A lightweight dynamic weighing method based on bridge monitoring response

Abstract

This invention discloses a lightweight dynamic weighing method based on bridge monitoring response. It utilizes a high-definition camera and image processing module to capture monitoring videos of randomly moving vehicles and classify the vehicles. A visual displacement sensor acquires data on the bridge deflection response as vehicles pass. A time synchronization module aligns and matches the video image signals with the bridge deflection response data signals to establish a corresponding dataset. A Bi-LSTM model is used to train the dataset corresponding to specific vehicles. The model's learned load-response mapping relationship for specific vehicles is used to analyze the structural response data under random vehicle loads. The trained Bi-LSTM model outputs the load values ​​of random vehicles. This invention not only effectively collects bridge dynamic response data but also, combined with the load information of specific vehicles, accurately identifies the dynamic load characteristics of random heavy-load traffic flows, providing technical support for load identification and safety management of small- and medium-span bridges.

Description

Technical Field

[0001] This invention belongs to the field of bridge health monitoring and vehicle dynamic weighing, specifically relating to a lightweight dynamic weighing method based on bridge monitoring response. By deploying lightweight structural response monitoring devices on the bridge and combining this with load information of specific vehicles on the bridge deck, dynamic weighing of random heavy-load traffic flows can be achieved. Background Technology

[0002] Traffic flow load is the main variable load borne by highway bridge structures, exhibiting a high degree of randomness in time and space. With the rapid development of transportation, the actual traffic flow load on bridges has far exceeded the design load standards. Vehicle overloading and the repeated action of traffic flow loads have become major factors affecting the operational safety of bridges and significantly shortening their service life. Therefore, accurate identification of traffic flow loads is crucial for bridge structural design and safety management.

[0003] In traffic flow load analysis, a key vehicle characteristic parameter is load information. The introduction of Weigh-In-Motion (WIM) systems has revolutionized vehicle load detection and has become an important tool for load measurement. WIM systems collect key data such as axle load, gross weight, speed, and arrival time of vehicles through piezoelectric sensors deployed on the road surface. The technology is mature and widely used in practical engineering. However, due to limitations such as high cost and complex construction, WIM systems are difficult to popularize in small-to-medium span bridges. Therefore, methods for vehicle load identification based on bridge dynamic response urgently need to overcome existing technological bottlenecks.

[0004] This invention aims to achieve dynamic vehicle weighing on small-to-medium span bridges through a lightweight bridge health monitoring system based on bridge monitoring responses. This system not only effectively collects bridge dynamic response data but also accurately identifies the dynamic load characteristics of random heavy-load traffic flows by combining specific vehicle load information. The purpose of this invention is to provide a low-cost, easy-to-deploy, and reliable weighing system, offering technical support for load identification and safety management of small-to-medium span bridges. Summary of the Invention

[0005] In view of the above-mentioned technical problems existing in the prior art, the purpose of this invention is to provide a lightweight dynamic weighing method based on bridge monitoring response.

[0006] The technical solution adopted in this invention is as follows:

[0007] A lightweight dynamic weighing method based on bridge monitoring response includes the following steps:

[0008] 1) A video surveillance system is installed on the bridge deck. The video surveillance system includes high-definition cameras and image processing modules. The high-definition cameras capture surveillance videos of randomly moving vehicles. The image processing module automatically identifies and classifies the vehicles in the surveillance videos. The vehicle classification includes ordinary vehicles and special vehicles transporting dangerous goods. Special vehicles are further subdivided according to the different types of dangerous goods being transported. The load of different types of special vehicles is a known fixed value.

[0009] 2) A high-precision visual displacement sensor is deployed at the mid-span of the bridge superstructure. The visual displacement sensor captures high-resolution images of the target area at the mid-span of the bridge superstructure under random vehicle loads using an industrial camera, and calculates the dynamic response data of the bridge structure, i.e., the deflection response data of the bridge, using DIC technology.

[0010] 3) The time synchronization module aligns and matches the random traffic flow monitoring video image signals obtained by the video surveillance system with the bridge structure dynamic response data signals obtained by the visual displacement sensor on a preset time axis. For random traffic flow on the bridge, the vehicle load is calibrated as follows: w j The corresponding bridge structure dynamic response data is calibrated as follows: x j Establish a spatiotemporal correlation dataset for constructing mapping relationships;

[0011] Identify and extract special vehicles from random traffic flow monitoring video images, and calibrate the load of the special vehicles as follows. w i And its corresponding bridge structure dynamic response data calibration is as follows x i Establish a spatiotemporally correlated subset for learning the feature values ​​of the junction response;

[0012] 4) Establish a bidirectional long short-term memory network (Bi-LSTM) model. The spatiotemporal correlation subset dataset used for learning the structural response feature values ​​established in step 3) is input into the model for training. Then, the total spatiotemporal correlation dataset used to construct the mapping relationship is input into the trained Bi-LSTM model. Using the specific vehicle load-response mapping relationship learned by the model, the structural response data under random vehicle action is analyzed, and the load value of random vehicle is output.

[0013] Further, in step 1), the image processing module uses multi-view regression detection technology to automatically identify and classify vehicle information in the surveillance video. First, it roughly classifies the vehicles into small vehicles, heavy-duty vehicles, and special vehicles. Then, it further detects the special vehicles in the rough classification by recognizing the side photos of the vehicles to detect whether there are dangerous goods transport signs. After confirming the presence of the signs, it analyzes the content of the signs for chemical dangerous goods vehicles to identify the specific type of the signs. Combining the vehicle shape and the sign information, it achieves accurate classification of special vehicles.

[0014] Furthermore, in step 2), the process of calculating the dynamic response data of the bridge structure using DIC technology involves employing an image processing algorithm with sub-pixel level precision to ensure the high accuracy and stability of the monitoring results. During the monitoring process, high-frequency image sampling and dynamic data processing are combined to generate a continuous time series of dynamic response data of the bridge structure for subsequent load identification and structural performance evaluation.

[0015] Furthermore, the specific processing procedure of the DIC technology is as follows:

[0016] S1: First, perform camera calibration to eliminate distortion and establish the relationship between image pixels and actual size to obtain calibration coefficients, in mm / pixel;

[0017] S2: Next, the target displacement video is acquired by the camera and the image is converted into a grayscale image. The pixel coordinates of the target point to be measured on the target are selected. The data is preprocessed and outlier removal is performed. The preprocessing includes data extraction and benchmark extraction. First, the time history data of the bridge deflection response before and after the vehicle passes is extracted, which is the change of the pixel coordinates of the target point over time. The initial position pixel coordinates of the target point before the vehicle passes are taken as the benchmark value. Outlier removal uses the interquartile range method to remove extreme deviation data in the original data.

[0018] S3: The cross-correlation function is used to calculate the positional relationship of the target point on the image. The pixel coordinate difference between the target point and its initial position when the vehicle passes is calculated, which is the pixel displacement, in mm / pixel. The pixel displacement is calculated for each frame of the monitoring video of the target point to be measured, and the pixel displacement time history of the target point is obtained as the extracted feature value.

[0019] S4: Finally, the extracted pixel displacement feature values ​​of the target point to be measured are multiplied by the calibration coefficient to convert them into actual physical displacement, thus obtaining the displacement time history of the target point to be measured.

[0020] Furthermore, in step S3, an image processing algorithm with subpixel-level precision is used to perform two searches on the target point: an integer pixel search and a subpixel search. The integer pixel search finds the integer pixel position of the target point in the target image, thus obtaining the integer pixel coordinates of the target point. The subpixel search is performed based on the integer pixel search to obtain the precise position of the pixel. The obtained pixel coordinates of the target point are decimals. This process improves the accuracy of the monitoring data.

[0021] Further, the spatiotemporal correlated total / sub-datasets established in step 3) specifically refer to matching vehicle information and structural response data obtained from multiple sensors during a single vehicle passage by corresponding acquisition times, and storing them in the dataset. The spatiotemporal correlated total / sub-datasets specifically include: the video surveillance image acquisition time, license plate, vehicle type, vehicle weight, and structural deflection response data for each vehicle's passage each time. The spatiotemporal correlated total dataset includes data information for all vehicles passing over the bridge; while the spatiotemporal correlated sub-dataset only contains vehicle data information for vehicles of a specific special type.

[0022] Furthermore, step 4) before training the Bi-LSTM model on the dataset also includes a feature extraction step, the specific steps of which are as follows:

[0023] M1: First, the Mel frequency cepstral coefficient features in the main beam deflection monitoring response signal when different vehicles pass are extracted. The signal is then denoised to remove high-frequency noise. Next, the signal is framed and smoothed using a Hanning window function. Then, a Fast Fourier Transform (FFT) is used to convert the signal from the time domain to the frequency domain. Next, a Mel filter bank is used to convert the spectrum to a Mel frequency scale, and the energy of each Mel filter channel is calculated to obtain the Mel spectrum. Finally, a Discrete Cosine Transform (DCT) is used to convert the Mel spectrum into Mel frequency cepstral coefficients (MFCCs). The first m MFCCs coefficients are selected as feature vectors. For each selected MFCC coefficient, its first and second order differences are further calculated to enhance dynamic change information. The MFCCs, first-order differences, and second-order difference features are concatenated to form the final feature vector. X i These feature vectors effectively characterize the dynamic response of the main beam deflection when the vehicle passes, and are suitable for subsequent signal analysis and classification.

[0024] The formulas for the Mel frequency conversion and Discrete Cosine Transform (DCT) are as follows:

[0025] Mel frequency conversion:

[0026] .

[0027] in, f It is a linear frequency. M(f) It is the Mel frequency;

[0028] DCT Transformation:

[0029] .

[0030] in C k These are cepstral coefficients. E(f n ) It is the Mel spectrum. N Represents signal length. n This represents the sampling point index of the time-domain signal, and its value range is 0 ≤ n ≤ N -1; k It is an index of MFCCs;

[0031] M2: Based on the MFCCs feature values ​​obtained in step M1, and using the chi-square test, features highly correlated with the classification task are selected. The specific process is as follows:

[0032] First, classify the MFCCs feature values ​​according to labels of vehicle type, load level, speed, and number of axles, and then calculate the expected frequency using the following formula:

[0033] .

[0034] in, O j There are a total of 3m features for MFCCs and first-order and second-order differences. X i The total number of samples with value j is given, and m is the number of MFCC coefficients selected. The total number of selected MFCC coefficients and their first-order and second-order difference features is 3m. O k For the tag k The total number of samples, O The total number of all samples; E jk Let j be the expected frequency of the sample with label k and value j.

[0035] Next, regarding the eigenvalues X i Calculate the chi-square statistic x using the following formula 2 ,in O jk It represents the number of samples with label k and value j, i.e., the actual frequency. E jk For the expected frequency, a It is a feature X i The number of different values, b It is the number of tag categories;

[0036] .

[0037] Next, based on the selected significance level and degrees of freedom, consult the chi-square distribution table to find the corresponding chi-square critical value χ². critical 2 ; Keep the calculated chi-square value x 2 Greater than the critical value χ critical 2 This step involves identifying significantly relevant feature values ​​and filtering out highly relevant features to reduce redundant features.

[0038] Finally, to verify the accuracy of the vehicle load dynamic identification method described in this invention, a field test was conducted on 500 sets of random vehicle sample data under test conditions of vehicle speed 30-100 km / h and total vehicle weight 10-60 tons. The verification results are shown in [the table below]. Figure 5 The Bi-LSTM model trained according to this invention is used to analyze the structural response of a random vehicle, and the output load value of the random vehicle is thus determined. Figure 5 The system identifies vehicle weight. Results show that 90% of the samples have a relative error of less than 10% between the identified load value and the actual load value, and the overall system error is stably controlled within ±10%.

[0039] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0040] Traditional dynamic weighing systems rely heavily on dedicated sensor arrays and complex hardware for vehicle load monitoring, resulting in three significant drawbacks: First, the procurement and deployment of high-precision weighing sensors and related equipment leads to high overall system costs; second, retrofitting existing bridges requires breaking through the bridge deck pavement for embedded installation, presenting engineering challenges such as long construction periods and significant traffic disruption; and third, sensors subjected to long-term vehicle pressure are prone to fatigue damage, limiting system reliability and lifespan. In contrast, this invention's lightweight dynamic weighing system based on structural response monitoring innovatively constructs a "vehicle-bridge" dynamic coupling inversion model by fusing random vehicle information from the bridge deck with structural deflection response data. This significantly reduces reliance on dedicated sensors while employing non-invasive monitoring equipment for lightweight system deployment. This system not only reduces hardware costs by approximately 90% and avoids destructive construction of the bridge deck structure, but also improves equipment durability by more than three times through optimized stress transmission paths, ultimately achieving 90% accuracy in dynamic vehicle load identification within a low-cost operation and maintenance framework. Attached Figure Description

[0041] Figure 1 This is a technical flowchart of a lightweight dynamic weighing monitoring method based on bridge monitoring response of a specific vehicle according to the present invention.

[0042] Figure 2 This is a schematic diagram of the lightweight bridge monitoring system.

[0043] Figure 3 This is a diagram of the Bidirectional Long Short-Term Memory (BiLSTM) network structure.

[0044] Figure 4 A diagram illustrating the markings for vehicles transporting hazardous chemicals.

[0045] Figure 5 Figure showing the results of vehicle weight calculation during the testing phase. Detailed Implementation

[0046] A lightweight dynamic weighing system based on bridge monitoring response of specific vehicles includes the following three modules: (1) a structural response monitoring module based on visual displacement sensor; (2) a bridge vehicle identification and classification module based on multi-view regression technology; and (3) a load-response mapping model establishment module based on bidirectional long short-term memory network.

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0048] like Figure 1 The lightweight dynamic weighing monitoring method based on bridge monitoring response under specific vehicles, as described in this embodiment, includes the following steps:

[0049] Step 1: Based on visual displacement sensors and high-definition cameras, a lightweight monitoring technology system for small and medium-span bridges is constructed to realize the monitoring of structural response under random vehicle loads and the dynamic capture of traffic flow.

[0050] For details, see Figure 2 High-precision visual displacement sensors are deployed at the mid-span of the bridge superstructure to monitor the dynamic displacement response of the structure. Simultaneously, high-definition cameras are positioned at both ends of the bridge deck to capture the shape characteristics and trajectories of vehicles, enabling the synchronous acquisition of vehicle data such as license plates, driving lanes, and vehicle types, as well as structural response data such as dynamic deflection at mid-span caused by vehicles. The high-definition cameras are connected to an image processing module.

[0051] To mitigate environmental interference, the visual displacement sensor and high-definition camera are designed to resist interference during installation and are equipped with sunshades or waterproof housings to withstand harsh conditions.

[0052] Step 2: The visual displacement sensor uses digital image correlation (DIC) technology to achieve real-time dynamic monitoring of the mid-span deflection of the bridge.

[0053] Specifically, the DIC technology uses high-resolution images of target areas at the mid-span of a bridge captured by an industrial camera, and calculates deflection changes based on correlation analysis of the image grayscale distribution. Visual displacement sensors record the vertical displacement response of the target points at the mid-span of the bridge under random vehicle loads in real time, and sub-pixel-level precision image processing algorithms ensure high accuracy and stability of the monitoring results. During monitoring, high-frequency image sampling and dynamic data processing are combined to generate a continuous time series of bridge mid-span deflection for subsequent load identification and structural performance evaluation.

[0054] Specifically, the DIC technology first performs camera calibration to eliminate distortion and establishes the relationship between image pixels and actual size, deriving calibration coefficients (unit: mm / pixel). Next, it acquires target displacement video using the camera and converts the image into a grayscale image. The pixel coordinates of the target point to be measured on the target are selected, and the data undergoes preprocessing and outlier removal. Preprocessing includes data truncation and baseline value extraction.

[0055] Specifically, the data preprocessing mainly includes the steps of data extraction and baseline value extraction. First, the time history data of the bridge deflection response 15 seconds before and after the vehicle passes are extracted, which is the change of the pixel coordinates of the target point under test over time. The initial position pixel coordinates of the target point under test before the vehicle passes are taken as the baseline value.

[0056] Outlier removal uses IQR (interquartile range) to remove extreme deviations from the original data.

[0057] The cross-correlation function is used to calculate the positional relationship of the target point on the image. The pixel coordinate difference between the target point and its initial position when the vehicle passes is calculated, which is the pixel displacement (unit: pixel). The pixel displacement is calculated for each frame of the monitoring video of the target point to obtain the pixel displacement time history of the target point. The maximum absolute value of the pixel displacement of the target point from its initial position is taken as the extracted feature value.

[0058] Finally, the extracted pixel displacement feature values ​​of the target point to be measured are multiplied by the calibration coefficient to convert them into actual physical displacement, thus obtaining the displacement time history of the target point to be measured.

[0059] The expression for the cross-correlation function is as follows:

[0060] .

[0061] in f(i , j) It is the grayscale function of the target point sub-region to be measured. (i , j) These are the pixel coordinates of the target point sub-region to be measured. g(i+y , j+x) It is the grayscale function of the search image sub-region, which is the distance from the sub-region of the target to be measured. (i+y , j+x) These are the pixel coordinates of the search image sub-region. The difference between the pixel coordinates of the search image sub-region and the target point sub-region is (x, ...). y).

[0062] w This is the current search window size. C CC It is the cross-relationship count between the two sub-regions of the current search window. (x,y) It is the displacement between two sub-regions, when C CC When the maximum value is found within the search area, the corresponding (x,y) It is the maximum displacement between two sub-regions, which is the absolute maximum value of the pixel displacement of the target point from its initial position.

[0063] The sub-pixel level precision image processing algorithm, based on the aforementioned cross-correlation function calculation method, performs two searches on the target point (i.e., using the aforementioned cross-correlation function to calculate the maximum absolute value of the pixel displacement of the target point from its initial position; the coordinates (x, y) for integer pixel searches are integers, while those for sub-pixel searches are decimals), namely, an integer pixel search and a sub-pixel search. The integer pixel search finds the integer pixel position of the target point in the target image, thus obtaining the integer pixel coordinates of the target point; the sub-pixel search, based on the integer pixel search, obtains the precise position of the pixel, and the resulting pixel coordinates of the target point are decimals. This process improves the accuracy of the monitoring data.

[0064] Step 3: Automatic identification and classification of vehicle flow information in surveillance videos is achieved based on multi-view regression detection technology. First, vehicle types are roughly categorized (e.g., small vehicles, heavy-duty vehicles, and special vehicles). Then, special vehicles in the rough classification are further inspected to check for hazardous materials transport markings (e.g., hazardous chemical markings affixed to the vehicle body). The content of these markings is analyzed to identify their specific types (e.g., explosives, compressed gases, and liquefied gases), achieving precise classification of special vehicles. A schematic diagram of hazardous chemical vehicle markings is shown below. Figure 4 The following are listed as hazardous chemical symbols: 1. Explosives, 2. Compressed and liquefied gases, 3. Flammable liquids, 4. Flammable solids, 5. Oxidizing agents and organic peroxides, 6. Toxic and infectious substances, 7. Radioactive materials.

[0065] Step 4: Based on the time-series correlation of vehicle-bridge monitoring data, establish a spatiotemporal correlation dataset of vehicle load-structure response.

[0066] The time synchronization module aligns and matches the random traffic flow monitoring video image signals obtained by the video surveillance system with the bridge structure dynamic response data signals obtained by the visual displacement sensor on a preset time axis. For random traffic flow on the bridge, the vehicle load is calibrated as follows: w j The corresponding bridge structure dynamic response data is calibrated as follows: x j Establish a spatiotemporal correlation dataset for constructing mapping relationships;

[0067] Using target detection technology, special vehicles are identified and extracted from random traffic flow monitoring video images, and the load of these special vehicles is calibrated as follows. w i And its corresponding bridge structure dynamic response data calibration is as follows x i To establish a spatiotemporally correlated subset for learning the response feature values ​​of the structure.

[0068] Since the specific vehicle type selected in this technical solution has a relatively constant load, especially chemical hazardous materials transport vehicles which have strict transportation standards and whose loads are basically fixed, and whose corresponding structural responses are known, the specific vehicle loads can be input as known values ​​during model training. w i and the known structural response x i Perform feature learning.

[0069] Step 5: Based on the Bidirectional Long Short-Term Memory (Bi-LSTM) network, the structural response time history features under specific vehicle loads are extracted through autonomous learning to construct a mapping relationship model between vehicle load and structural response.

[0070] Specifically, firstly, Mel-Frequency Cepstral Coefficients (MFCCs) features are extracted from the main beam deflection monitoring response signals when different vehicles pass by, and then the chi-square test is used to select features highly correlated with the classification task. Subsequently, a Bi-LSTM model is constructed to capture the temporal dependence of the signals, and an attention mechanism is introduced to improve the accuracy and robustness of the mapping model between the monitoring signals and vehicle loads. Through model training, specific vehicle load parameters are accurately reconstructed, thereby achieving efficient inversion of heavy-duty vehicle load parameters.

[0071] The extraction of Mel frequency cepstral coefficients (MFCCs) features first employs Variational Mode Decomposition (VMD) for signal denoising to remove high-frequency noise. Then, the signal is framed, and each frame is smoothed using a Hanning window function to reduce spectral leakage. Next, a Fast Fourier Transform (FFT) is used to convert the signal from the time domain to the frequency domain. Then, a Mel filter bank is used to convert the linear spectrum to a Mel frequency scale, and the energy of each Mel filter channel is calculated to obtain the Mel spectrum. Finally, a Discrete Cosine Transform (DCT) is used to convert the Mel spectrum into Mel frequency cepstral coefficients (MFCCs), and the first 13 MFCC coefficients are selected as feature vectors. For each selected MFCC coefficient, its first and second order differences (Δ and ΔΔ) are further calculated to enhance dynamic variation information. The MFCCs, first-order differences, and second-order difference features are concatenated to form the final feature vector. X i A total of 39 features were identified. These feature vectors effectively characterize the dynamic response of the main beam deflection when a vehicle passes, and are suitable for subsequent signal analysis and classification. The key steps involved are formulated as follows:

[0072] Mel filter bank, Mel frequency conversion:

[0073] .

[0074] in, f It is a linear frequency. M(f) It is the Mel frequency.

[0075] DCT Transformation:

[0076] .

[0077] in C k These are cepstral coefficients. E(f n ) It is the Mel spectrum (calculated from the energy of each Mel filter channel). k It is an index of MFCC. N Represents the signal length (or block size). n The sampling point index of the time-domain signal (value range 0 ≤ n ≤ N -1). In the formula for one-dimensional DCT, n represents the sampling point index of the time-domain signal. For example, for a signal sequence of length N, the value of n ranges from 0 ≤ n ≤ N-1, corresponding to the nth sampling point of the signal.

[0078] The chi-square test first categorizes the MFCCs feature values ​​according to labels such as vehicle type, load level, speed, and number of axles, and then calculates the expected frequency using the following formula. O j Features X i The total number of samples with value j, where j is the vector set of the feature values. O k For the tag k The total number of samples, O The total number of all samples.

[0079] .

[0080] O j A total of 39 features are derived from MFCCs and their first-order and second-order differences. X i The total number of samples with a value of j (where the value of j for MFCCs is in the range of [-50, 50], and the value of j for first-order and second-order differences is in the range of [-20, 20]).

[0081] Next, regarding the eigenvalues X i Calculate the chi-square statistic x using the following formula 2 ,in O jk It represents the number of samples with label k and value j, i.e., the actual frequency. E jk For the expected frequency, a It is a feature X i The number of different values, b It represents the number of tag categories.

[0082] .

[0083] Next, based on the selected significance level (usually 0.05) and degrees of freedom, find the corresponding chi-square critical value χ² by referring to the table. critical 2 Retain the calculated chi-square value (x). 2 ) greater than the critical value (χ) critical 2 This step involves selecting highly relevant features and reducing redundant features.

[0084] The Bidirectional Long Short-Term Memory (Bi-LSTM) model combines forward and backward LSTM networks to comprehensively capture the contextual information of the input feature value sequence. The network structure diagram of BiLSTM is shown below. Figure 3First, the input time-series data sequence is fed into the Bi-LSTM network. Then, the input sequence is processed from left to right by the forward LSTM unit to capture forward dependencies; simultaneously, the input sequence is also processed from right to left by the backward LSTM unit to capture backward dependencies. Subsequently, at each time step, the Bi-LSTM combines the outputs of the forward and backward LSTMs to enhance its expressive power for the time-series data. Finally, the output combining the forward and backward information is used for subsequent tasks.

[0085] Step 6: Input the structural response data under random vehicle loads into the Bi-LSTM model established in Step 5. Utilize the specific vehicle load-response mapping relationship learned by the model to analyze the structural response of the random vehicles and output the load values ​​of the random vehicles. Experimental results show that the model exhibits significant differences in identifying vehicles of different load classes: for heavy-load vehicles, the identification accuracy is significantly better than that for medium-load and light-load vehicles, with an accuracy rate exceeding 95%. Through this process, dynamic identification of random vehicle loads on bridges is achieved, providing data support for bridge health monitoring and overload control.

Claims

1. A lightweight dynamic weighing method based on bridge monitoring response, characterized in that... Includes the following steps: Step 1: A video surveillance system is installed on the bridge deck. The video surveillance system includes high-definition cameras and an image processing module. The high-definition cameras capture surveillance videos of randomly moving vehicles. The image processing module automatically identifies and classifies the vehicles in the surveillance video. The vehicle classification includes ordinary vehicles and special vehicles transporting dangerous goods. Special vehicles are further subdivided according to the different types of dangerous goods being transported. The load of different types of special vehicles is a known fixed value. Step 2: Deploy high-precision visual displacement sensors at the mid-span of the bridge superstructure. These sensors use industrial cameras to capture high-resolution images of the target area at the mid-span of the bridge superstructure under random vehicle loads, and use DIC technology to calculate the dynamic response data of the bridge structure, i.e., the deflection response data of the bridge. Step 3: Using a time synchronization module, align and match the random traffic flow monitoring video image signal obtained by the video surveillance system with the bridge structure dynamic response data signal obtained by the visual displacement sensor on a preset time axis. For random traffic flow on the bridge, calibrate its vehicle load as w. j The corresponding bridge structure dynamic response data is calibrated as x. j Establish a spatiotemporal correlation dataset for constructing mapping relationships; Special vehicles are identified and extracted from random traffic flow monitoring video images, and their load is calibrated as w. i And its corresponding bridge structure dynamic response data is calibrated as x i Establish a spatiotemporally correlated subset for learning the feature values ​​of the junction response; Step 4: Establish a bidirectional long short-term memory network (Bi-LSTM) model. The spatiotemporal correlation subset dataset used for learning structural response feature values ​​established in Step 3 is input into the model for training. Then, the total spatiotemporal correlation dataset used to construct the mapping relationship is input into the trained Bi-LSTM model. Using the specific vehicle load-response mapping relationship learned by the model, the structural response data under random vehicle action is analyzed, and the load value of the random vehicle is output. Step 4, before training the Bi-LSTM model on the dataset, also includes a feature extraction step, the specific steps of which are as follows: M1: First, extract the Mel frequency cepstral coefficient features from the main beam deflection monitoring response signal when different vehicles pass by, and then denoise the signal to remove high-frequency noise. Subsequently, the signal is framed and smoothed using the Hanning window function. Then, a Fast Fourier Transform (FFT) is used to convert the signal from the time domain to the frequency domain. Next, a Mel filter bank is used to convert the spectrum to a Mel frequency scale, and the energy of each Mel filter channel is calculated to obtain the Mel spectrum. Finally, a Discrete Cosine Transform (DCT) is used to convert the Mel spectrum into Mel frequency cepstral coefficients (MFCCs), and the first m MFCCs coefficients are selected as feature vectors. For each selected MFCC coefficient, its first and second order differences are further calculated to enhance dynamic change information. The MFCCs, first-order differences, and second-order difference features are concatenated to form the final feature vector X. i These feature vectors effectively characterize the dynamic response of the main beam deflection when the vehicle passes, and are suitable for subsequent signal analysis and classification. The spectrum is converted to a Mel frequency scale using a Mel filter bank, and the conversion formula is as follows: Where f is the linear frequency and M(f) is the Mel frequency; The formula for DCT transform is as follows: Where C k These are cepstral coefficients, E(f) n ) is the Mel spectrum, N represents the signal length, n represents the sampling point index of the time domain signal, and its value range is 0≤n≤N-1; k is the index of MFCCs; M2: Based on the MFCCs feature values ​​obtained in step M1, and using the chi-square test, features highly correlated with the classification task are selected. The specific process is as follows: First, classify the MFCCs feature values ​​according to labels of vehicle type, load level, speed, and number of axles, and then calculate the expected frequency using the following formula: Among them, O j For MFCCs and a total of 3m features X with first-order and second-order differences i The total number of samples with value j, m is the number of MFCC coefficients selected, and the total number of selected MFCC coefficients and their first and second difference features is 3m. k Let O be the total number of samples with label k, and O be the total number of all samples; E jk Let j be the expected frequency of the sample with label k and value j. Next, for the eigenvalue X i Calculate the chi-square statistic x using the following formula 2 O jk E represents the number of samples with label k and value j, i.e., the actual frequency. jk Let be the expected frequency, and 'a' be the feature X. i The number of different values ​​for , where b is the number of label categories; Next, based on the selected significance level and degrees of freedom, consult the chi-square distribution table to find the corresponding chi-square critical value χ². critical 2 ; Keep the calculated chi-square value x 2 Greater than the critical value χ critical 2 This step involves identifying significantly relevant feature values ​​and filtering out highly relevant features to reduce redundant features.

2. The lightweight dynamic weighing method based on bridge monitoring response as described in claim 1, characterized in that... In step 1, the image processing module uses multi-view regression detection technology to automatically identify and classify vehicle information in the surveillance video. First, it roughly classifies the vehicles into small vehicles, heavy-duty vehicles, and special vehicles. Then, it further detects the special vehicles in the rough classification by recognizing the side photos of the vehicles to check for the presence of dangerous goods transport signs. After confirming the presence of the signs, it analyzes the content of the chemical dangerous goods vehicle signs to identify their specific types. Combining the vehicle shape and sign information, it achieves accurate classification of special vehicles.

3. The lightweight dynamic weighing method based on bridge monitoring response as described in claim 1, characterized in that... In step 2, the process of calculating the dynamic response data of the bridge structure using DIC technology involves employing an image processing algorithm with sub-pixel precision to ensure the high accuracy and stability of the monitoring results. During the monitoring process, high-frequency image sampling and dynamic data processing are combined to generate a continuous time series of dynamic response data of the bridge structure for subsequent load identification and structural performance evaluation.

4. The lightweight dynamic weighing method based on bridge monitoring response as described in claim 3, characterized in that... The specific processing procedure of the DIC technology is as follows: S1: First, perform camera calibration to eliminate distortion and establish the relationship between image pixels and actual size to obtain calibration coefficients, in mm / pixel; S2: Next, the target displacement video is acquired by the camera and the image is converted into a grayscale image. The pixel coordinates of the target point to be measured on the target are selected. The data is preprocessed and outlier removal is performed. The preprocessing includes data extraction and benchmark extraction. First, the time history data of the bridge deflection response before and after the vehicle passes is extracted, which is the change of the pixel coordinates of the target point over time. The initial position pixel coordinates of the target point before the vehicle passes are taken as the benchmark value. Outlier removal uses the interquartile range method to remove extreme deviation data in the original data. S3: The cross-correlation function is used to calculate the positional relationship of the target point on the image. The pixel coordinate difference between the target point and its initial position when the vehicle passes is calculated, which is the pixel displacement, in mm / pixel. The pixel displacement is calculated for each frame of the monitoring video of the target point to be measured, and the pixel displacement time history of the target point is obtained. The maximum absolute value of the pixel displacement of the target point from its initial position is taken as the extracted feature value. S4: Finally, the extracted pixel displacement feature values ​​of the target point to be measured are multiplied by the calibration coefficient to convert them into actual physical displacement, thus obtaining the displacement time history of the target point to be measured.

5. A lightweight dynamic weighing method based on bridge monitoring response as described in claim 4, characterized in that... In step S3, an image processing algorithm with subpixel-level precision is used to perform two searches on the target point: an integer pixel search and a subpixel search. The integer pixel search finds the integer pixel position of the target point in the target image, thus obtaining the integer pixel coordinates of the target point. The subpixel search is performed based on the integer pixel search to obtain the precise position of the pixel. The obtained pixel coordinates of the target point are decimals. This process improves the accuracy of the monitoring data.

Images