower bridge time steps of the axle by using first-order difference, so as to obtain the average action time of each axle load on the bridge ; B3: The wheel base is solved by substituting and solving the following linear non-homogeneous equation set and the vehicle speed estimate : ； The damage indicators in step S10 include long-term damage indicators and short-term damage indicators The error vector composed of the average REI values of each sensor in the sliding window is obtained by calculating the reconstruction error of the test data in the PTAD model , and compared with the reference error vector in the healthy state , to calculate and ; The calculation formula is as follows: 。 2. The vehicle-bridge parameter synchronous inversion identification method based on a deep time series model according to claim 1, characterized in that, The generation of simulation response samples in step S3 includes: A1: correcting the finite element model to randomly generate road roughness data; A2: generating a vehicle model with random key to-be-identified parameters to generate simulation training samples.

3. The vehicle-bridge parameter synchronous inversion identification method based on a deep time series model according to claim 1, characterized in that, The bridge state is evaluated according to the damage index in the step S10, and the damage degree of the bridge is quantified, specifically including: the DIs values obtained under the healthy state are counted to establish the short-term and long-term damage thresholds with a confidence of 95%; the long-term damage index is compared with the preset damage threshold to determine whether the bridge has potential damage; meanwhile, the short-term damage index is regularly analyzed to evaluate the damage degree in real time. According to and the results of the damage degree quantification of the bridge damage, the basis for the subsequent maintenance decision is provided.

Images